results.json

{
 "schema": 1,
 "updated": "2026-10-16",
 "about": "Every benchmark number this repository has measured or relies on. Model output, not readings. AUC: pixel AUC against the segment's 20260918 human labels inside its supervision mask (kit auc). auc_reversed: the same map read with depth reversed, the control. auc_shuffled: a map made with the depth layers shuffled, a second control. window crop: the segment's 640 px densest-labelled-ink window; inner_px: edge left out; map_from: full = cut from a whole-segment map, bare-crop = inferred on the crop alone (equal to full when inner_px is 64).",
 "held_out": {
  "w045": "held out from v8in only. ink_9um trained on this surface through its 2.399 um render pooled to 9.6 um (label dataset segment pherc0139-w029); d9v2 and Reader v2 start from ink_9um, and Reader v2 also trained on w045. Our scores use the native 9.362 um render: a cross-scan check on a training surface for those readers. See docs/logs/2026-10-08-w045-not-held-out.md. Not held out from Hecate or the team's canonical 2.4 um model either: the canonical trainer (ScrollPrize/villa 94602c91 ink-detection/train_resnet3d_3d_decoder.py, lines 200-207) lists this surface as w029_202601261946, and the published Hecate split rule (villa branch distilled_3d_ink, 5ff00ca0, vesuvius/src/vesuvius/ink_detection/data/multiteacher.py) puts 86.8 % of its supervised pixels under training patches (our computation on the ink dataset's masks, 2026-10-16). See docs/logs/2026-10-16-reader-exposure.md",
  "PHerc0841": "in the training data of none of ink_9um (its label dataset has no PHerc0841 segment), v8in and v8in-1447 (v8in's training/configs/v8in.json at d89166b and its patch pack list none; v8in-1447 adds only PHerc1447 w058 and w060), d9v2 (all three are in TAUIL's reader/test.txt at c156944, Community report) and Reader v2 (its train_config.json lists none, Community report): the held-out test for those readers. Not held out from Hecate (Interpretation, settled on 2026-10-16 as far as public sources go). Its own published trainer's split rule (ScrollPrize/villa branch distilled_3d_ink, 5ff00ca0, data/multiteacher.py), applied to the ink dataset's published masks, makes all three segments validation-only, since every supervised pixel lies inside their validation_mask_v2 (0 training patches; our computation; the archived manifest of the released run is not public). But its starting weights, scrollprize/ink_canonical_2um, come from a committed trainer (villa 94602c91 ink-detection/train_resnet3d_3d_decoder.py, lines 200-207, 292-293 and 794) that lists w00, ag405 and ag896 as training fragments and reads no validation mask; it keeps tiles only inside each segment's own _mask file, whose content in the team's 2um_dataset is not public (Sourced fact; that the released checkpoint used this list is Interpretation, from its matching name, epoch and step count). Weaker channels: checkpoint selection on a validation mean that would include PHerc0841, and the Paris 4 3D teacher (scrollprize/ink_3d_dino_guided), whose first stage trained on eight scrolls' labels including 0841 but which made Hecate's targets for Paris 4 only, so indirect at most (Interpretation). Team members also said on the community Discord in September 2026 that Hecate's training used almost all of the available ink labels and that the released label sets had been at least partly used to train the published models (Community reports, paraphrased). Treat Hecate's PHerc0841 scores, ensembles that include Hecate, and the team's new_canon_autoresearch_recipe maps of PHerc0841 as possibly in-sample. See docs/logs/2026-10-16-reader-exposure.md",
  "w043": "PHerc0139 w043 (20260112000000-w043_2026011217): not held out from ink_9um (label dataset segment pherc0139-w043), d9v2 (reader/train.txt, and inherited from ink_9um), Reader v2 (train_config.json, twice, and inherited) or, very likely, Hecate: the ink dataset holds w043 with no validation mask, so the published Hecate split rule (villa distilled_3d_ink 5ff00ca0) holds out two generated 1024 px blocks (15.3 % of its supervised pixels) and puts 74.5 % under training patches (our computation, 2026-10-16; 68.2 % to 93.7 % over seeds 0 to 39, so the verdict does not depend on the seed; the archived manifest is not published). It is not in the canonical 2.4 um trainer's committed list (villa 94602c91, lines 200-207) or in v8in's. A ranking of these readers on w043 is mostly in-sample and favours Hecate over canonical 2um. See docs/logs/2026-10-16-reader-exposure.md",
  "PHerc1447": "w058 and w060: held out from every released reader except v8in-1447, which fine-tuned on both (training/configs/loo_w062.json at 2bf9f42). Neither is in the label lists of ink_9um, v8in, d9v2 or Reader v2, in the ink dataset behind Hecate (it has no PHerc1447 folder) or in the canonical trainer's committed list. w062's labels are not released. See docs/logs/2026-10-16-reader-exposure.md"
 },
 "held_out_by_reader": {
  "about": "Per reader family, whether each test group was outside its training data, in one word, for python -m kit leaderboard and python -m kit integrity. The reasons and sources are in held_out above and in docs/logs/2026-10-16-reader-exposure.md; keep the two in step. A row for a family with no status here fails kit leaderboard check.",
  "w045": {
   "v8in": "held out",
   "ink_9um": "not held out",
   "d9v2": "not held out",
   "Reader v2": "not held out",
   "Hecate": "not held out"
  },
  "PHerc0841": {
   "ink_9um": "held out",
   "v8in": "held out",
   "v8in-1447": "held out",
   "d9v2": "held out",
   "Reader v2": "held out",
   "Hecate": "not held out"
  }
 },
 "segments": {
  "w045": {
   "scroll": "PHerc0139",
   "segment": "20260126000000-w045_2026012619",
   "voxel_um": 9.362,
   "surface_shape": [
    5980,
    8240
   ],
   "crop": [
    3840,
    4480,
    2560,
    3200
   ],
   "labels": "ink-labels/2.399um-volume-20260102150214/20260918",
   "label_level": "2"
  },
  "0841-w00": {
   "segment": "20260220213127-w00",
   "surface_shape": [
    4220,
    4760
   ],
   "crop": [
    2624,
    3264,
    2688,
    3328
   ],
   "scroll": "PHerc0841",
   "voxel_um": 9.366,
   "labels": "ink-labels/2.403um-volume-20260319124803/20260918",
   "label_level": "2"
  },
  "0841-ag896": {
   "segment": "20260220214732-auto_grown_20260220144552896",
   "surface_shape": [
    4640,
    4720
   ],
   "crop": [
    2496,
    3136,
    1600,
    2240
   ],
   "scroll": "PHerc0841",
   "voxel_um": 9.366,
   "labels": "ink-labels/2.403um-volume-20260319124803/20260918",
   "label_level": "2"
  },
  "0841-ag405": {
   "segment": "20260221022814-auto_grown_20260220174252405",
   "surface_shape": [
    3760,
    4900
   ],
   "crop": [
    1024,
    1664,
    2496,
    3136
   ],
   "scroll": "PHerc0841",
   "voxel_um": 9.366,
   "labels": "ink-labels/2.403um-volume-20260319124803/20260918",
   "label_level": "2"
  }
 },
 "results": [
  {
   "segment": "w045",
   "window": "whole",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.8722,
   "auc_reversed": 0.4425,
   "ink_px": 149192,
   "row_score_as_stored": 79.8,
   "row_score_reversed": 8.1,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "whole",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed43",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.8866,
   "auc_reversed": 0.5024,
   "ink_px": 149192,
   "row_score_as_stored": 68.3,
   "row_score_reversed": 10.6,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "whole",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seeds 42+43 mean",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "row_score_as_stored": 95.9,
   "row_score_reversed": 9.5,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.9136,
   "auc_reversed": 0.3704,
   "ink_px": 78047,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed43",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.9098,
   "auc_reversed": 0.4347,
   "ink_px": 78047,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "mps (M1 Pro, villa PR #1865 6723ad158)",
   "auc_as_stored": 0.9136,
   "auc_reversed": 0.3704,
   "ink_px": 78047,
   "measured_by": "user's Mac",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed43",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "mps (M1 Pro, villa PR #1865 6723ad158)",
   "auc_as_stored": 0.9098,
   "auc_reversed": 0.4347,
   "ink_px": 78047,
   "measured_by": "user's Mac",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "v8in d89166b",
   "settings": "fp32, stride 21 both directions, batch 4",
   "device": "mps (M1 Pro)",
   "auc_as_stored": 0.7382,
   "auc_reversed": 0.3269,
   "ink_px": 73200,
   "measured_by": "user's Mac",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.9074,
   "auc_reversed": 0.3835,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop and full (identical)",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.9199,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "w045",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.86,
   "auc_reversed": 0.4004,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "whole",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.7479,
   "auc_reversed": 0.501,
   "row_score_as_stored": 13.7,
   "row_score_reversed": 23.6,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "whole",
   "inner_px": 0,
   "map_from": "AutoPipeline render, from the released scan and the published segment",
   "reader": "ink_9um seed42",
   "settings": "AutoPipeline package defaults, ink on one L4 on Modal, 9 October 2026",
   "device": "cuda",
   "auc_as_stored": 0.7479,
   "auc_reversed": 0.5012,
   "auc_shuffled": 0.4851,
   "measured_by": "this repo",
   "source": "https://github.com/ChaseHendrick/Scrolls/blob/faada83287d39bbd4c7fd88c7ab0e75f3f1ff284/submissions/progress-2026-10/autopipeline/README.md",
   "date": "2026-10-09"
  },
  {
   "segment": "0841-ag896",
   "window": "whole",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.7195,
   "auc_reversed": 0.5697,
   "row_score_as_stored": 17.5,
   "row_score_reversed": 8.8,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag405",
   "window": "whole",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.7508,
   "auc_reversed": 0.6012,
   "row_score_as_stored": 46.4,
   "row_score_reversed": 7.5,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.7696,
   "auc_reversed": 0.5143,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag896",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.6553,
   "auc_reversed": 0.5397,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag405",
   "window": "crop",
   "inner_px": 0,
   "map_from": "full",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.793,
   "auc_reversed": 0.6527,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.7704,
   "auc_reversed": 0.4958,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag896",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.6604,
   "auc_reversed": 0.5575,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag405",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.8212,
   "auc_reversed": 0.6711,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.8755,
   "auc_reversed": 0.5573,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag896",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.8049,
   "auc_reversed": 0.5129,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag405",
   "window": "crop",
   "inner_px": 0,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.8308,
   "auc_reversed": 0.5883,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.8061,
   "auc_reversed": 0.5361,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag896",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.6594,
   "auc_reversed": 0.5579,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag405",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "ink_9um seed42",
   "settings": "step-075000, overlap 0.5, hann, villa e0bbb8b",
   "device": "cpu",
   "auc_as_stored": 0.7838,
   "auc_reversed": 0.6423,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.8994,
   "auc_reversed": 0.5855,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag896",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.823,
   "auc_reversed": 0.5323,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-ag405",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "d9v2",
   "settings": "TAUIL v1.0 d9v2_ft-012000, overlap 0.5, hann",
   "device": "cpu",
   "auc_as_stored": 0.8319,
   "auc_reversed": 0.6572,
   "measured_by": "this repo",
   "source": "docs/logs/2026-10-07-community-scan.md"
  },
  {
   "segment": "0841-w00",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "Reader v2",
   "settings": "domenicor046/reader-v2 reader-v2-step040000.pth (sha256 654ec5ac...d5d3a6), default window, overlap 0.5, hann, villa e0bbb8b40 + PR #1865 (6723ad158)",
   "device": "cpu",
   "auc_as_stored": 0.9269,
   "auc_reversed": 0.7294,
   "auc_shuffled": 0.5681,
   "measured_by": "this repo",
   "source": "scripts/experiments/2026-10-07-cloud/tricks/results.json"
  },
  {
   "segment": "0841-ag896",
   "window": "crop",
   "inner_px": 64,
   "map_from": "bare-crop",
   "reader": "Reader v2",
   "settings": "domenicor046/reader-v2 reader-v2-step040000.pth (sha256 654ec5ac...d5d3a6), default window, overlap 0.5, hann, villa e0bbb8b40 + PR #1865 (6723ad158)",
   "device": "cpu",
   "auc_as_stored": 0.7972,
   "auc_reversed": 0.5988,
   "auc_shuffled": 0.617,
   "measured_by": "this repo",
   "source": "scripts/experiments/2026-10-07-cloud/tricks/results.json"
  }
 ],
 "finetune_checks": {
  "about": "Our own fine-tunes of released readers, scored like the readers in results but kept out of that list so kit gate ranks released readers only. Model output.",
  "results": [
   {
    "segment": "0841-w00",
    "window": "crop",
    "inner_px": 64,
    "map_from": "bare-crop",
    "reader": "d9v2 fine-tune, arm A step 1500",
    "settings": "fine-tuned from d9v2_ft-012000 on 20260918 human labels of nine public 9.362 um segments (no PHerc0841, no w045): anchor 1.0, depth jitter 5, lr 2e-5, 1500 steps, batch 16, bf16, one seed; best of the 24 sweep checkpoints by selection mean (28 checkpoints tried with two earlier CPU runs); inference overlap 0.5, hann, villa PR #1865 6723ad158",
    "device": "cuda (Modal B200)",
    "auc_as_stored": 0.8838,
    "auc_reversed": 0.5458,
    "measured_by": "this repo",
    "source": "docs/logs/2026-10-08-d9v2-finetune-sweep.md"
   },
   {
    "segment": "0841-ag896",
    "window": "crop",
    "inner_px": 64,
    "map_from": "bare-crop",
    "reader": "d9v2 fine-tune, arm A step 1500",
    "settings": "fine-tuned from d9v2_ft-012000 on 20260918 human labels of nine public 9.362 um segments (no PHerc0841, no w045): anchor 1.0, depth jitter 5, lr 2e-5, 1500 steps, batch 16, bf16, one seed; best of the 24 sweep checkpoints by selection mean (28 checkpoints tried with two earlier CPU runs); inference overlap 0.5, hann, villa PR #1865 6723ad158",
    "device": "cuda (Modal B200)",
    "auc_as_stored": 0.8394,
    "auc_reversed": 0.5236,
    "measured_by": "this repo",
    "source": "docs/logs/2026-10-08-d9v2-finetune-sweep.md"
   }
  ]
 },
 "community": [
  {
   "reader": "Hecate",
   "test": "PHerc0841 whole, human labels",
   "auc": 0.855,
   "source": "https://huggingface.co/domenicor046/reader-v2",
   "note": "PHerc0841 is not held out from Hecate: its base model's committed trainer lists all three segments, although Hecate's own published split rule, applied to the published masks, makes them validation-only (see held_out.PHerc0841)"
  },
  {
   "reader": "Reader v2",
   "test": "PHerc0841 whole, human labels",
   "auc": 0.824,
   "source": "https://huggingface.co/domenicor046/reader-v2"
  },
  {
   "reader": "ink_9um",
   "test": "PHerc0841 whole, human labels",
   "auc": 0.733,
   "source": "https://huggingface.co/domenicor046/reader-v2"
  },
  {
   "reader": "Hecate + Reader v2 mean",
   "test": "PHerc0841 whole, human labels",
   "auc": 0.866,
   "source": "https://huggingface.co/domenicor046/reader-v2",
   "note": "PHerc0841 is not held out from Hecate: its base model's committed trainer lists all three segments, although Hecate's own published split rule, applied to the published masks, makes them validation-only (see held_out.PHerc0841)"
  },
  {
   "reader": "ink_9um seed42",
   "test": "PHerc0841, TAUIL held-out benchmark",
   "auc": 0.736,
   "source": "https://github.com/TAUIL-Abd-Elilah/pherc0826-first-letters-search"
  },
  {
   "reader": "d9v2",
   "test": "PHerc0841, TAUIL held-out benchmark",
   "auc": 0.828,
   "source": "https://github.com/TAUIL-Abd-Elilah/pherc0826-first-letters-search"
  },
  {
   "reader": "Reader v2",
   "test": "PHerc0841, TAUIL held-out benchmark",
   "auc": 0.826,
   "source": "https://github.com/TAUIL-Abd-Elilah/pherc0826-first-letters-search"
  },
  {
   "reader": "Reader v2 + d9v2 mean",
   "test": "PHerc0841, TAUIL held-out benchmark",
   "auc": 0.84,
   "source": "https://github.com/TAUIL-Abd-Elilah/pherc0826-first-letters-search"
  },
  {
   "reader": "v8in d89166b",
   "test": "PHerc0841 w00, labels' bounding box, human labels",
   "auc": 0.837,
   "auc_reversed": 0.562,
   "source": "https://github.com/Bullo27/v8in-12gb",
   "date": "2026-10-01"
  },
  {
   "reader": "v8in d89166b",
   "test": "PHerc0841 ag896, labels' bounding box, human labels",
   "auc": 0.807,
   "auc_reversed": 0.618,
   "source": "https://github.com/Bullo27/v8in-12gb",
   "date": "2026-10-01"
  },
  {
   "reader": "v8in d89166b",
   "test": "PHerc0841 ag405, labels' bounding box, human labels",
   "auc": 0.81,
   "auc_reversed": 0.586,
   "source": "https://github.com/Bullo27/v8in-12gb",
   "date": "2026-10-01"
  }
 ],
 "device_checks": [
  {
   "what": "ink_9um seed42, CPU vs MPS, w035",
   "machine": "Apple M1 Pro",
   "verdict": "pass",
   "max_abs_diff": 1,
   "pearson": 0.99999999,
   "source": "docs/logs/2026-10-07-w035-cpu.md"
  },
  {
   "what": "v8in, CPU vs MPS, w045 crop, stride 64",
   "machine": "Apple M1 Pro",
   "verdict": "pass",
   "max_abs_diff": 1,
   "pearson": 0.99999998,
   "source": "docs/logs/2026-10-07-community-scan.md"
  }
 ],
 "pending": [
  "v8in on 0841-w00, ag896, ag405 (crop, inner 64) on the user's Mac: a reproduction of Bullo27's v8in-12gb (2026-10-01) and the device check",
  "v8in-1447 (Youssef's PHerc1447 fine-tune) on the PHerc0841 crops: w00 scored 2026-10-07 (forward 0.8381 at stride 42, reverse 0.5932 at stride 64, so a matched-stride rerun is still needed; PR #7); ag405 interrupted at 172/196 tiles (PR #9); ag896 has scripts only, no saved result (PR #10)",
  "Reader v2 and the tricks table on PHerc0841 ag405, plus Reader v2 depth windows (w00 and ag896 scored 2026-10-07 in scripts/experiments/2026-10-07-cloud/tricks/, PR #8; that job stopped early before ag405, and w045 was dropped)"
 ],
 "surfacefix_checks": [
  {
   "date": "2026-10-07",
   "source": "docs/logs/2026-10-07-real-surfacefix-validation.md",
   "scope": "Published real PHerc0841 model-output consistency, not labelled accuracy or a reading",
   "reader": "d9v2_ft-012000@sha256:50d2ad0ef7690a6a422640804d38f01409da8bc96cfc1ab72f45324a4a18f966",
   "voxel_um": 9.366,
   "stride": 42,
   "maps_completed": 12,
   "patches_per_map": 196,
   "thresholds": {
    "region_points": 8,
    "min_points": 24,
    "min_corr": 0.5,
    "min_gain": 0.1,
    "control_margin": 0.1,
    "max_step_vox": 1,
    "max_shift_um": 50,
    "high_pass_um": 48.0,
    "spatial_displacement_um": [
     1000,
     3000
    ],
    "map_sampling": "bilinear at native row/scale_y-.5, col/scale_x-.5",
    "normal_tolerance_vox": 0.0033829117335329633
   },
   "normal": {
    "status": "no_corrections",
    "accepted_regions": 0,
    "flagged_regions": 16,
    "regions_examined": 16,
    "evaluable_regions": 4,
    "candidate_selection_attempted_regions": 4,
    "selection_points": 215,
    "held_back_points": 217,
    "covered_points": 432,
    "native_valid_points": 962
   },
   "supervised_pixels": 0,
   "label_accuracy_improvement_established": false,
   "nulls": [
    {
     "axis": 0,
     "shift_px": -300,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 2,
     "candidate_selection_attempted_regions": 1,
     "selection_points": 113,
     "held_back_points": 126,
     "covered_points": 239,
     "native_valid_points": 962
    },
    {
     "axis": 0,
     "shift_px": -150,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 4,
     "candidate_selection_attempted_regions": 3,
     "selection_points": 179,
     "held_back_points": 186,
     "covered_points": 365,
     "native_valid_points": 962
    },
    {
     "axis": 0,
     "shift_px": 150,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 4,
     "candidate_selection_attempted_regions": 3,
     "selection_points": 190,
     "held_back_points": 198,
     "covered_points": 388,
     "native_valid_points": 962
    },
    {
     "axis": 0,
     "shift_px": 300,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 2,
     "candidate_selection_attempted_regions": 1,
     "selection_points": 102,
     "held_back_points": 112,
     "covered_points": 214,
     "native_valid_points": 962
    },
    {
     "axis": 1,
     "shift_px": -300,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 2,
     "candidate_selection_attempted_regions": 2,
     "selection_points": 144,
     "held_back_points": 143,
     "covered_points": 287,
     "native_valid_points": 962
    },
    {
     "axis": 1,
     "shift_px": -150,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 4,
     "candidate_selection_attempted_regions": 4,
     "selection_points": 210,
     "held_back_points": 212,
     "covered_points": 422,
     "native_valid_points": 962
    },
    {
     "axis": 1,
     "shift_px": 150,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 4,
     "candidate_selection_attempted_regions": 3,
     "selection_points": 181,
     "held_back_points": 187,
     "covered_points": 368,
     "native_valid_points": 962
    },
    {
     "axis": 1,
     "shift_px": 300,
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 2,
     "candidate_selection_attempted_regions": 1,
     "selection_points": 83,
     "held_back_points": 87,
     "covered_points": 170,
     "native_valid_points": 962
    }
   ],
   "source_hashes_unchanged": true,
   "tangential_geometry_negative": {
    "refused": true,
    "changed_vox": 0.5,
    "changed_axis": "x",
    "native_node": [
     8,
     12
    ],
    "candidate_hashes_recomputed": true,
    "error": "candidate is not the declared bounded normal offset, or changes holes",
    "output_created": false
   },
   "regions": [
    {
     "id": 5,
     "box": [
      8,
      16,
      8,
      16
     ],
     "points": 64,
     "selection_points": 32,
     "held_back_points": 32,
     "accepted": false,
     "offset_vox": 0,
     "flagged": true,
     "reasons": [
      "weak baseline agreement",
      "selection: insufficient correlation or gain",
      "selection: controls agree or are inadequate",
      "held_back: insufficient correlation or gain",
      "held_back: controls agree or are inadequate"
     ],
     "baseline": {
      "selection": {
       "r": 0.4387141246866009,
       "control_max_abs": 0.38148932872392544,
       "spatial_controls": 4,
       "reason": null
      },
      "held_back": {
       "r": 0.5062978447824745,
       "control_max_abs": 0.5868860539781303,
       "spatial_controls": 5,
       "reason": null
      }
     },
     "winner": {
      "offset_vox": 1,
      "selection": {
       "r": 0.4282174152109277,
       "control_max_abs": 0.3344407065358476,
       "spatial_controls": 4,
       "reason": null
      },
      "held_back": {
       "r": 0.5561387038504649,
       "control_max_abs": 0.4677618348637407,
       "spatial_controls": 5,
       "reason": null
      }
     }
    },
    {
     "id": 6,
     "box": [
      8,
      16,
      16,
      24
     ],
     "points": 64,
     "selection_points": 28,
     "held_back_points": 30,
     "accepted": false,
     "offset_vox": 0,
     "flagged": true,
     "reasons": [
      "weak baseline agreement",
      "selection: insufficient correlation or gain",
      "selection: controls agree or are inadequate",
      "held_back: insufficient correlation or gain",
      "held_back: controls agree or are inadequate"
     ],
     "baseline": {
      "selection": {
       "r": 0.5213564250260587,
       "control_max_abs": 0.526313375619026,
       "spatial_controls": 2,
       "reason": null
      },
      "held_back": {
       "r": 0.39032566789421713,
       "control_max_abs": 0.3898062849175623,
       "spatial_controls": 3,
       "reason": null
      }
     },
     "winner": {
      "offset_vox": -1,
      "selection": {
       "r": 0.6025462683077615,
       "control_max_abs": 0.5033948230516222,
       "spatial_controls": 2,
       "reason": null
      },
      "held_back": {
       "r": 0.39948311627463784,
       "control_max_abs": 0.4506269095299158,
       "spatial_controls": 3,
       "reason": null
      }
     }
    },
    {
     "id": 9,
     "box": [
      16,
      24,
      8,
      16
     ],
     "points": 64,
     "selection_points": 32,
     "held_back_points": 31,
     "accepted": false,
     "offset_vox": 0,
     "flagged": true,
     "reasons": [
      "weak baseline agreement",
      "selection: insufficient correlation or gain",
      "held_back: insufficient correlation or gain",
      "held_back: controls agree or are inadequate"
     ],
     "baseline": {
      "selection": {
       "r": 0.33672626445978776,
       "control_max_abs": 0.3745772900733683,
       "spatial_controls": 4,
       "reason": null
      },
      "held_back": {
       "r": 0.4417598107495779,
       "control_max_abs": 0.2640971966156233,
       "spatial_controls": 4,
       "reason": null
      }
     },
     "winner": {
      "offset_vox": -1,
      "selection": {
       "r": 0.4402678023078828,
       "control_max_abs": 0.33333888442481774,
       "spatial_controls": 4,
       "reason": null
      },
      "held_back": {
       "r": 0.3755952562813157,
       "control_max_abs": 0.4653327409002419,
       "spatial_controls": 4,
       "reason": null
      }
     }
    },
    {
     "id": 10,
     "box": [
      16,
      24,
      16,
      24
     ],
     "points": 64,
     "selection_points": 30,
     "held_back_points": 27,
     "accepted": false,
     "offset_vox": 0,
     "flagged": true,
     "reasons": [
      "weak baseline agreement",
      "selection: insufficient correlation or gain",
      "selection: controls agree or are inadequate",
      "held_back: insufficient correlation or gain",
      "held_back: controls agree or are inadequate"
     ],
     "baseline": {
      "selection": {
       "r": 0.17411655481782953,
       "control_max_abs": 0.7201273476411033,
       "spatial_controls": 3,
       "reason": null
      },
      "held_back": {
       "r": 0.40737491375056234,
       "control_max_abs": 0.43527380793743925,
       "spatial_controls": 3,
       "reason": null
      }
     },
     "winner": {
      "offset_vox": -1,
      "selection": {
       "r": 0.1568140611365692,
       "control_max_abs": 0.637058555529443,
       "spatial_controls": 3,
       "reason": null
      },
      "held_back": {
       "r": 0.3059372422041614,
       "control_max_abs": 0.5493950865719324,
       "spatial_controls": 3,
       "reason": null
      }
     }
    }
   ],
   "limits": "One same-sheet patch; no labels; dependent null tests; no accepted recovery or significance established."
  }
 ],
 "reconstruction_checks": [
  {
   "scope": "Public PHerc0139 w045 structural CT only; no model/label selection; no targets",
   "status": "registration_feasibility_complete_no_ink_inference",
   "rule_sha256": "37ed675e2edf44d9010cefbb434ecbffc504e8e9692c6dca358ee46934839dbd",
   "rows": [
    {
     "volume": "pag0",
     "patch": 0,
     "quality": {
      "source_nonzero_fraction": 0.9993963241577148,
      "source_std": 38.914131428100376,
      "search_nonzero_fraction": 0.9040805675365307,
      "search_std": 43.12134541019717,
      "source_clipped255_fraction": 0.0004611015319824219,
      "search_clipped255_fraction": 0.0007428769712094907
     },
     "gradient": {
      "shift_zyx_vox": [
       1,
       0,
       0
      ],
      "ncc": 0.8478612450707881,
      "competing_peak_ncc": 0.6595112722209818,
      "peak_margin": 0.18834997284980626,
      "boundary": false
     },
     "raw": {
      "shift_zyx_vox": [
       1,
       0,
       0
      ],
      "ncc": 0.762514640669362,
      "competing_peak_ncc": 0.591498588785872,
      "peak_margin": 0.17101605188349,
      "boundary": false
     }
    },
    {
     "volume": "pag0",
     "patch": 1,
     "quality": {
      "source_nonzero_fraction": 0.9999194145202637,
      "source_std": 41.17582922325454,
      "search_nonzero_fraction": 0.9266023989076968,
      "search_std": 40.592275127529206,
      "source_clipped255_fraction": 0.00018024444580078125,
      "search_clipped255_fraction": 0.000354766845703125
     },
     "gradient": {
      "shift_zyx_vox": [
       -1,
       0,
       0
      ],
      "ncc": 0.8832125513174112,
      "competing_peak_ncc": 0.7204041198537716,
      "peak_margin": 0.16280843146363955,
      "boundary": false
     },
     "raw": {
      "shift_zyx_vox": [
       -1,
       0,
       0
      ],
      "ncc": 0.8384771281038401,
      "competing_peak_ncc": 0.6661783326720139,
      "peak_margin": 0.17229879543182625,
      "boundary": false
     }
    },
    {
     "volume": "pag50",
     "patch": 0,
     "quality": {
      "source_nonzero_fraction": 0.9993963241577148,
      "source_std": 38.914131428100376,
      "search_nonzero_fraction": 0.8872364185474537,
      "search_std": 39.60397934631725,
      "source_clipped255_fraction": 0.0004611015319824219,
      "search_clipped255_fraction": 0.0
     },
     "gradient": {
      "shift_zyx_vox": [
       1,
       0,
       0
      ],
      "ncc": 0.7983506592494876,
      "competing_peak_ncc": 0.6219411850068278,
      "peak_margin": 0.17640947424265974,
      "boundary": false
     },
     "raw": {
      "shift_zyx_vox": [
       1,
       0,
       0
      ],
      "ncc": 0.6980895540264925,
      "competing_peak_ncc": 0.5223580158852901,
      "peak_margin": 0.17573153814120246,
      "boundary": false
     }
    },
    {
     "volume": "pag50",
     "patch": 1,
     "quality": {
      "source_nonzero_fraction": 0.9999194145202637,
      "source_std": 41.17582922325454,
      "search_nonzero_fraction": 0.9102993717900029,
      "search_std": 37.002759645052336,
      "source_clipped255_fraction": 0.00018024444580078125,
      "search_clipped255_fraction": 0.0
     },
     "gradient": {
      "shift_zyx_vox": [
       -1,
       0,
       0
      ],
      "ncc": 0.823588006720883,
      "competing_peak_ncc": 0.6676792078670158,
      "peak_margin": 0.1559087988538671,
      "boundary": false
     },
     "raw": {
      "shift_zyx_vox": [
       -1,
       0,
       0
      ],
      "ncc": 0.7702961506823381,
      "competing_peak_ncc": 0.5865054235423434,
      "peak_margin": 0.18379072713999478,
      "boundary": false
     }
    }
   ],
   "controls": [
    {
     "shift_zyx_vox": [
      0,
      0,
      0
     ],
     "ncc": 0.9956367610729221,
     "competing_peak_ncc": 0.7691388989438767,
     "peak_margin": 0.22649786212904544,
     "boundary": false,
     "patch": 0,
     "control": "self",
     "expected_zyx_vox": [
      0,
      0,
      0
     ],
     "pass": true
    },
    {
     "shift_zyx_vox": [
      3,
      -4,
      5
     ],
     "ncc": 0.9956367610729228,
     "competing_peak_ncc": 0.7691387223981794,
     "peak_margin": 0.22649803867474338,
     "boundary": false,
     "patch": 0,
     "control": "known_shift",
     "expected_zyx_vox": [
      3,
      -4,
      5
     ],
     "pass": true
    },
    {
     "shift_zyx_vox": [
      0,
      0,
      0
     ],
     "ncc": 0.9950509149877048,
     "competing_peak_ncc": 0.8424505818821971,
     "peak_margin": 0.1526003331055077,
     "boundary": false,
     "patch": 1,
     "control": "self",
     "expected_zyx_vox": [
      0,
      0,
      0
     ],
     "pass": true
    },
    {
     "shift_zyx_vox": [
      3,
      -4,
      5
     ],
     "ncc": 0.9950509149877053,
     "competing_peak_ncc": 0.842450665446233,
     "peak_margin": 0.15260024954147222,
     "boundary": false,
     "patch": 1,
     "control": "known_shift",
     "expected_zyx_vox": [
      3,
      -4,
      5
     ],
     "pass": true
    }
   ],
   "verdict": {
    "pag0": {
     "pass": false,
     "fit_shift_zyx_vox": [
      1,
      0,
      0
     ],
     "heldback_residual_zyx_vox": [
      -2,
      0,
      0
     ],
     "reasons": [
      "heldback residual exceeds0.5voxel"
     ]
    },
    "pag50": {
     "pass": false,
     "fit_shift_zyx_vox": [
      1,
      0,
      0
     ],
     "heldback_residual_zyx_vox": [
      -2,
      0,
      0
     ],
     "reasons": [
      "heldback residual exceeds0.5voxel"
     ]
    }
   },
   "all_pass": false,
   "actual_chunks": 82,
   "actual_download_bytes": 171966464,
   "peak_rss_bytes": 974397440,
   "limits": "Two local structural cubes only; does not prove global registration, ink identity or geometry correction quality.",
   "source": "docs/logs/2026-10-07-reconstruction-tests.md",
   "date": "2026-10-07",
   "voxel_um": 9.362
  }
 ],
 "followup_checks": [
  {
   "date": "2026-10-07",
   "kind": "reconstruction_transfer",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/physics/result.json",
   "evidence_sha256": "12a98ba9b6a88208dfc2d14f9ac700191a351d673595cfce239da5b51392c03c",
   "outcome": "TRANSFER_PREDICTABLE",
   "primary": {
    "rms_uint8_units": 2.7172861665532877,
    "target_std_uint8_units": 35.81703106544626,
    "normalized_rms": 0.07586575675655968,
    "correlation": 0.9972415686653727,
    "rms_reduction_vs_affine": 0.5575069048626593
   },
   "controls": {
    "self": {
     "rms_uint8_units": 6.508392482636255e-06,
     "target_std_uint8_units": 39.69610406220731,
     "normalized_rms": 1.6395544692338137e-07,
     "correlation": 0.9999999999999892,
     "passed": true
    },
    "misregistered": {
     "rms_uint8_units": 50.41781355615378,
     "target_std_uint8_units": 35.021058936088636,
     "normalized_rms": 1.4396427489004062,
     "correlation": -0.015493771731763854,
     "correlation_drop": 1.0127353403971366,
     "passed": true
    },
    "depth_shuffle": {
     "rms_uint8_units": 48.03849028313286,
     "target_std_uint8_units": 36.23266649364926,
     "normalized_rms": 1.3258337001377716,
     "correlation": 0.10851748810500667,
     "correlation_drop": 0.888724080560366,
     "passed": true
    }
   },
   "rule_sha256": "f506cac999a1189e4eff33fc014964e56800bbd49bdd9ea2a8c20a5218fb9f49"
  },
  {
   "date": "2026-10-07",
   "kind": "transfer_robustness_and_failed_transport",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/physics/supplement-result.json",
   "evidence_sha256": "d87cfef01138674ff3c46661bd76fc9c041bc355f8995689d49d0cc9420f9ba5",
   "readout": {
    "supplement_rule_sha256": "c4bb51f47aec5a8898cb4db14a6864e9bb959345eb4962c53361c44f7f81b4a1",
    "boundary": [
     {
      "rms_uint8_units": 2.7104543482052783,
      "target_std_uint8_units": 35.81703106544626,
      "normalized_rms": 0.0756750145832197,
      "correlation": 0.9972561272523617,
      "reflection_pad_vox": 32,
      "correlation_drop": -1.4558586988955291e-05,
      "normalized_rms_increase": -0.00019074217333998889,
      "pass": true
     },
     {
      "rms_uint8_units": 2.7070283505277093,
      "target_std_uint8_units": 35.81703106544626,
      "normalized_rms": 0.07557936182877142,
      "correlation": 0.9972636976977209,
      "reflection_pad_vox": 64,
      "correlation_drop": -2.212903234821706e-05,
      "normalized_rms_increase": -0.00028639492778825826,
      "pass": true
     }
    ],
    "boundary_all_pass": true,
    "transportability": {
     "local_original_to_pag50_shift_zyx": [
      [
       1,
       0,
       0
      ],
      [
       -1,
       0,
       0
      ]
     ],
     "unaltered_original_vs_pag50": {
      "rms_uint8_units": 37.97422313340617,
      "target_std_uint8_units": 35.84023378229306,
      "normalized_rms": 1.0595417252040193,
      "correlation": 0.7700780778788229
     },
     "reuse_pag0_to_pag50_response": {
      "rms_uint8_units": 26.202700154385873,
      "target_std_uint8_units": 35.84023378229306,
      "normalized_rms": 0.7310973559366504,
      "correlation": 0.826032116069194,
      "meets_main_prediction_gate": false
     },
     "new_original_to_pag50_response": {
      "rms_uint8_units": 14.877056321028931,
      "target_std_uint8_units": 35.84023378229306,
      "normalized_rms": 0.41509372989578464,
      "correlation": 0.9162328134819623,
      "meets_main_prediction_gate": false
     },
     "new_original_to_pag50_affine_baseline": {
      "rms_uint8_units": 23.144581538942198,
      "target_std_uint8_units": 35.84023378229306,
      "normalized_rms": 0.6457709422190439,
      "correlation": 0.770078076098251
     },
     "new_radial_response": [
      0.35731608267892945,
      0.4246066960891954,
      0.5342380065149127,
      0.7843969553108562,
      1.1336330013153784,
      1.4427832619340673,
      1.714579104322068,
      1.8916182713273852,
      2.084863990522495,
      2.2828395536613195,
      2.215548085772253,
      2.378324195170307,
      2.520417370676034,
      2.8039534305234186,
      2.860778081951922,
      2.9432688071361297,
      3.0203800875165943,
      2.9382258498162552,
      3.0525854977984785,
      2.375778460802748,
      1.8853187875719328,
      1.6242616464737876,
      1.3057280032939365,
      1.247538787475447,
      1.019512066666303,
      0.8524376127482227,
      0.37035551507980324,
      0.33297207359333336,
      0.12454588304290581,
      0.42633484117234016,
      0.6152649877899703,
      1.3462497734949692
     ]
    },
    "limits": "Local shift audits only; global original→phase translation gate failed. No mesh/render/label inference permitted from these diagnostics.",
    "elapsed_seconds": 2.09436192999965,
    "peak_rss_bytes": 839704576
   }
  },
  {
   "date": "2026-10-07",
   "kind": "published_ct_orientation",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/orientation-pherc0841/orientation-result.json",
   "evidence_sha256": "924543b3452ac0e745b67a843b25e75ef1d8470dd49655dc7596ebae4d08d291",
   "existing_canonical_negative_normal": {
    "correlation": 0.9999982592883446,
    "mean_absolute_error_uint8": 0.005617822919573102,
    "rms_error_uint8": 0.07495213752504395,
    "equal_fraction": 0.9943821770804269,
    "voxels": 7340032,
    "passes_frozen_reproduction_bar": true
   },
   "existing_order_reversed_positive_normal": {
    "correlation": 0.09173207985689143,
    "mean_absolute_error_uint8": 42.459928376334055,
    "rms_error_uint8": 54.14116455136405,
    "equal_fraction": 0.01311847141810826,
    "voxels": 7340032,
    "passes_frozen_reproduction_bar": false
   }
  },
  {
   "date": "2026-10-07",
   "kind": "published_ct_orientation",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/orientation-w045/orientation-result.json",
   "evidence_sha256": "1e1e94e20b36bc1c57d22cc29a88b374d3d158bc08e08804bbcf601482d5fac2",
   "official_raw_positive_normal": {
    "correlation": 0.01403328481568698,
    "mean_absolute_error_uint8": 39.806175640651155,
    "rms_error_uint8": 52.43670110290429,
    "equal_fraction": 0.013877732413155692,
    "voxels": 7340032,
    "passes_frozen_reproduction_bar": false
   },
   "official_raw_reversed_negative_normal": {
    "correlation": 0.9999988669466128,
    "mean_absolute_error_uint8": 0.0031597954886300223,
    "rms_error_uint8": 0.05621205821378561,
    "equal_fraction": 0.99684020451137,
    "voxels": 7340032,
    "passes_frozen_reproduction_bar": true
   }
  },
  {
   "date": "2026-10-07",
   "kind": "labelled_reader_baseline",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/w045-baseline/original-baseline-results.json",
   "evidence_sha256": "967bd813497b15c7369a6d22666822c2bb512278e97c98ecbbc590b0706dd4fd",
   "rows": {
    "original/forward": {
     "auc": {
      "auc": 0.9255,
      "ink_px": 46757,
      "background_px": 123909,
      "zero_px_in_mask": 0,
      "zero_px_counted": true,
      "inner": 64
     },
     "auc_bootstrap": {
      "finite_draws": 1000,
      "finite_fraction": 1.0,
      "ci95": [
       0.883332675185323,
       0.9616414943174794
      ]
     },
     "hp": {
      "hp_r": 0.0177,
      "raw_r": 0.6626,
      "null_max_abs": 0.0578,
      "px": 95772,
      "sigma_px": 5.127,
      "inner": 0
     },
     "hp_bootstrap": {
      "finite_draws": 1000,
      "finite_fraction": 1.0,
      "ci95": [
       -0.006043657395329331,
       0.03061072107694212
      ]
     },
     "zero_fraction_primary": 0.0
    },
    "original/reverse": {
     "auc": {
      "auc": 0.3346,
      "ink_px": 46757,
      "background_px": 123909,
      "zero_px_in_mask": 0,
      "zero_px_counted": true,
      "inner": 64
     },
     "auc_bootstrap": {
      "finite_draws": 1000,
      "finite_fraction": 1.0,
      "ci95": [
       0.204900879550216,
       0.49254205363596004
      ]
     },
     "hp": {
      "hp_r": 0.0214,
      "raw_r": -0.2676,
      "null_max_abs": 0.0609,
      "px": 95772,
      "sigma_px": 5.127,
      "inner": 0
     },
     "hp_bootstrap": {
      "finite_draws": 1000,
      "finite_fraction": 1.0,
      "ci95": [
       0.0008470377876892148,
       0.050421117358723656
      ]
     },
     "zero_fraction_primary": 0.0
    },
    "original/shuffle": {
     "auc": {
      "auc": 0.5814,
      "ink_px": 46757,
      "background_px": 123909,
      "zero_px_in_mask": 0,
      "zero_px_counted": true,
      "inner": 64
     },
     "auc_bootstrap": {
      "finite_draws": 1000,
      "finite_fraction": 1.0,
      "ci95": [
       0.47171973457556776,
       0.696060708514822
      ]
     },
     "hp": {
      "hp_r": 0.0046,
      "raw_r": -0.0239,
      "null_max_abs": 0.0237,
      "px": 95772,
      "sigma_px": 5.127,
      "inner": 0
     },
     "hp_bootstrap": {
      "finite_draws": 1000,
      "finite_fraction": 1.0,
      "ci95": [
       -0.006376258300949498,
       0.015892512628444096
      ]
     },
     "zero_fraction_primary": 0.0
    }
   },
   "limits": "One fixed known labelled crop; AUC control advantage without passing high-pass detail controls; not a new reading or improvement."
  },
  {
   "date": "2026-10-07",
   "kind": "published_reader_baseline",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/w045-baseline/published-baseline-results.json",
   "evidence_sha256": "5a8d4ec9e276fb19614872d865fa5d16e2234be63195b9172427ec39250a94bf",
   "rows": {
    "forward": {
     "auc": {
      "auc": 0.9258,
      "ink_px": 46757,
      "background_px": 123909,
      "zero_px_in_mask": 0,
      "zero_px_counted": true,
      "inner": 64
     },
     "hp": {
      "hp_r": 0.0179,
      "raw_r": 0.6632,
      "null_max_abs": 0.0571,
      "px": 95772,
      "sigma_px": 5.127,
      "inner": 0
     },
     "zero_fraction_primary": 0.0
    },
    "reverse": {
     "auc": {
      "auc": 0.3356,
      "ink_px": 46757,
      "background_px": 123909,
      "zero_px_in_mask": 0,
      "zero_px_counted": true,
      "inner": 64
     },
     "hp": {
      "hp_r": 0.0221,
      "raw_r": -0.2666,
      "null_max_abs": 0.0578,
      "px": 95772,
      "sigma_px": 5.127,
      "inner": 0
     },
     "zero_fraction_primary": 0.0
    },
    "shuffle": {
     "auc": {
      "auc": 0.5814,
      "ink_px": 46757,
      "background_px": 123909,
      "zero_px_in_mask": 0,
      "zero_px_counted": true,
      "inner": 64
     },
     "hp": {
      "hp_r": 0.0041,
      "raw_r": -0.0236,
      "null_max_abs": 0.0231,
      "px": 95772,
      "sigma_px": 5.127,
      "inner": 0
     },
     "zero_fraction_primary": 0.0
    }
   },
   "limits": "One fixed known labelled crop; AUC control advantage without passing high-pass detail controls; not a new reading or improvement."
  },
  {
   "date": "2026-10-07",
   "kind": "continuous_geometry_coverage",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/surfacefix-design/certified-crop-geometry.json",
   "evidence_sha256": "c098dbee77c0cf626a5d08a072684adc062228983c722022cf943564f245b9ad",
   "geometrically_matched_native": 479,
   "ambiguous_native": 0,
   "unresolved_native": 0,
   "geometric_and_canvas_covered_native": 444,
   "evaluable_regions_before_map_coverage": 4,
   "limits": "Native label samples are geometric eligibility counts only, not final prediction coverage. Interleaved subsets are correlated local checks, not statistical independence. Labels reused from one source sheet are not independent ag896 annotations."
  },
  {
   "date": "2026-10-07",
   "kind": "subpixel_estimator_control_failure",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/affine-v2/result.json",
   "evidence_sha256": "ab99bcb7c487ea04991dcfd668ff3ee4bca3a20101489c5e17d94d6664dbc3d2",
   "readout": {
    "status": "stopped_algorithm_controls_failed",
    "controls": [
     {
      "cube": 0,
      "kind": "gradient",
      "expected_zyx": [
       0,
       0,
       0
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.01781515758244024,
        -0.03378773519965352,
        -0.010231970851091142
       ],
       "ncc": 0.9999995695456397,
       "margin": 0.21161636102686487,
       "competing_ncc": 0.7883832085187749,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.039543445943754855,
      "pass": true
     },
     {
      "cube": 0,
      "kind": "raw",
      "expected_zyx": [
       0,
       0,
       0
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.008761772429173777,
        -0.013249488336246689,
        -0.008880485323477905
       ],
       "ncc": 1.0000007087635323,
       "margin": 0.16879384568667766,
       "competing_ncc": 0.8312068630768547,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.01819836852175143,
      "pass": true
     },
     {
      "cube": 0,
      "kind": "gradient",
      "expected_zyx": [
       3,
       -4,
       5
      ],
      "peak": {
       "integer_shift_zyx": [
        3,
        -4,
        5
       ],
       "shift_zyx": [
        3.0178150020741206,
        -4.033788228873377,
        4.989768024039355
       ],
       "ncc": 0.9999996960720139,
       "margin": 0.21161648755323903,
       "competing_ncc": 0.7883832085187749,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.03954379902538217,
      "pass": true
     },
     {
      "cube": 0,
      "kind": "raw",
      "expected_zyx": [
       3,
       -4,
       5
      ],
      "peak": {
       "integer_shift_zyx": [
        3,
        -4,
        5
       ],
       "shift_zyx": [
        3.0087614346530493,
        -4.013249525718723,
        4.991119555083712
       ],
       "ncc": 1.0000007087635323,
       "margin": 0.1687937712495644,
       "competing_ncc": 0.8312069375139679,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.018198213397528393,
      "pass": true
     },
     {
      "cube": 0,
      "kind": "gradient",
      "expected_zyx": [
       0.25,
       -0.35,
       0.4
      ],
      "peak": {
       "integer_shift_zyx": [
        1,
        -1,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9565976114462643,
       "margin": 0.10861854159421525,
       "competing_ncc": 0.8479790698520491,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     },
     {
      "cube": 0,
      "kind": "raw",
      "expected_zyx": [
       0.25,
       -0.35,
       0.4
      ],
      "peak": {
       "integer_shift_zyx": [
        1,
        0,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9709418506939486,
       "margin": 0.10065253217739611,
       "competing_ncc": 0.8702893185165524,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     },
     {
      "cube": 0,
      "kind": "gradient",
      "expected_zyx": [
       -0.4,
       0.2,
       -0.3
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9733379846703814,
       "margin": 0.15727457072203377,
       "competing_ncc": 0.8160634139483476,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     },
     {
      "cube": 0,
      "kind": "raw",
      "expected_zyx": [
       -0.4,
       0.2,
       -0.3
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9831266402004422,
       "margin": 0.12583314422131275,
       "competing_ncc": 0.8572934959791294,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     },
     {
      "cube": 1,
      "kind": "gradient",
      "expected_zyx": [
       0,
       0,
       0
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.032535844078940175,
        0.013681845248974105,
        -0.0147306657171002
       ],
       "ncc": 0.9999996762333515,
       "margin": 0.13255914546140146,
       "competing_ncc": 0.86744053077195,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.03824613120061862,
      "pass": true
     },
     {
      "cube": 1,
      "kind": "raw",
      "expected_zyx": [
       0,
       0,
       0
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.010202307872980295,
        0.0037576501002064967,
        -0.004647328405867994
       ],
       "ncc": 0.9999999966478278,
       "margin": 0.09870172029394919,
       "competing_ncc": 0.9012982763538786,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.01182390297332675,
      "pass": true
     },
     {
      "cube": 1,
      "kind": "gradient",
      "expected_zyx": [
       3,
       -4,
       5
      ],
      "peak": {
       "integer_shift_zyx": [
        3,
        -4,
        5
       ],
       "shift_zyx": [
        3.0325356732049493,
        -3.986318888946053,
        4.98526965125798
       ],
       "ncc": 0.9999996762333491,
       "margin": 0.13255914546139902,
       "competing_ncc": 0.86744053077195,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.03824560111478477,
      "pass": true
     },
     {
      "cube": 1,
      "kind": "raw",
      "expected_zyx": [
       3,
       -4,
       5
      ],
      "peak": {
       "integer_shift_zyx": [
        3,
        -4,
        5
       ],
       "shift_zyx": [
        3.0102025656858853,
        -3.9962427782560743,
        4.995352726042864
       ],
       "ncc": 0.9999999273277793,
       "margin": 0.09870165097390071,
       "competing_ncc": 0.9012982763538786,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.011823967905931997,
      "pass": true
     },
     {
      "cube": 1,
      "kind": "gradient",
      "expected_zyx": [
       0.25,
       -0.35,
       0.4
      ],
      "peak": {
       "integer_shift_zyx": [
        1,
        0,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9805691757423648,
       "margin": 0.06856430948622927,
       "competing_ncc": 0.9120048662561355,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     },
     {
      "cube": 1,
      "kind": "raw",
      "expected_zyx": [
       0.25,
       -0.35,
       0.4
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.23208533869405373,
        -0.40147374256389934,
        0.4170871404523452
       ],
       "ncc": 0.9851188027710025,
       "margin": 0.07112694943367392,
       "competing_ncc": 0.9139918533373286,
       "refinement_error": null,
       "boundary": false
      },
      "error_vox": 0.05711787489113652,
      "pass": false
     },
     {
      "cube": 1,
      "kind": "gradient",
      "expected_zyx": [
       -0.4,
       0.2,
       -0.3
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9876768332571997,
       "margin": 0.08494287717780558,
       "competing_ncc": 0.9027339560793941,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     },
     {
      "cube": 1,
      "kind": "raw",
      "expected_zyx": [
       -0.4,
       0.2,
       -0.3
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": null,
       "ncc": 0.9909431701840109,
       "margin": 0.06804648406549885,
       "competing_ncc": 0.922896686118512,
       "refinement_error": "quadratic_offset_outside_halfvoxel",
       "boundary": false
      },
      "error_vox": Infinity,
      "pass": false
     }
    ],
    "all_pass": false
   }
  },
  {
   "date": "2026-10-07",
   "kind": "fresh_affine_registration_failure",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/affine-v3/result.json",
   "evidence_sha256": "3986f8b438af155f58b94ad47a5c3e6a3833705c8d6a3b421bfb7390d60a69fe",
   "readout": {
    "status": "affine_followup_completed_no_render_or_ink",
    "all_pass": false,
    "reasons": [
     "pag0cal4raw_gradient_disagree",
     "pag50cal4raw_gradient_disagree",
     "pag50cal5raw_gradient_disagree",
     "pag50cal6raw_gradient_disagree",
     "pag0audit8raw_fail",
     "pag50audit8gradient_fail",
     "pag50audit8raw_fail",
     "pag0audit10gradient_fail",
     "pag50audit10gradient_fail",
     "pag0audit11gradient_fail",
     "pag50audit11gradient_fail"
    ],
    "fits": {
     "pag0": {
      "matrix_original_xyz_to_phase_xyz": [
       [
        0.9994308543471703,
        -0.0003073651693106053,
        -0.0001039847834849211,
        4.2528366616661515
       ],
       [
        -0.0005285749031873115,
        1.0007866124107359,
        -0.00034767901714558453,
        1.7580674690293563
       ],
       [
        0.0003215116415617198,
        0.00023135703063864748,
        0.9987197816182064,
        5.9081799022078485
       ],
       [
        0.0,
        0.0,
        0.0,
        1.0
       ]
      ],
      "normalized_design_rank": 4,
      "normalized_design_condition": 1.0000000000000002,
      "singular_values": [
       1.0009085007326999,
       0.9993253223885015,
       0.9987035666530119
      ],
      "determinant": 0.9989364744876597,
      "calibration_residuals_euclidean_vox": [
       0.4339448747972224,
       0.2763975237016825,
       0.21837087842122738,
       0.09264673255467396,
       0.19346365711334923,
       0.027239770832756895,
       0.07443334329286848,
       0.2041183011158512
      ]
     },
     "pag50": {
      "matrix_original_xyz_to_phase_xyz": [
       [
        0.9993655952100916,
        -0.0003026453808888393,
        -0.00011560933525042425,
        4.594066387434395
       ],
       [
        -0.0005540530810734849,
        1.000808020701035,
        -0.00034888122447204343,
        1.7987250985856917
       ],
       [
        0.00034375903495374363,
        0.00023076206871407322,
        0.9987188840997457,
        5.819081630699657
       ],
       [
        0.0,
        0.0,
        0.0,
        1.0
       ]
      ],
      "normalized_design_rank": 4,
      "normalized_design_condition": 1.0000000000000002,
      "singular_values": [
       1.0009291133861908,
       0.9992642336286339,
       0.9986993056406788
      ],
      "determinant": 0.998891718448916,
      "calibration_residuals_euclidean_vox": [
       0.44383172833246554,
       0.2729365863872713,
       0.22858222814375137,
       0.09334666545106256,
       0.19702905088143574,
       0.030532579750315668,
       0.07156755331138423,
       0.21491787741865503
      ]
     }
    },
    "audits": [
     {
      "volume": "pag0",
      "point": 8,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.903274526906,
       3703.1099673092835,
       6286.162315555886
      ],
      "observed_xyz": [
       4494.2181985326215,
       3703.21335133561,
       6286.521740448827
      ],
      "residual_euclidean_vox": 0.48892907455015544,
      "pass": true
     },
     {
      "volume": "pag0",
      "point": 8,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.903274526906,
       3703.1099673092835,
       6286.162315555886
      ],
      "observed_xyz": [
       4494.3010529402,
       3703.3243121784385,
       6286.508166726996
      ],
      "residual_euclidean_vox": 0.5690205809754427,
      "pass": false
     },
     {
      "volume": "pag50",
      "point": 8,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.89563513477,
       3703.107843831143,
       6286.165352125232
      ],
      "observed_xyz": [
       4494.227371913699,
       3703.232507778052,
       6286.52420645206
      ],
      "residual_euclidean_vox": 0.5043479136832142,
      "pass": false
     },
     {
      "volume": "pag50",
      "point": 8,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.89563513477,
       3703.107843831143,
       6286.165352125232
      ],
      "observed_xyz": [
       4494.301324049438,
       3703.320558699832,
       6286.506096497365
      ],
      "residual_euclidean_vox": 0.57090965834104,
      "pass": false
     },
     {
      "volume": "pag0",
      "point": 9,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.784316456269,
       3953.2675058691316,
       6286.243946675022
      ],
      "observed_xyz": [
       4567.686451909591,
       3953.5283917465313,
       6286.448562589208
      ],
      "residual_euclidean_vox": 0.34569724161690407,
      "pass": true
     },
     {
      "volume": "pag0",
      "point": 9,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.784316456269,
       3953.2675058691316,
       6286.243946675022
      ],
      "observed_xyz": [
       4567.639795518848,
       3953.5067494956143,
       6286.470583628746
      ],
      "residual_euclidean_vox": 0.3598445816723486,
      "pass": true
     },
     {
      "volume": "pag50",
      "point": 9,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.773027835095,
       3953.268849078402,
       6286.248480810997
      ],
      "observed_xyz": [
       4567.673023206801,
       3953.5389901292597,
       6286.4564480303425
      ],
      "residual_euclidean_vox": 0.35528506492809697,
      "pass": true
     },
     {
      "volume": "pag50",
      "point": 9,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.773027835095,
       3953.268849078402,
       6286.248480810997
      ],
      "observed_xyz": [
       4567.635716867759,
       3953.507827424083,
       6286.467082641172
      ],
      "residual_euclidean_vox": 0.3517836147566103,
      "pass": true
     },
     {
      "volume": "pag0",
      "point": 10,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.780160005927,
       3953.1519032493375,
       6730.650457633648
      ],
      "observed_xyz": [
       4494.281065476323,
       3952.708058283991,
       6730.3029384600595
      ],
      "residual_euclidean_vox": 0.754104912825838,
      "pass": false
     },
     {
      "volume": "pag0",
      "point": 10,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.780160005927,
       3953.1519032493375,
       6730.650457633648
      ],
      "observed_xyz": [
       4493.727322680465,
       3953.397206190075,
       6730.462320007361
      ],
      "residual_euclidean_vox": 0.31362570385974364,
      "pass": true
     },
     {
      "volume": "pag50",
      "point": 10,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.768527635361,
       3953.1545968615114,
       6730.652946066797
      ],
      "observed_xyz": [
       4494.290454472577,
       3952.692159704353,
       6730.3085134420635
      ],
      "residual_euclidean_vox": 0.7777464758568348,
      "pass": false
     },
     {
      "volume": "pag50",
      "point": 10,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4493.768527635361,
       3953.1545968615114,
       6730.652946066797
      ],
      "observed_xyz": [
       4493.730396009987,
       3953.3885288511196,
       6730.45651661474
      ],
      "residual_euclidean_vox": 0.30783555066143686,
      "pass": true
     },
     {
      "volume": "pag0",
      "point": 11,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.814884519946,
       3702.916135603818,
       6730.616410237464
      ],
      "observed_xyz": [
       4568.16007365358,
       3703.1710765251482,
       6730.279471440594
      ],
      "residual_euclidean_vox": 0.5455989041269781,
      "pass": false
     },
     {
      "volume": "pag0",
      "point": 11,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.814884519946,
       3702.916135603818,
       6730.616410237464
      ],
      "observed_xyz": [
       4567.763265355824,
       3703.35644156907,
       6730.455684131583
      ],
      "residual_euclidean_vox": 0.4715578037235089,
      "pass": true
     },
     {
      "volume": "pag50",
      "point": 11,
      "kind": "gradient",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.797243026131,
       3702.911591758253,
       6730.620693718205
      ],
      "observed_xyz": [
       4568.1693885637,
       3703.1744320571725,
       6730.277781405508
      ],
      "residual_euclidean_vox": 0.5702334417304685,
      "pass": false
     },
     {
      "volume": "pag50",
      "point": 11,
      "kind": "raw",
      "structural_quality_pass": true,
      "predicted_xyz": [
       4567.797243026131,
       3702.911591758253,
       6730.620693718205
      ],
      "observed_xyz": [
       4567.766631046989,
       3703.3458389557463,
       6730.451753157791
      ],
      "residual_euclidean_vox": 0.4669567803884022,
      "pass": true
     }
    ],
    "phase_pair_composition_xyz": [
     [
      0.9999347099308664,
      4.698718521844823e-06,
      -1.1644614943467747e-05,
      0.3415679315836977
     ],
     [
      -2.548099180247713e-05,
      1.000021383914864,
      -1.198957150354664e-06,
      0.040735485341904853
     ],
     [
      2.226004033877614e-05,
      -5.874504432127577e-07,
      0.9999991034442133,
      -0.08918661003344042
     ],
     [
      0.0,
      0.0,
      0.0,
      1.0
     ]
    ],
    "phase_pair_independent_closure": [
     {
      "point": 8,
      "kind": "gradient",
      "phase0_measured_fractional_xyz": [
       4494.2181985326215,
       3703.21335133561,
       6286.521740448827
      ],
      "phase50_predicted_xyz": [
       4494.210534879485,
       3703.2112216127125,
       6286.524783645415
      ],
      "phase50_direct_observed_xyz": [
       4494.231137119559,
       3703.2416453449778,
       6286.528767541781
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.5287675417813462,
        0.24164534497786047,
        0.23113711955921165
       ],
       "ncc": 0.9912617086311223,
       "integer_grid_ncc": 0.9565991625660457,
       "margin": 0.1847311394717992,
       "competing_ncc": 0.7718680230942465,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 107,
       "continuous_Hessian_eigenvalues": [
        -0.9086418122114083,
        -0.15858837882144966,
        -0.10637438126380205
       ],
       "continuous_Hessian_condition": 8.541923359892749
      },
      "residual_euclidean_vox": 0.03695845250129618,
      "pass": true
     },
     {
      "point": 8,
      "kind": "raw",
      "phase0_measured_fractional_xyz": [
       4494.2181985326215,
       3703.21335133561,
       6286.521740448827
      ],
      "phase50_predicted_xyz": [
       4494.210534879485,
       3703.2112216127125,
       6286.524783645415
      ],
      "phase50_direct_observed_xyz": [
       4494.230793879037,
       3703.2305388204795,
       6286.523694817279
      ],
      "peak": {
       "integer_shift_zyx": [
        1,
        0,
        0
       ],
       "shift_zyx": [
        0.5236948172784456,
        0.23053882047949453,
        0.23079387903683624
       ],
       "ncc": 0.9903095216439213,
       "integer_grid_ncc": 0.937211253758535,
       "margin": 0.2835263276553489,
       "competing_ncc": 0.6536849261031861,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 117,
       "continuous_Hessian_eigenvalues": [
        -0.8794884586190095,
        -0.24245801918731982,
        -0.1949852959460317
       ],
       "continuous_Hessian_condition": 4.510537342582158
      },
      "residual_euclidean_vox": 0.028013695320122253,
      "pass": true
     },
     {
      "point": 9,
      "kind": "gradient",
      "phase0_measured_fractional_xyz": [
       4567.686451909591,
       3953.5283917465313,
       6286.448562589208
      ],
      "phase50_predicted_xyz": [
       4567.675168521157,
       3953.5297427829223,
       6286.453094210006
      ],
      "phase50_direct_observed_xyz": [
       4567.652166763496,
       3953.573849663776,
       6286.450734988603
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.45073498860296246,
        0.5738496637760375,
        -0.34783323650362896
       ],
       "ncc": 0.9903324717825457,
       "integer_grid_ncc": 0.9554970884207373,
       "margin": 0.11919786360256823,
       "competing_ncc": 0.8362992248181691,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 106,
       "continuous_Hessian_eigenvalues": [
        -1.3765161037694214,
        -0.13312888159559103,
        -0.09747359172134094
       ],
       "continuous_Hessian_condition": 14.12193887042377
      },
      "residual_euclidean_vox": 0.049800238149440235,
      "pass": true
     },
     {
      "point": 9,
      "kind": "raw",
      "phase0_measured_fractional_xyz": [
       4567.686451909591,
       3953.5283917465313,
       6286.448562589208
      ],
      "phase50_predicted_xyz": [
       4567.675168521157,
       3953.5297427829223,
       6286.453094210006
      ],
      "phase50_direct_observed_xyz": [
       4567.6605132897375,
       3953.552937511449,
       6286.448370079437
      ],
      "peak": {
       "integer_shift_zyx": [
        0,
        0,
        0
       ],
       "shift_zyx": [
        0.44837007943713386,
        0.5529375114489011,
        -0.33948671026209404
       ],
       "ncc": 0.9890781102223921,
       "integer_grid_ncc": 0.9330421566032908,
       "margin": 0.15748786516240787,
       "competing_ncc": 0.775554291440883,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 106,
       "continuous_Hessian_eigenvalues": [
        -1.027263390601692,
        -0.22585620482641128,
        -0.14142202556378233
       ],
       "continuous_Hessian_condition": 7.263814717025029
      },
      "residual_euclidean_vox": 0.027840413951654472,
      "pass": true
     },
     {
      "point": 10,
      "kind": "gradient",
      "phase0_measured_fractional_xyz": [
       4494.281065476323,
       3952.708058283991,
       6730.3029384600595
      ],
      "phase50_predicted_xyz": [
       4494.26940236283,
       3952.7107300581133,
       6730.305438615691
      ],
      "phase50_direct_observed_xyz": [
       4494.308828170031,
       3952.6557624461902,
       6730.308079104248
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": [
        -0.6919208957521489,
        -0.3442375538096817,
        0.30882817003107393
       ],
       "ncc": 0.9862687308586394,
       "integer_grid_ncc": 0.94864282252321,
       "margin": 0.20688418326539038,
       "competing_ncc": 0.7417586392578196,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 143,
       "continuous_Hessian_eigenvalues": [
        -1.0675909897119875,
        -0.1693057801133904,
        -0.16142860939469975
       ],
       "continuous_Hessian_condition": 6.613393956096608
      },
      "residual_euclidean_vox": 0.06769641655063503,
      "pass": true
     },
     {
      "point": 10,
      "kind": "raw",
      "phase0_measured_fractional_xyz": [
       4494.281065476323,
       3952.708058283991,
       6730.3029384600595
      ],
      "phase50_predicted_xyz": [
       4494.26940236283,
       3952.7107300581133,
       6730.305438615691
      ],
      "phase50_direct_observed_xyz": [
       4494.302601874301,
       3952.6810998788374,
       6730.304330365578
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": [
        -0.6956696344225134,
        -0.31890012116257155,
        0.3026018743015837
       ],
       "ncc": 0.9817510727035333,
       "integer_grid_ncc": 0.9333767911670587,
       "margin": 0.3273787063041663,
       "competing_ncc": 0.6059980848628924,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 116,
       "continuous_Hessian_eigenvalues": [
        -0.744805426914172,
        -0.3404708654115668,
        -0.25858265762156835
       ],
       "continuous_Hessian_condition": 2.880337891817103
      },
      "residual_euclidean_vox": 0.0445127319332462,
      "pass": true
     },
     {
      "point": 11,
      "kind": "gradient",
      "phase0_measured_fractional_xyz": [
       4568.16007365358,
       3703.1710765251482,
       6730.279471440594
      ],
      "phase50_predicted_xyz": [
       4568.142414743763,
       3703.166529739432,
       6730.283762757577
      ],
      "phase50_direct_observed_xyz": [
       4568.175222561818,
       3703.1833579812305,
       6730.281966293447
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": [
        -0.7180337065528863,
        0.18335798123060276,
        0.17522256181836604
       ],
       "ncc": 0.9887676269879468,
       "integer_grid_ncc": 0.9574646302173234,
       "margin": 0.22794151668328877,
       "competing_ncc": 0.7295231135340347,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 135,
       "continuous_Hessian_eigenvalues": [
        -0.8452440894491335,
        -0.16037540617288915,
        -0.13963313351581616
       ],
       "continuous_Hessian_condition": 6.053320355754913
      },
      "residual_euclidean_vox": 0.036915713875859846,
      "pass": true
     },
     {
      "point": 11,
      "kind": "raw",
      "phase0_measured_fractional_xyz": [
       4568.16007365358,
       3703.1710765251482,
       6730.279471440594
      ],
      "phase50_predicted_xyz": [
       4568.142414743763,
       3703.166529739432,
       6730.283762757577
      ],
      "phase50_direct_observed_xyz": [
       4568.180547330947,
       3703.178521673002,
       6730.275638083275
      ],
      "peak": {
       "integer_shift_zyx": [
        -1,
        0,
        0
       ],
       "shift_zyx": [
        -0.7243619167249975,
        0.17852167300200425,
        0.18054733094704384
       ],
       "ncc": 0.9849182792577222,
       "integer_grid_ncc": 0.9517913277538985,
       "margin": 0.37972126201785006,
       "competing_ncc": 0.5720700657360485,
       "refinement_error": null,
       "boundary": false,
       "optimization_success": true,
       "optimization_evaluations": 100,
       "continuous_Hessian_eigenvalues": [
        -0.7096703985822805,
        -0.33167819323464665,
        -0.22981695678438904
       ],
       "continuous_Hessian_condition": 3.08798101111435
      },
      "residual_euclidean_vox": 0.040791065303601874,
      "pass": true
     }
    ],
    "rule_sha256": "b3ae4daddd7653322b1b4c407b1d766c102fa59b9675f4b000e0da175585bc5e",
    "actual_new_bytes": 448790528,
    "reused_bytes": 58720256,
    "peak_rss_bytes": 365002752,
    "limits": "Local finite checkpoints; passing does not validate global warp or ink identity. Allfailedchecks retained; noalternatefit."
   }
  },
  {
   "date": "2026-10-07",
   "kind": "supervised_surfacefix_rejection",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/supervised-surfacefix/actual-validation-with-nulls.json",
   "evidence_sha256": "1ccf9157d6f7ae1060c91ca3c44d489439256dc63349e8e73d025fe6fb397bc8",
   "readout": {
    "scope": "Fixed supervised public PHerc0841 same-sheet model-output consistency; labels never choose offsets and final merged geometry requires official rerender before label-improvement assessment",
    "normal": {
     "status": "no_corrections",
     "accepted_regions": 0,
     "flagged_regions": 16,
     "regions_examined": 16,
     "evaluable_regions": 4,
     "candidate_selection_attempted_regions": 4,
     "selection_points": 221,
     "held_back_points": 223,
     "covered_points": 444,
     "native_valid_points": 1024
    },
    "tangential_geometry_negative": {
     "refused": true,
     "changed_vox": 0.5,
     "changed_axis": "x",
     "native_node": [
      8,
      12
     ],
     "candidate_hashes_recomputed": true,
     "error": "candidate is not the declared bounded normal offset, or changes holes",
     "output_created": false
    },
    "source_hashes_unchanged": true,
    "null_rule_sha256": "2f2bd0fa614a365107a8c9f7580d81b468c31eba1251b5670e59b7dc456637a1",
    "nulls": [
     {
      "axis": 0,
      "shift_px": 150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 2,
      "candidate_selection_attempted_regions": 2,
      "selection_points": 143,
      "held_back_points": 146,
      "covered_points": 289,
      "native_valid_points": 1024
     },
     {
      "axis": 0,
      "shift_px": -150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 3,
      "selection_points": 188,
      "held_back_points": 203,
      "covered_points": 391,
      "native_valid_points": 1024
     },
     {
      "axis": 0,
      "shift_px": 300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 0,
      "candidate_selection_attempted_regions": 0,
      "selection_points": 56,
      "held_back_points": 50,
      "covered_points": 106,
      "native_valid_points": 1024
     },
     {
      "axis": 0,
      "shift_px": -300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 1,
      "selection_points": 175,
      "held_back_points": 183,
      "covered_points": 358,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": 150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 3,
      "selection_points": 180,
      "held_back_points": 179,
      "covered_points": 359,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": -150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 4,
      "selection_points": 192,
      "held_back_points": 195,
      "covered_points": 387,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": 300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 2,
      "candidate_selection_attempted_regions": 1,
      "selection_points": 115,
      "held_back_points": 111,
      "covered_points": 226,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": -300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 2,
      "candidate_selection_attempted_regions": 2,
      "selection_points": 128,
      "held_back_points": 133,
      "covered_points": 261,
      "native_valid_points": 1024
     }
    ],
    "significance": "No statistical significance claim: eight null cases are dependent and some may lack coverage."
   }
  },
  {
   "date": "2026-10-07",
   "kind": "supervised_surfacefix_label_audit",
   "source": "docs/logs/2026-10-07-followup-tests.md",
   "evidence": "docs/evidence/2026-10-07-followup/supervised-surfacefix/label-audit-results.json",
   "evidence_sha256": "bfbba1e556f833f841e56d11dc107194cac37c504fb81e16b780de3b7632c12c",
   "readout": {
    "status": "no_accepted_corrections_descriptive_offset_audit_only",
    "scope": "One label-selected public known-sheet crop. Candidate offset scores are descriptive only and never choose or retry a correction. No independent-scroll or legibility claim.",
    "label_rule_sha256": "a897301ed8a1ffe5099c0e9ff85cdff9affcc2a956962b64f85f32a05f618f26",
    "evaluator_sha256": "31a61bbce494275a7d33443cd71711039c1d607bd9c600e1dafde0bed2b1a8c1",
    "frozen_math_sha256": "a95e703eb73dcfe02a97b48184cbf1ef965ef1872219a3d0774780b7b9d4e72c",
    "actual_manifest_sha256": "1fd130c0491ad4d317743473aedfabad85a271c563d2c843d88ca7a043f525f1",
    "actual_apply_report_sha256": "6e8d956a7e4972f847bf524da2107c48f51c2e8e61123295abb92b5847e4d95b",
    "null_results_sha256": "1ccf9157d6f7ae1060c91ca3c44d489439256dc63349e8e73d025fe6fb397bc8",
    "code_amendment_sha256": "0fc0f44046eeb8f9d2cc3506c078708bacd1795bac32a402320cb87cbe695529",
    "inference_proofs": {
     "reference": {
      "manifest_sha256": "099716e4c69ab06b8df881e88baaeabe622dfe23dd24e0b9deb0edc9a7a933f4",
      "maps": {
       "forward": "e81e5e709cad4cadcb8b33ef76fa9f797d1d5257fb1dd6f01e89772424d22ff8",
       "reverse": "099fc7a06c338281ad1de4320e0f0afbf772685e26b2f3a92965e9121803c058",
       "shuffle": "876911bf3ea1aa785984cb13c98fef90d9e57c9b2c74adb4b88d4ee0c300a937"
      }
     },
     "baseline": {
      "manifest_sha256": "4b8441f49c5ae1bf585f792c927ef9fe5ed05dd606609c8e5a245c4fbc9b21de",
      "maps": {
       "forward": "5482eada009b0df000d2d34b3e9c5e8e35091abd5ade8c884aecffc4fa721a12",
       "reverse": "579c7d0fd214048bb742e1092a0bf7343d89fcc629592bc8f1a6e1269f828ea6",
       "shuffle": "ad77bd97dea3b1b8be7823d629caba2bd1afca9b75080b7bf7a8278650ee8041"
      }
     },
     "minus1": {
      "manifest_sha256": "12650579cfeafa34fcd2d29d48e40122bf2f1ef052879fc516ddaea65d4b9dd9",
      "maps": {
       "forward": "83d558f767cb75a37224dab29532f815214f3d2f3dff0434f7c1d64596301077",
       "reverse": "5df2b564bed9387eaeedb325ecbf3d76e0291c3ba1bf4e16b30fb1d26047ac89",
       "shuffle": "2d0bf8a2dc991356bef5363ed8495ed0902a3b09112486053e90f022b013cc2a"
      }
     },
     "plus1": {
      "manifest_sha256": "129e81870fd7dc1ea9da69da92525d53841d1ff7e021fd2fc98c9780206876f7",
      "maps": {
       "forward": "4f38382c6d8849794327a7dba03192402fe06b8b8c026cbf5097d24c16fe68d1",
       "reverse": "79c1f0f0dfa40578a16f1eb375c4ee93c0344af22bb2f0b40cba35dd91a58c36",
       "shuffle": "18fbed42cb64de7b2fbc10527fb19e15cc0d486ae5c1b050056a5c8f3ef25984"
      }
     }
    },
    "combined_render_proof": null,
    "primary_supervised_pixels": 262130,
    "common_hp_core_pixels": 200680,
    "bootstrap_draws": 1000,
    "bootstrap_blocks": 36,
    "rows": {
     "baseline/forward": {
      "auc": {
       "auc": 0.9211,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.8835986925163625,
        0.9532167557878806
       ]
      },
      "hp": {
       "hp_r": 0.03304449730060359,
       "null_max_abs": 0.0071014361532177,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.0105077259294234,
        0.05398288681261318
       ]
      },
      "zero_fraction": 0.0
     },
     "baseline/reverse": {
      "auc": {
       "auc": 0.5951,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.4920567575698745,
        0.7043636293171107
       ]
      },
      "hp": {
       "hp_r": -0.0005581627980643473,
       "null_max_abs": 0.011116705354633447,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        -0.012354891912870737,
        0.010453402042686247
       ]
      },
      "zero_fraction": 0.0
     },
     "baseline/shuffle": {
      "auc": {
       "auc": 0.6439,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.5303614970601143,
        0.7552907282955225
       ]
      },
      "hp": {
       "hp_r": -0.007767375778516303,
       "null_max_abs": 0.006269926644425831,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        -0.018393686117214442,
        0.002395241532657731
       ]
      },
      "zero_fraction": 0.0
     },
     "minus1/forward": {
      "auc": {
       "auc": 0.9086,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.8635830632449587,
        0.9437274480263244
       ]
      },
      "hp": {
       "hp_r": 0.03325589214050017,
       "null_max_abs": 0.012776906871730847,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.006572142087426752,
        0.05920710404111539
       ]
      },
      "zero_fraction": 0.0
     },
     "minus1/reverse": {
      "auc": {
       "auc": 0.5952,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.48864866701127335,
        0.7104040002978523
       ]
      },
      "hp": {
       "hp_r": -0.005471283491188925,
       "null_max_abs": 0.02396355970884853,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        -0.014037184783955109,
        0.0026898715906785334
       ]
      },
      "zero_fraction": 0.0
     },
     "minus1/shuffle": {
      "auc": {
       "auc": 0.5474,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.4309876176036894,
        0.6696283217668592
       ]
      },
      "hp": {
       "hp_r": 0.010829553592106023,
       "null_max_abs": 0.0209239394012872,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        -0.0007260685885440959,
        0.023058016441392007
       ]
      },
      "zero_fraction": 0.0
     },
     "plus1/forward": {
      "auc": {
       "auc": 0.93,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.8923897371871317,
        0.9595123663793989
       ]
      },
      "hp": {
       "hp_r": 0.03953139626645315,
       "null_max_abs": 0.008587090082687628,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.019237936086902045,
        0.06101771034413416
       ]
      },
      "zero_fraction": 0.0
     },
     "plus1/reverse": {
      "auc": {
       "auc": 0.5518,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.45555838105278934,
        0.6570636633477676
       ]
      },
      "hp": {
       "hp_r": -0.0012049454100113888,
       "null_max_abs": 0.011656338202589453,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        -0.010933027341337059,
        0.009184849860974157
       ]
      },
      "zero_fraction": 0.0
     },
     "plus1/shuffle": {
      "auc": {
       "auc": 0.5911,
       "ink_px": 83199,
       "background_px": 178931,
       "zero_px_in_mask": 0,
       "zero_px_counted": true,
       "inner": 0
      },
      "auc_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        0.5024214245040864,
        0.6833733504543881
       ]
      },
      "hp": {
       "hp_r": -0.010108105978907805,
       "null_max_abs": 0.009761010826488451,
       "usable_null_shifts": 3,
       "px": 200680,
       "same_global_smoothing_support": true
      },
      "hp_bootstrap": {
       "finite_draws": 1000,
       "finite_fraction": 1.0,
       "ci95": [
        -0.020769610617989504,
        0.0003930586105450283
       ]
      },
      "zero_fraction": 0.0
     }
    },
    "candidate_descriptive_contrasts": {
     "auc": {
      "minus1-baseline": {
       "difference": -0.012520045915068057,
       "simultaneous_ci95": [
        -0.026020846213817728,
        0.0009807543836816146
       ],
       "finite_fraction": 1.0,
       "family_size": 2
      },
      "plus1-baseline": {
       "difference": 0.008890214074979585,
       "simultaneous_ci95": [
        -0.004610586223770086,
        0.022391014373729257
       ],
       "finite_fraction": 1.0,
       "family_size": 2
      }
     },
     "hp": {
      "minus1-baseline": {
       "difference": 0.00021139483989657953,
       "simultaneous_ci95": [
        -0.01102890111364999,
        0.011451690793443148
       ],
       "finite_fraction": 1.0,
       "family_size": 2
      },
      "plus1-baseline": {
       "difference": 0.006486898965849558,
       "simultaneous_ci95": [
        -0.004753396987697011,
        0.017727194919396127
       ],
       "finite_fraction": 1.0,
       "family_size": 2
      }
     }
    },
    "depth_controls": {
     "auc": {
      "baseline/forward-minus-reverse": {
       "difference": 0.32604489089009414,
       "simultaneous_ci95": [
        0.1778997488068215,
        0.4741900329733668
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "baseline/forward-minus-shuffle": {
       "difference": 0.27723021418356786,
       "simultaneous_ci95": [
        0.1290850721002952,
        0.42537535626684053
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "minus1/forward-minus-reverse": {
       "difference": 0.31337821720879455,
       "simultaneous_ci95": [
        0.1652330751255219,
        0.4615233592920672
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "minus1/forward-minus-shuffle": {
       "difference": 0.3612241607933093,
       "simultaneous_ci95": [
        0.21307901871003668,
        0.509369302876582
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "plus1/forward-minus-reverse": {
       "difference": 0.3781725458774159,
       "simultaneous_ci95": [
        0.23002740379414324,
        0.5263176879606886
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "plus1/forward-minus-shuffle": {
       "difference": 0.33894835706493853,
       "simultaneous_ci95": [
        0.1908032149816659,
        0.4870934991482112
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      }
     },
     "hp": {
      "baseline/forward-minus-reverse": {
       "difference": 0.033602660098667934,
       "simultaneous_ci95": [
        0.004847302292072829,
        0.06235801790526304
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "baseline/forward-minus-shuffle": {
       "difference": 0.04081187307911989,
       "simultaneous_ci95": [
        0.012056515272524784,
        0.069567230885715
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "minus1/forward-minus-reverse": {
       "difference": 0.03872717563168909,
       "simultaneous_ci95": [
        0.009971817825093987,
        0.0674825334382842
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "minus1/forward-minus-shuffle": {
       "difference": 0.022426338548394147,
       "simultaneous_ci95": [
        -0.006329019258200959,
        0.05118169635498925
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "plus1/forward-minus-reverse": {
       "difference": 0.04073634167646454,
       "simultaneous_ci95": [
        0.011980983869869434,
        0.06949169948305964
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      },
      "plus1/forward-minus-shuffle": {
       "difference": 0.049639502245360956,
       "simultaneous_ci95": [
        0.02088414443876585,
        0.07839486005195606
       ],
       "finite_fraction": 1.0,
       "family_size": 6
      }
     }
    },
    "fixed_spatial_halves": {
     "upper": {
      "eligible": true,
      "ink_px": 47374,
      "background_px": 83684,
      "metrics": {
       "baseline/forward": {
        "auc": 0.8957,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "baseline/reverse": {
        "auc": 0.6157,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "baseline/shuffle": {
        "auc": 0.7666,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "minus1/forward": {
        "auc": 0.8714,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "minus1/reverse": {
        "auc": 0.622,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "minus1/shuffle": {
        "auc": 0.6239,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "plus1/forward": {
        "auc": 0.9099,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "plus1/reverse": {
        "auc": 0.6038,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "plus1/shuffle": {
        "auc": 0.5846,
        "ink_px": 47374,
        "background_px": 83684,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       }
      }
     },
     "lower": {
      "eligible": true,
      "ink_px": 35825,
      "background_px": 95247,
      "metrics": {
       "baseline/forward": {
        "auc": 0.9484,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "baseline/reverse": {
        "auc": 0.5615,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "baseline/shuffle": {
        "auc": 0.5798,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "minus1/forward": {
        "auc": 0.9449,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "minus1/reverse": {
        "auc": 0.556,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "minus1/shuffle": {
        "auc": 0.5023,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "plus1/forward": {
        "auc": 0.9558,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "plus1/reverse": {
        "auc": 0.4832,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       },
       "plus1/shuffle": {
        "auc": 0.6087,
        "ink_px": 35825,
        "background_px": 95247,
        "zero_px_in_mask": 0,
        "zero_px_counted": true,
        "inner": 0
       }
      }
     }
    },
    "combined_vs_baseline": {},
    "accepted_regions": 0,
    "changed_native_points": 0,
    "edited_support_supervised_pixels": 0,
    "edited_support_rows": {},
    "edited_support_combined_vs_baseline": {},
    "correction_success_gate": {
     "pixel_ranking_improvement": false,
     "ink_detail_improvement": false,
     "coverage_ok": null,
     "both_halves_improve": null,
     "matched_auc_controls_pass": null
    },
    "nulls": [
     {
      "axis": 0,
      "shift_px": 150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 2,
      "candidate_selection_attempted_regions": 2,
      "selection_points": 143,
      "held_back_points": 146,
      "covered_points": 289,
      "native_valid_points": 1024
     },
     {
      "axis": 0,
      "shift_px": -150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 3,
      "selection_points": 188,
      "held_back_points": 203,
      "covered_points": 391,
      "native_valid_points": 1024
     },
     {
      "axis": 0,
      "shift_px": 300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 0,
      "candidate_selection_attempted_regions": 0,
      "selection_points": 56,
      "held_back_points": 50,
      "covered_points": 106,
      "native_valid_points": 1024
     },
     {
      "axis": 0,
      "shift_px": -300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 1,
      "selection_points": 175,
      "held_back_points": 183,
      "covered_points": 358,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": 150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 3,
      "selection_points": 180,
      "held_back_points": 179,
      "covered_points": 359,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": -150,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 4,
      "candidate_selection_attempted_regions": 4,
      "selection_points": 192,
      "held_back_points": 195,
      "covered_points": 387,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": 300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 2,
      "candidate_selection_attempted_regions": 1,
      "selection_points": 115,
      "held_back_points": 111,
      "covered_points": 226,
      "native_valid_points": 1024
     },
     {
      "axis": 1,
      "shift_px": -300,
      "status": "no_corrections",
      "accepted_regions": 0,
      "flagged_regions": 16,
      "regions_examined": 16,
      "evaluable_regions": 2,
      "candidate_selection_attempted_regions": 2,
      "selection_points": 128,
      "held_back_points": 133,
      "covered_points": 261,
      "native_valid_points": 1024
     }
    ]
   }
  },
  {
   "date": "2026-10-07",
   "kind": "upstream_coarse_support_review",
   "source": "docs/logs/2026-10-07-coarse-support-mask.md",
   "evidence": "docs/evidence/2026-10-07-followup/pr1996-review/review-receipt.json",
   "evidence_sha256": "95023806156c94ddeef11f43817ce1c7eaa0718eff52f4e06b4ef161d4818733",
   "readout": {
    "pr_url": "https://github.com/ScrollPrize/villa/pull/1996",
    "head": "d651f04d2f5c73fc5ba46c5f859392391cc59b43",
    "base": "e0bbb8b40a2db58b1d71864f286eb85717e59e64",
    "evidence_repository": "https://github.com/ibarapascal/ink-surface-support",
    "evidence_commit": "e8623546a3e147e656f3059cf5c8d78db2fe7473",
    "base_grower_sha256": "8156f58f3a7cf1ff5301a211935e63280304328e729f4bbc6f4699dbb0c734e6",
    "evidence_declared_base_sha256": "8156f58f3a7cf1ff5301a211935e63280304328e729f4bbc6f4699dbb0c734e6",
    "tests": {
     "upstream_pytest": {
      "passed": 12,
      "seconds": 0.46,
      "pytest": "9.1.1"
     },
     "linked_helper_unittest": {
      "passed": 11
     },
     "randomized_base_and_helper_parity": {
      "passed": 240
     },
     "saved_report_rebuild": {
      "status": "PASS",
      "derived_files": 6,
      "scope": "Saved report integrity and arithmetic, not new measurement"
     }
    },
    "new_ct_or_predictions_downloaded": false,
    "notes": [
     "Only public source and two S3 zarray metadata documents were fetched.",
     "The isolated pytest runtime is outside the shared repository."
    ],
    "runtime": {
     "python": "3.12.14",
     "numpy": "2.5.3",
     "scipy": "1.18.1"
    }
   }
  },
  {
   "date": "2026-10-07",
   "kind": "real_mesh_coarse_support",
   "source": "docs/logs/2026-10-07-coarse-support-mask.md",
   "evidence": "docs/evidence/2026-10-07-followup/pr1996-geometry/result.json",
   "evidence_sha256": "5709a5764f9ddef99722edafed0af3a50595d0cd0a5c973a96baf38f56933c62",
   "readout": {
    "rule_sha256": "8d6fc3431201282c959b83e6437750002ba999c4ed1ca2a52d2bfd513cc01646",
    "rows": [
     {
      "mode": "erode",
      "factor": 2,
      "coarse_shape": [
       106,
       119
      ],
      "retainedvertices": 8870,
      "retainedquads": 8644,
      "fine_support_violations": 0,
      "fine_valid_cells": 36338,
      "represented_fine_valid_cells": 34576,
      "represented_fine_valid_cell_fraction": 0.9515108151246628,
      "retainedquads_failing_local_jacobian": 13,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       1.6194734601423662,
       20.43917049627685,
       66.89782245596513,
       206.911799406566
      ],
      "mesh": "w00"
     },
     {
      "mode": "support",
      "factor": 2,
      "coarse_shape": [
       106,
       119
      ],
      "retainedvertices": 9199,
      "retainedquads": 8966,
      "fine_support_violations": 0,
      "fine_valid_cells": 36338,
      "represented_fine_valid_cells": 35864,
      "represented_fine_valid_cell_fraction": 0.9869558038417084,
      "retainedquads_failing_local_jacobian": 18,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       1.6514926925253204,
       20.68831854359533,
       68.62861423880503,
       206.911799406566
      ],
      "mesh": "w00"
     },
     {
      "mode": "erode",
      "factor": 4,
      "coarse_shape": [
       53,
       60
      ],
      "retainedvertices": 2137,
      "retainedquads": 2031,
      "fine_support_violations": 0,
      "fine_valid_cells": 36338,
      "represented_fine_valid_cells": 32496,
      "represented_fine_valid_cell_fraction": 0.8942704606747757,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       17.161703370617868,
       75.4301921387718,
       172.93376010666557,
       373.7301539695508
      ],
      "mesh": "w00"
     },
     {
      "mode": "support",
      "factor": 4,
      "coarse_shape": [
       53,
       60
      ],
      "retainedvertices": 2296,
      "retainedquads": 2185,
      "fine_support_violations": 0,
      "fine_valid_cells": 36338,
      "represented_fine_valid_cells": 34960,
      "represented_fine_valid_cell_fraction": 0.9620782651769497,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       17.617051755850817,
       76.47599078337332,
       174.69621150126392,
       373.7301539695508
      ],
      "mesh": "w00"
     },
     {
      "mode": "erode",
      "factor": 2,
      "coarse_shape": [
       116,
       118
      ],
      "retainedvertices": 8910,
      "retainedquads": 8670,
      "fine_support_violations": 0,
      "fine_valid_cells": 36509,
      "represented_fine_valid_cells": 34680,
      "represented_fine_valid_cell_fraction": 0.9499027637021008,
      "retainedquads_failing_local_jacobian": 19,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       1.6120435451538957,
       21.005933379870186,
       71.89965556832662,
       195.78267603751303
      ],
      "mesh": "ag896"
     },
     {
      "mode": "support",
      "factor": 2,
      "coarse_shape": [
       116,
       118
      ],
      "retainedvertices": 9251,
      "retainedquads": 9002,
      "fine_support_violations": 0,
      "fine_valid_cells": 36509,
      "represented_fine_valid_cells": 36008,
      "represented_fine_valid_cell_fraction": 0.9862773562683174,
      "retainedquads_failing_local_jacobian": 27,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       1.6262381861029072,
       21.299334650344797,
       73.13480094476563,
       198.01491217465815
      ],
      "mesh": "ag896"
     },
     {
      "mode": "erode",
      "factor": 4,
      "coarse_shape": [
       58,
       59
      ],
      "retainedvertices": 2144,
      "retainedquads": 2031,
      "fine_support_violations": 0,
      "fine_valid_cells": 36509,
      "represented_fine_valid_cells": 32496,
      "represented_fine_valid_cell_fraction": 0.8900818976142869,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       18.115201999501117,
       77.15671165825667,
       184.4090719574988,
       413.04746340139843
      ],
      "mesh": "ag896"
     },
     {
      "mode": "support",
      "factor": 4,
      "coarse_shape": [
       58,
       59
      ],
      "retainedvertices": 2308,
      "retainedquads": 2190,
      "fine_support_violations": 0,
      "fine_valid_cells": 36509,
      "represented_fine_valid_cells": 35040,
      "represented_fine_valid_cell_fraction": 0.9597633460242679,
      "retainedquads_failing_local_jacobian": 1,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       18.651713767164974,
       79.09319518045162,
       184.52626079299577,
       413.04746340139843
      ],
      "mesh": "ag896"
     }
    ],
    "controls": [
     {
      "mode": "erode",
      "factor": 4,
      "coarse_shape": [
       5,
       5
      ],
      "retainedvertices": 9,
      "retainedquads": 4,
      "fine_support_violations": 0,
      "fine_valid_cells": 256,
      "represented_fine_valid_cells": 64,
      "represented_fine_valid_cell_fraction": 0.25,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       0.0,
       0.0,
       0.0,
       0.0
      ],
      "phantom": "full_plane"
     },
     {
      "mode": "support",
      "factor": 4,
      "coarse_shape": [
       5,
       5
      ],
      "retainedvertices": 25,
      "retainedquads": 16,
      "fine_support_violations": 0,
      "fine_valid_cells": 256,
      "represented_fine_valid_cells": 256,
      "represented_fine_valid_cell_fraction": 1.0,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       0.0,
       0.0,
       0.0,
       0.0
      ],
      "phantom": "full_plane"
     },
     {
      "mode": "erode",
      "factor": 4,
      "coarse_shape": [
       5,
       5
      ],
      "retainedvertices": 9,
      "retainedquads": 4,
      "fine_support_violations": 1,
      "fine_valid_cells": 252,
      "represented_fine_valid_cells": 60,
      "represented_fine_valid_cell_fraction": 0.23809523809523808,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       0.0,
       0.0,
       0.0,
       0.0
      ],
      "phantom": "interior_hole"
     },
     {
      "mode": "support",
      "factor": 4,
      "coarse_shape": [
       5,
       5
      ],
      "retainedvertices": 24,
      "retainedquads": 12,
      "fine_support_violations": 0,
      "fine_valid_cells": 252,
      "represented_fine_valid_cells": 192,
      "represented_fine_valid_cell_fraction": 0.7619047619047619,
      "retainedquads_failing_local_jacobian": 0,
      "bilinear_vs_working_vertex_error_um_quantiles": [
       0.0,
       0.0,
       0.0,
       0.0,
       0.0
      ],
      "phantom": "interior_hole"
     }
    ],
    "elapsed_seconds": 3.1433862040003078,
    "peak_rss_bytes": 70533120,
    "limits": "Public PHerc0841 cached mesh resampling/mask behavior; not original track-graph growth, flattening benchmark, ink or targets; same-UV errors are geometry discrepancy, notnormaldistance, flatten distortion orink readability."
   }
  },
  {
   "date": "2026-10-07",
   "kind": "coarse_support_continuous_correspondence",
   "source": "docs/logs/2026-10-07-coarse-support-mask.md",
   "evidence": "docs/evidence/2026-10-07-followup/pr1996-geometry/correspondence-result.json",
   "evidence_sha256": "65e0571d272db6377e40ba91ee107f405e1a11d0e9d22915fbfc03d036db8eba",
   "readout": {
    "rule_sha256": "a306b4aaff6181bb2165876dd18c3ee5b7cfa3533b3a2a87e8ae39066399ab9c",
    "rows": [
     {
      "mesh": "w00",
      "factor": 2,
      "mode": "erode",
      "source_native_indices": [
       358,
       2992,
       4424,
       5782,
       7001,
       8182,
       9252,
       10362,
       11385,
       12440,
       13417,
       14374,
       15316,
       16245,
       17126,
       18026,
       18912,
       19790,
       20642,
       21508,
       22349,
       23201,
       24031,
       24869,
       25690,
       26518,
       27331,
       28148,
       28956,
       29767,
       30574,
       31383,
       32199,
       33012,
       33839,
       34660,
       35505,
       36342,
       37207,
       38102,
       38982,
       39905,
       40858,
       41841,
       42939,
       44143,
       45999,
       49826
      ],
      "points": 48,
      "certified_eligible": 32,
      "ambiguous": 9,
      "unresolved": 2,
      "no_match_or_distance_fail": 5,
      "finite_uppergap_um_quantiles": [
       0.0,
       4.529627900144695,
       16.854552815760144,
       34.22259141892183
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "w00",
      "factor": 2,
      "mode": "support",
      "source_native_indices": [
       358,
       2992,
       4424,
       5782,
       7001,
       8182,
       9252,
       10362,
       11385,
       12440,
       13417,
       14374,
       15316,
       16245,
       17126,
       18026,
       18912,
       19790,
       20642,
       21508,
       22349,
       23201,
       24031,
       24869,
       25690,
       26518,
       27331,
       28148,
       28956,
       29767,
       30574,
       31383,
       32199,
       33012,
       33839,
       34660,
       35505,
       36342,
       37207,
       38102,
       38982,
       39905,
       40858,
       41841,
       42939,
       44143,
       45999,
       49826
      ],
      "points": 48,
      "certified_eligible": 33,
      "ambiguous": 9,
      "unresolved": 2,
      "no_match_or_distance_fail": 4,
      "finite_uppergap_um_quantiles": [
       0.0,
       4.756361278028184,
       18.321379511820552,
       34.22259141892183
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "w00",
      "factor": 4,
      "mode": "erode",
      "source_native_indices": [
       358,
       2992,
       4424,
       5782,
       7001,
       8182,
       9252,
       10362,
       11385,
       12440,
       13417,
       14374,
       15316,
       16245,
       17126,
       18026,
       18912,
       19790,
       20642,
       21508,
       22349,
       23201,
       24031,
       24869,
       25690,
       26518,
       27331,
       28148,
       28956,
       29767,
       30574,
       31383,
       32199,
       33012,
       33839,
       34660,
       35505,
       36342,
       37207,
       38102,
       38982,
       39905,
       40858,
       41841,
       42939,
       44143,
       45999,
       49826
      ],
      "points": 48,
      "certified_eligible": 37,
      "ambiguous": 2,
      "unresolved": 0,
      "no_match_or_distance_fail": 9,
      "finite_uppergap_um_quantiles": [
       0.0,
       17.86548435238069,
       42.18929785564872,
       48.77270160416609
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "w00",
      "factor": 4,
      "mode": "support",
      "source_native_indices": [
       358,
       2992,
       4424,
       5782,
       7001,
       8182,
       9252,
       10362,
       11385,
       12440,
       13417,
       14374,
       15316,
       16245,
       17126,
       18026,
       18912,
       19790,
       20642,
       21508,
       22349,
       23201,
       24031,
       24869,
       25690,
       26518,
       27331,
       28148,
       28956,
       29767,
       30574,
       31383,
       32199,
       33012,
       33839,
       34660,
       35505,
       36342,
       37207,
       38102,
       38982,
       39905,
       40858,
       41841,
       42939,
       44143,
       45999,
       49826
      ],
      "points": 48,
      "certified_eligible": 40,
      "ambiguous": 2,
      "unresolved": 0,
      "no_match_or_distance_fail": 6,
      "finite_uppergap_um_quantiles": [
       0.0,
       19.956339593344744,
       42.27111533177704,
       48.77270160416609
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "ag896",
      "factor": 2,
      "mode": "erode",
      "source_native_indices": [
       625,
       7203,
       8672,
       10038,
       11255,
       12449,
       13599,
       14711,
       15739,
       16800,
       17786,
       18754,
       19704,
       20638,
       21560,
       22462,
       23330,
       24213,
       25090,
       25939,
       26802,
       27658,
       28492,
       29333,
       30158,
       30988,
       31806,
       32626,
       33436,
       34250,
       35058,
       35876,
       36686,
       37502,
       38332,
       39157,
       40002,
       40858,
       41703,
       42594,
       43507,
       44435,
       45387,
       46372,
       47476,
       48683,
       50543,
       54362
      ],
      "points": 48,
      "certified_eligible": 32,
      "ambiguous": 9,
      "unresolved": 1,
      "no_match_or_distance_fail": 6,
      "finite_uppergap_um_quantiles": [
       0.0,
       2.3726686212326693,
       14.369070803109718,
       36.73150626082486
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "ag896",
      "factor": 2,
      "mode": "support",
      "source_native_indices": [
       625,
       7203,
       8672,
       10038,
       11255,
       12449,
       13599,
       14711,
       15739,
       16800,
       17786,
       18754,
       19704,
       20638,
       21560,
       22462,
       23330,
       24213,
       25090,
       25939,
       26802,
       27658,
       28492,
       29333,
       30158,
       30988,
       31806,
       32626,
       33436,
       34250,
       35058,
       35876,
       36686,
       37502,
       38332,
       39157,
       40002,
       40858,
       41703,
       42594,
       43507,
       44435,
       45387,
       46372,
       47476,
       48683,
       50543,
       54362
      ],
      "points": 48,
      "certified_eligible": 34,
      "ambiguous": 9,
      "unresolved": 1,
      "no_match_or_distance_fail": 4,
      "finite_uppergap_um_quantiles": [
       0.0,
       2.3726686212326693,
       14.288305904541486,
       36.73150626082486
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "ag896",
      "factor": 4,
      "mode": "erode",
      "source_native_indices": [
       625,
       7203,
       8672,
       10038,
       11255,
       12449,
       13599,
       14711,
       15739,
       16800,
       17786,
       18754,
       19704,
       20638,
       21560,
       22462,
       23330,
       24213,
       25090,
       25939,
       26802,
       27658,
       28492,
       29333,
       30158,
       30988,
       31806,
       32626,
       33436,
       34250,
       35058,
       35876,
       36686,
       37502,
       38332,
       39157,
       40002,
       40858,
       41703,
       42594,
       43507,
       44435,
       45387,
       46372,
       47476,
       48683,
       50543,
       54362
      ],
      "points": 48,
      "certified_eligible": 27,
      "ambiguous": 3,
      "unresolved": 0,
      "no_match_or_distance_fail": 18,
      "finite_uppergap_um_quantiles": [
       0.0,
       16.5152048717107,
       33.24357209052666,
       37.01192888058795
      ],
      "tolerance_vox": 0.0033829117335329633
     },
     {
      "mesh": "ag896",
      "factor": 4,
      "mode": "support",
      "source_native_indices": [
       625,
       7203,
       8672,
       10038,
       11255,
       12449,
       13599,
       14711,
       15739,
       16800,
       17786,
       18754,
       19704,
       20638,
       21560,
       22462,
       23330,
       24213,
       25090,
       25939,
       26802,
       27658,
       28492,
       29333,
       30158,
       30988,
       31806,
       32626,
       33436,
       34250,
       35058,
       35876,
       36686,
       37502,
       38332,
       39157,
       40002,
       40858,
       41703,
       42594,
       43507,
       44435,
       45387,
       46372,
       47476,
       48683,
       50543,
       54362
      ],
      "points": 48,
      "certified_eligible": 29,
      "ambiguous": 3,
      "unresolved": 0,
      "no_match_or_distance_fail": 16,
      "finite_uppergap_um_quantiles": [
       0.0,
       16.5152048717107,
       32.631176468144446,
       37.01192888058795
      ],
      "tolerance_vox": 0.0033829117335329633
     }
    ],
    "elapsed_seconds": 9.11014940899986,
    "peak_rss_bytes": 72138752,
    "limits": "Source native20vox grid;40/80vox resampling stress, notParis4spacing5/10 orflattenCPU reproduction. Samplecoverage geometric only; mapsnotprovided."
   }
  }
 ],
 "adaptive_geometry_checks": {
  "date": "2026-10-07",
  "source": "docs/logs/2026-10-07-adaptive-resolution.md",
  "original_primary": "ag896",
  "original_primary_passed": false,
  "interpretation": "Promising local geometric discrimination; failed original primary retained. No exported mesh, global topology, speedup or ink accuracy claim.",
  "experiments": [
   {
    "experiment": "original",
    "rule_sha256": "ad7c395cbe658679fe84481fd5d190fe45ca622ac6817977b73dc897614e0cab",
    "evidence": "docs/evidence/2026-10-07-adaptive-resolution/result.json",
    "evidence_sha256": "5daaf8828a8adb31801e099c4a9052e576c37335daad31198778a485e04b68ea",
    "elapsed_seconds": 9.273897722000015,
    "peak_rss_bytes": 74723328,
    "traces": [
     {
      "trace": "w00",
      "native_regular_quads": 36237,
      "selected_blocks": 6304,
      "hypothetical_primitive_reduction": 0.5218975080718603,
      "query_summaries": {
       "adaptive": {
        "native_eligible": 120,
        "retained": 116,
        "retention": 0.9666666666666667,
        "total_eligible": 116,
        "ambiguous": 0,
        "unresolved": 0
       },
       "uniform2": {
        "native_eligible": 120,
        "retained": 104,
        "retention": 0.8666666666666667,
        "total_eligible": 104,
        "ambiguous": 0,
        "unresolved": 0
       },
       "count_matched_random": {
        "native_eligible": 120,
        "retained": 110,
        "retention": 0.9166666666666666,
        "total_eligible": 110,
        "ambiguous": 0,
        "unresolved": 0
       }
      },
      "dense_all_pass": true,
      "primary_hypothesis_passed": true
     },
     {
      "trace": "ag896",
      "native_regular_quads": 36381,
      "selected_blocks": 6261,
      "hypothetical_primitive_reduction": 0.5162859734476787,
      "query_summaries": {
       "adaptive": {
        "native_eligible": 120,
        "retained": 112,
        "retention": 0.9333333333333333,
        "total_eligible": 112,
        "ambiguous": 0,
        "unresolved": 0
       },
       "uniform2": {
        "native_eligible": 120,
        "retained": 104,
        "retention": 0.8666666666666667,
        "total_eligible": 104,
        "ambiguous": 0,
        "unresolved": 0
       },
       "count_matched_random": {
        "native_eligible": 120,
        "retained": 108,
        "retention": 0.9,
        "total_eligible": 108,
        "ambiguous": 0,
        "unresolved": 0
       }
      },
      "dense_all_pass": true,
      "primary_hypothesis_passed": false
     }
    ]
   },
   {
    "experiment": "separately_frozen_ag405_followup",
    "rule_sha256": "e149b3f78af6620aa07041d4c2530e1ed9ffd064de2ff59dc60ef7b4baf4bd9a",
    "evidence": "docs/evidence/2026-10-07-adaptive-resolution/ag405-primary/result.json",
    "evidence_sha256": "97b0414de4f927d5f60e4f41a497708325dc900de7072f70680dfaeba28be6f3",
    "elapsed_seconds": 4.339847684999995,
    "peak_rss_bytes": 66490368,
    "traces": [
     {
      "trace": "ag405",
      "native_regular_quads": 34236,
      "selected_blocks": 6167,
      "hypothetical_primitive_reduction": 0.5403960743077463,
      "query_summaries": {
       "adaptive": {
        "native_eligible": 120,
        "retained": 116,
        "retention": 0.9666666666666667,
        "total_eligible": 116,
        "ambiguous": 0,
        "unresolved": 0
       },
       "uniform2": {
        "native_eligible": 120,
        "retained": 106,
        "retention": 0.8833333333333333,
        "total_eligible": 106,
        "ambiguous": 0,
        "unresolved": 0
       },
       "count_matched_random": {
        "native_eligible": 120,
        "retained": 111,
        "retention": 0.925,
        "total_eligible": 111,
        "ambiguous": 0,
        "unresolved": 0
       }
      },
      "dense_all_pass": true,
      "primary_hypothesis_passed": true
     }
    ]
   }
  ],
  "independent_audit": "docs/evidence/2026-10-07-adaptive-resolution/independent-review/adaptive-independent-review.md",
  "margin_followup": {
   "primary": "ag896",
   "source": "docs/evidence/2026-10-07-adaptive-margin/result.json",
   "source_sha256": "42a1f0e6c43d2a6a9667cf830ce5c9c2a964f900513994b8f47ba90606cb58e0",
   "rule_sha256": "563513c0b47acac2698c1d128813ea30b55847c5f7e6ef2e78d7326f504a544d",
   "elapsed_seconds": 14.671653177000053,
   "peak_rss_bytes": 72364032,
   "traces": [
    {
     "trace": "ag896",
     "native_regular_quads": 36381,
     "selected_blocks": 1887,
     "hypothetical_primitive_reduction": 0.155603199472252,
     "query_summaries": {
      "adaptive": {
       "native_eligible": 120,
       "retained": 120,
       "retention": 1.0,
       "total_eligible": 120,
       "ambiguous": 0,
       "unresolved": 0
      },
      "uniform2": {
       "native_eligible": 120,
       "retained": 103,
       "retention": 0.8583333333333333,
       "total_eligible": 103,
       "ambiguous": 0,
       "unresolved": 0
      },
      "count_matched_random": {
       "native_eligible": 120,
       "retained": 117,
       "retention": 0.975,
       "total_eligible": 117,
       "ambiguous": 0,
       "unresolved": 0
      }
     },
     "primary_hypothesis_passed": false
    },
    {
     "trace": "w00",
     "native_regular_quads": 36237,
     "selected_blocks": 1922,
     "hypothetical_primitive_reduction": 0.15911913237850817,
     "query_summaries": {
      "adaptive": {
       "native_eligible": 118,
       "retained": 118,
       "retention": 1.0,
       "total_eligible": 118,
       "ambiguous": 0,
       "unresolved": 2
      },
      "uniform2": {
       "native_eligible": 118,
       "retained": 95,
       "retention": 0.8050847457627118,
       "total_eligible": 97,
       "ambiguous": 0,
       "unresolved": 0
      },
      "count_matched_random": {
       "native_eligible": 118,
       "retained": 114,
       "retention": 0.9661016949152542,
       "total_eligible": 116,
       "ambiguous": 0,
       "unresolved": 0
      }
     },
     "primary_hypothesis_passed": false
    },
    {
     "trace": "ag405",
     "native_regular_quads": 34236,
     "selected_blocks": 2112,
     "hypothetical_primitive_reduction": 0.18506834910620398,
     "query_summaries": {
      "adaptive": {
       "native_eligible": 120,
       "retained": 120,
       "retention": 1.0,
       "total_eligible": 120,
       "ambiguous": 0,
       "unresolved": 0
      },
      "uniform2": {
       "native_eligible": 120,
       "retained": 105,
       "retention": 0.875,
       "total_eligible": 105,
       "ambiguous": 0,
       "unresolved": 0
      },
      "count_matched_random": {
       "native_eligible": 120,
       "retained": 117,
       "retention": 0.975,
       "total_eligible": 117,
       "ambiguous": 0,
       "unresolved": 0
      }
     },
     "primary_hypothesis_passed": false
    }
   ],
   "interpretation": "All combined endpoints failed because reduction was below20percent. All baseline-eligible queries retained on a fresh fixed query set; two native w00 abstentions preserved. Distance bound proves existence, not uniqueness."
  }
 },
 "production_speed_checks": {
  "date": "2026-10-07",
  "source": "docs/logs/2026-10-07-production-speed.md",
  "baseline_commit": "e2fd0c0afb5b3bb4e9f956f8cba648e674299949",
  "scope": "Single-core CPU on cached public PHerc0841 test crops; optimized existing operations, no model inference/flattening/whole-scroll throughput claim.",
  "geometry": {
   "evidence": "docs/evidence/2026-10-07-production-speed/geometry/process-result.json",
   "sha256": "1938ce7d3a49be65e8fee8bd1f9a4e9b978b867137251fa8be6dea614bad0769",
   "alternating_pairs_per_case": 5,
   "medians": {
    "supervised_schema2": {
     "baseline": {
      "process_wall_seconds": 5.291266432000157,
      "full_seconds": 5.173491796000235,
      "geometry_seconds": 4.5208181810003225
     },
     "optimized": {
      "process_wall_seconds": 2.550777377000486,
      "full_seconds": 2.4328744889999143,
      "geometry_seconds": 1.8074414839993551
     }
    },
    "earlier_schema1_control": {
     "baseline": {
      "process_wall_seconds": 0.7533203149996552,
      "full_seconds": 0.6416048600003705,
      "geometry_seconds": 0.0
     },
     "optimized": {
      "process_wall_seconds": 0.7541534380006851,
      "full_seconds": 0.6430288149995249,
      "geometry_seconds": 0.0
     }
    }
   },
   "ratios_of_medians": {
    "supervised_schema2": {
     "process_wall_seconds": 2.07437406325999,
     "full_seconds": 2.1264935036278465,
     "geometry_seconds": 2.501225196512052
    },
    "earlier_schema1_control": {
     "process_wall_seconds": 0.9988952871404544,
     "full_seconds": 0.9977855502491666,
     "geometry_seconds": null
    }
   },
   "all_exact": true,
   "target_met": true,
   "rule_sha256": "582429ecdb0ca0492697af3440a46158c9c8117d64c155d4cec49e2de5035829",
   "process_amendment_sha256": "9dcd939d50a5583ad9cc4fa0a85f1d96c816b80704958bcc0012699846ea90b1"
  },
  "scoring": {
   "evidence": "docs/evidence/2026-10-07-production-speed/scoring/scoring-benchmark-summary.json",
   "sha256": "2b22ecb64e3baaf46ef09425f1cabb27c2ed9971247ac19c5dcd95cfef292ba1",
   "alternating_pairs": 6,
   "all_exact": true,
   "operations": {
    "auc_files": {
     "baseline_median_s": 0.10636817200020232,
     "candidate_median_s": 0.04063367850039867,
     "ratio_of_medians": 2.617734252121887,
     "paired_speedup_range": [
      2.3447765096028825,
      2.8664110383831893
     ]
    },
    "hp_files": {
     "baseline_median_s": 0.1751219709999532,
     "candidate_median_s": 0.17784679750002397,
     "ratio_of_medians": 0.9846787991778687,
     "paired_speedup_range": [
      0.867554158928713,
      1.0239235914389602
     ]
    },
    "rows_files": {
     "baseline_median_s": 0.021223528000064107,
     "candidate_median_s": 0.02078325650018087,
     "ratio_of_medians": 1.0211839516044758,
     "paired_speedup_range": [
      0.8049135405685044,
      1.2228809908883114
     ]
    },
    "full_pipeline": {
     "baseline_median_s": 0.2985465844999453,
     "candidate_median_s": 0.23521334950009987,
     "ratio_of_medians": 1.2692586757275806,
     "paired_speedup_range": [
      1.0834862474426779,
      1.318074300651249
     ]
    },
    "cli_process_wall_s": {
     "baseline_median_s": 0.3719989534997694,
     "candidate_median_s": 0.29812229849994765,
     "ratio_of_medians": 1.2478065390329558,
     "paired_speedup_range": [
      1.0592894161965047,
      1.3990237316319871
     ]
    },
    "worker_process_wall_s": {
     "baseline_median_s": 1.3394777639996391,
     "candidate_median_s": 1.148786525000105,
     "ratio_of_medians": 1.1659936244460358,
     "paired_speedup_range": [
      1.07119925884295,
      1.1957839323521549
     ]
    }
   },
   "scope": "full_pipeline here means sequential AUC/HP/row file scoring, not rendering/inference. Fresh CLI and warm file-operation timers are distinct."
  },
  "integration": {
   "repository_tests": 193,
   "experiment_regressions": 39,
   "actual_cli": "docs/evidence/2026-10-07-production-speed/integration/integrated-cli-check.json"
  },
  "limits": [
   "One real schema2 crop plus unchanged legacy case; larger-surface scaling unmeasured.",
   "Wall time includes warm local data/cache; no cold-storage benchmark.",
   "No measured peak-memory improvement; scoring adds O(pixels) temporary workspace.",
   "Geometry still rejects uncertain cases; zero real corrections accepted on the benchmark crop."
  ]
 },
 "production_scaling_checks": {
  "date": "2026-10-07",
  "log": "docs/logs/2026-10-07-production-scaling.md",
  "geometry_hp_baseline_commit": "c43c41a59ac3793b9a653fa236165f6f6b9d47a4",
  "row_comparison": "Dense and sparse modes in the same updated build; dense algorithm unchanged from PR24.",
  "geometry": {
   "evidence": {
    "path": "docs/evidence/2026-10-07-production-scaling/geometry/benchmark-result.json",
    "sha256": "c681f8230b1c109c3e083665cb96cd89fa25aef40e7a01e1e5bf13fd3d5f95e3"
   },
   "pairs_per_case": 5,
   "all_exact": true,
   "medians": {
    "actual_full_correct": {
     "baseline": {
      "process_wall_seconds": 2.513779447000161,
      "full_seconds": 2.4102645299999494,
      "geometry_seconds": 1.7768814190003468,
      "peak_rss_bytes": 84959232
     },
     "optimized": {
      "process_wall_seconds": 2.2761411589999625,
      "full_seconds": 2.160458063999613,
      "geometry_seconds": 1.5368202269992253,
      "peak_rss_bytes": 84955136
     }
    },
    "larger_geometry": {
     "baseline": {
      "process_wall_seconds": 7.423814374000358,
      "full_seconds": 7.285576299999775,
      "geometry_seconds": 7.285543480000342,
      "peak_rss_bytes": 69419008
     },
     "optimized": {
      "process_wall_seconds": 6.621332063999944,
      "full_seconds": 6.483136817999366,
      "geometry_seconds": 6.483104298999933,
      "peak_rss_bytes": 69431296
     }
    }
   },
   "ratios": {
    "actual_full_correct": {
     "process_wall_seconds": 1.1044040204011802,
     "full_seconds": 1.1156266211147254,
     "geometry_seconds": 1.1562064240069652,
     "peak_rss_bytes": 1.0000482136830433
    },
    "larger_geometry": {
     "process_wall_seconds": 1.1211965058154227,
     "full_seconds": 1.1237733375875343,
     "geometry_seconds": 1.1237739120014143,
     "peak_rss_bytes": 0.9998230192908973
    }
   }
  },
  "high_pass": {
   "evidence": {
    "path": "docs/evidence/2026-10-07-production-scaling/hp/hp-benchmark-summary.json",
    "sha256": "b4f5ed98fcf558c00e1b061917ecd3dc24bbbc6cb63a8cf8f8e7d04bd32c1a0a"
   },
   "pairs_per_case": 5,
   "all_exact": true,
   "cases": {
    "real640": {
     "hp_files": {
      "baseline_median": 0.18018964000020787,
      "candidate_median": 0.1371999399998458,
      "ratio_of_medians": 1.3133361428613626,
      "paired_ratios": [
       1.351545680010002,
       1.5613024671954447,
       1.3133361428613626,
       1.3795488206472235,
       1.2274548022665006
      ]
     },
     "cli_process_wall_s": {
      "baseline_median": 0.6501116549998187,
      "candidate_median": 0.6024023539994232,
      "ratio_of_medians": 1.0791983973562647,
      "paired_ratios": [
       1.066921659142475,
       1.1674431254075044,
       1.186106811397936,
       1.0871390485213306,
       1.0636054671578887
      ]
     },
     "cli_hp_operation_s": {
      "baseline_median": 0.20789358699948934,
      "candidate_median": 0.15321340199989208,
      "ratio_of_medians": 1.3568890468187358,
      "paired_ratios": [
       1.2932112528689632,
       1.3568890468187358,
       1.3962180782799787,
       1.2555005567327324,
       1.3161661551187223
      ]
     },
     "cli_peak_rss_kib": {
      "baseline_median": 88916,
      "candidate_median": 88600,
      "ratio_of_medians": 1.003566591422122,
      "paired_ratios": [
       1.0099045239582405,
       0.9988721465307228,
       1.0020316027088036,
       1.004471948685518,
       1.004065224264872
      ]
     },
     "auc_files": {
      "baseline_median": 0.04066087299997889,
      "candidate_median": 0.038315881000016816,
      "ratio_of_medians": 1.0612015680902926,
      "paired_ratios": [
       1.0725365083953302,
       1.348129382683101,
       1.014937357842189,
       1.0320942841125011,
       1.0424968859139783
      ]
     },
     "rows_files": {
      "baseline_median": 0.020770308000464865,
      "candidate_median": 0.019737038999664946,
      "ratio_of_medians": 1.0523517737801227,
      "paired_ratios": [
       0.8998853874064952,
       1.3542975215557833,
       1.0635062004659013,
       1.0288675641489178,
       0.9657737201998209
      ]
     },
     "full_pipeline": {
      "baseline_median": 0.2543045060001532,
      "candidate_median": 0.19648761399912473,
      "ratio_of_medians": 1.2942520946958316,
      "paired_ratios": [
       1.1039874327646282,
       1.3922845317053312,
       1.2942520946958316,
       1.2486435166406311,
       1.1670043412221807
      ]
     },
     "worker_process_wall_s": {
      "baseline_median": 1.2318850259998726,
      "candidate_median": 1.0863607210003465,
      "ratio_of_medians": 1.1339557866796988,
      "paired_ratios": [
       1.1410846895775602,
       1.2783739902017917,
       1.1219067879023186,
       1.2203695602518772,
       1.0445625880558858
      ]
     },
     "process_peak_rss_kib": {
      "baseline_median": 89684,
      "candidate_median": 89436,
      "ratio_of_medians": 1.002772932599848,
      "paired_ratios": [
       1.005153715156404,
       0.9952674345923743,
       1.0014759157386288,
       0.9981746950405128,
       1.0056820723904971
      ]
     }
    },
    "public2560": {
     "hp_files": {
      "baseline_median": 10.755387158999838,
      "candidate_median": 7.485103060999791,
      "ratio_of_medians": 1.4369056873831785,
      "paired_ratios": [
       1.4645958036323454,
       1.5138911033034341,
       1.395900871701165,
       1.4369056873831785,
       1.4528855654691133
      ]
     },
     "cli_process_wall_s": {
      "baseline_median": 11.276090796999597,
      "candidate_median": 7.9505747089997385,
      "ratio_of_medians": 1.4182736732522625,
      "paired_ratios": [
       1.4431490231440045,
       1.4856740356287568,
       1.3725533914379386,
       1.4182736732522625,
       1.4330759835235998
      ]
     },
     "cli_hp_operation_s": {
      "baseline_median": 10.755387158999838,
      "candidate_median": 7.485103060999791,
      "ratio_of_medians": 1.4369056873831785,
      "paired_ratios": [
       1.4645958036323454,
       1.5138911033034341,
       1.395900871701165,
       1.4369056873831785,
       1.4528855654691133
      ]
     },
     "cli_peak_rss_kib": {
      "baseline_median": 532248,
      "candidate_median": 525868,
      "ratio_of_medians": 1.0121323221797105,
      "paired_ratios": [
       1.0121477199254554,
       1.0121171092365384,
       1.0120862236826091,
       1.0122315784669451,
       1.0122307411464038
      ]
     }
    }
   },
   "large_control": "Synthetic rolled workload control; not a reverse-depth map or accuracy test."
  },
  "rows": {
   "evidence": {
    "path": "docs/evidence/2026-10-07-production-scaling/rows/benchmark-summary.json",
    "sha256": "a9779c63211e9557ec763fb8262ce53242bacfde250f5c85a7e24a9614f575e8"
   },
   "pairs_per_final_case": 3,
   "default_dense_unchanged": true,
   "sparse_opt_in": true,
   "universal_exactness": false,
   "cases": {
    "central8192-three-pairs": {
     "median": {
      "baseline": {
       "worker_operation_seconds": 5.279071661999296,
       "resize_seconds": 3.917248194000422,
       "parent_wall_seconds": 5.420129512999665,
       "peak_rss_bytes_max": 1660817408
      },
      "fast": {
       "worker_operation_seconds": 1.96820784800002,
       "resize_seconds": 0.4845775860003414,
       "parent_wall_seconds": 2.1101680850006233,
       "peak_rss_bytes_max": 1655980032
      }
     },
     "worker_speedup": 2.6821718383876916,
     "parent_speedup": 2.568577143937832,
     "resize_float_hashes_equal": true,
     "old_fields_equal_all_runs": true
    },
    "native-full-three-fresh-pairs": {
     "median": {
      "baseline": {
       "worker_operation_seconds": 39.73196149500018,
       "resize_seconds": 34.403317003000666,
       "parent_wall_seconds": 39.960150818999864,
       "peak_rss_bytes_max": 7097917440
      },
      "fast": {
       "worker_operation_seconds": 5.670121790999474,
       "resize_seconds": 0.38628741899992747,
       "parent_wall_seconds": 5.924856167000144,
       "peak_rss_bytes_max": 6536126464
      }
     },
     "worker_speedup": 7.007250101412128,
     "parent_speedup": 6.744492978845152,
     "resize_float_hashes_equal": true,
     "old_fields_equal_all_runs": true
    }
   },
   "timer_scope": "Fresh-worker cli.main operation including imports, reads, scoring and JSON construction; excludes interpreter startup/shutdown, hashing and evidence writes. Parent wall recorded separately.",
   "memory_scope": "Instrumented worker retains resize array; peak RSS through scoring, before hashing and archival."
  },
  "rejected_inference_layout": {
   "evidence": {
    "path": "docs/evidence/2026-10-07-production-scaling/inference-layout/screening/result.json",
    "sha256": "ae9428804d8a9cb1afa29e01f05356cc55432f7f9ce99a0845fae5bfba20b19c"
   },
   "pairs": 3,
   "baseline_median_seconds": 5.35897,
   "candidate_median_seconds": 6.97131,
   "shuffled_pixels_changed_per_pair": 1,
   "maximum_uint8_difference": 1,
   "default_unchanged": true
  },
  "integration": {
   "repository_tests": 229,
   "experiment_regressions": 39,
   "cli_equivalence": {
    "path": "docs/evidence/2026-10-07-production-scaling/integration/equivalence.json",
    "sha256": "a2b185bfc68750d3085c50f196daeaf0c9a137ebd286097bcf84aae2f4e0327b"
   },
   "rows_cli": {
    "path": "docs/evidence/2026-10-07-production-scaling/integration/rows-cli-check.json",
    "sha256": "a9b79eafccd349748bb8a533419dda787beacf179178635634c29e4adfdd3cc4"
   },
   "prior_seal_check": {
    "path": "docs/evidence/2026-10-07-production-scaling/integration/prior-seal-check.json",
    "sha256": "36047466e0db409cfc3431ceb1c18b0db398501f23dd97c524d7f80a3f837590"
   },
   "initial_performance_integration_tests": 210,
   "final_integrated_base": "a4410a34d9692c390e64a9c87b141e6247feaf8c",
   "final_cli_equivalence": {
    "path": "docs/evidence/2026-10-07-production-scaling/integration/after-merge/equivalence.json",
    "sha256": "2dc179664d21e77c905e9ea8e06ad0d46b3c32a0fe50cc998cf07c4b00f84f52"
   },
   "final_rows_cli": {
    "path": "docs/evidence/2026-10-07-production-scaling/integration/after-merge/rows-cli-check.json",
    "sha256": "a9b79eafccd349748bb8a533419dda787beacf179178635634c29e4adfdd3cc4"
   }
  },
  "limits": [
   "Public PHerc0841 workloads, cached local inputs and existing cloud CPUs only.",
   "Full native row map is 305,497,600 pixels; no full-canvas HP or complete larger-surface correction test.",
   "Sparse resize can change float32 rounding and near-tied FFT decisions on other inputs.",
   "No GPU, training, new text, improved correction accuracy or overall reading-pipeline speed claim."
  ]
 },
 "upstream_reports": [
  {
   "what": "VC3D caches an unconfirmed S3 404 as an empty chunk: w035 crop render, villa main 25f3c93, 6 false 404s replayed by scripts/false404/fault_relay.py (rate 0.3, seed 1, bad 1)",
   "measurement": "public data, no ink",
   "main": {
    "empty_markers": 12,
    "zero_share": 0.1918,
    "same_as_clean": false,
    "rerender_zero_share": 0.1918
   },
   "fix": {
    "empty_markers": 0,
    "zero_share": 0.0003,
    "same_as_clean": true,
    "rerender_zero_share": 0.0003
   },
   "tests": {
    "fix": "test_http_s3_missing 15 cases pass, test_s3_chunk_cache 3 cases pass",
    "main": "65 and 5 checks fail"
   },
   "machines": [
    "Linux x86_64, g++ 13.3",
    "Apple Silicon, macOS 27.0, Apple clang 21"
   ],
   "issue": "https://github.com/ScrollPrize/villa/issues/2058",
   "pr": "https://github.com/ScrollPrize/villa/pull/2060",
   "pr_state": "closed unmerged, 2026-10-30 (rendering moving to vc_reader)",
   "source": "docs/evidence/2026-10-12-false-404/README.md",
   "date": "2026-10-12"
  },
  {
   "what": "vc_grow_seg_from_seed does not hold an explicit seed: PHerc0139 scan 20250728140407, seed (4201, 3186, 4602) on published winding w031, villa e0bbb8b, scored against published windings w023 to w040",
   "measurement": "public geometry, no ink",
   "unanchored_10_generations": {
    "vertices": 324,
    "on_w031": 0,
    "on_w033": 265,
    "off": 58,
    "unscored": 1
   },
   "anchored_10_generations": {
    "vertices": 324,
    "share_on_w031_4vox": 0.9907,
    "switch_edges": 0
   },
   "unanchored_60_generations": {
    "area_cm2": 4.88,
    "dominant": "w033",
    "dominant_share_4vox": 0.511,
    "area_on_w031_cm2": 0.0
   },
   "anchored_60_generations": {
    "area_cm2": 4.88,
    "dominant": "w031",
    "dominant_share_4vox": 0.972,
    "area_on_w031_cm2": 4.74,
    "switch_edges": 30,
    "edges": 26612
   },
   "issue": "https://github.com/ScrollPrize/villa/issues/2059",
   "source": "docs/logs/2026-10-12-upstream-reports.md",
   "date": "2026-10-12"
  },
  {
   "what": "PHerc1667 transform 20260323082859 (1.129 um) to 20251217075048 (2.399 um): CT check at the six landmarks",
   "measurement": "public CT and geometry, no ink",
   "ct_offset_from_published_matrix_um": [
    61.0,
    73.7,
    215.9,
    43.8,
    42.3,
    181.8
   ],
   "landmark_residual_um": [
    60.2,
    76.3,
    212.5,
    43.8,
    41.9,
    181.4
   ],
   "landmark_to_ct_um": "3.3 to 5.7 (4.8 um, 8 October); median 6.8 (9.6 um, 12 October catalogue)",
   "control_swapped_pairs_matched": "0 of 6",
   "other_transforms_passing": "14 of 14 landmarks (median 13.4 um) and 16 of 16 (median 3.8 um)",
   "comment": "https://github.com/ScrollPrize/villa/issues/1843#issuecomment-6130119937",
   "prior_report": "https://github.com/ScrollPrize/villa/issues/1843",
   "source": "docs/logs/2026-10-12-upstream-reports.md",
   "date": "2026-10-12"
  }
 ],
 "letter_reader_checks": [
  {
   "date": "2026-10-08",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "claim": "Model output on synthetic images and public community drawings; not a reading, no hidden text",
   "model": {
    "format": "linenet-v1",
    "params": 593849,
    "channels": [
     32,
     64,
     96,
     128,
     192
    ],
    "dilations": [
     1,
     2
    ],
    "train_lines": 36000,
    "val_lines": 1200,
    "steps": 4000,
    "batch": 48,
    "lr_max": 0.002,
    "device": "4 CPU threads",
    "seconds": 1578,
    "weights_in_git": false
   },
   "synthetic_validation": [
    {
     "step": 500,
     "val_cer": 0.206,
     "letters_per_blank_line": 0.69
    },
    {
     "step": 1000,
     "val_cer": 0.207,
     "letters_per_blank_line": 0.5
    },
    {
     "step": 1500,
     "val_cer": 0.15,
     "letters_per_blank_line": 0.6
    },
    {
     "step": 2000,
     "val_cer": 0.14,
     "letters_per_blank_line": 0.79
    },
    {
     "step": 2500,
     "val_cer": 0.12,
     "letters_per_blank_line": 0.94
    },
    {
     "step": 3000,
     "val_cer": 0.119,
     "letters_per_blank_line": 0.62
    },
    {
     "step": 3500,
     "val_cer": 0.111,
     "letters_per_blank_line": 0.88
    },
    {
     "step": 4000,
     "val_cer": 0.112,
     "letters_per_blank_line": 0.9
    }
   ],
   "language_model": {
    "order": 5,
    "discount": 0.75,
    "corpus": "PerseusDL canonical-greekLit tlg0004.tlg001 and tlg0059.tlg030 (CC BY-SA 4.0)",
    "train_letters": 937054,
    "bits_per_letter_held_out_chunks": 2.556,
    "bits_per_letter_xenophon_anabasis": 2.826,
    "uniform_bits": 4.585
   },
   "synthetic_pages": [
    {
     "text": "uniform random letters",
     "pages": 18,
     "seed": 11,
     "cer_matched_lines": {
      "map": {
       "cer_matched_lines": 0.158,
       "letters": 672
      },
      "mask": {
       "cer_matched_lines": 0.013,
       "letters": 306
      }
     },
     "cer_with_lm": null,
     "lines_found": "81 of 81",
     "letter_px_ratio_median": 1.036,
     "blank_pages": [
      {
       "kind": "texture",
       "letters_kept": 0,
       "null_kept": [
        2,
        0
       ]
      },
      {
       "kind": "clutter",
       "letters_kept": 0,
       "null_kept": [
        11,
        4
       ]
      },
      {
       "kind": "texture",
       "letters_kept": 0,
       "null_kept": [
        10,
        0
       ]
      },
      {
       "kind": "clutter",
       "letters_kept": 0,
       "null_kept": [
        0,
        0
       ]
      },
      {
       "kind": "texture",
       "letters_kept": 1,
       "null_kept": [
        0,
        0
       ]
      },
      {
       "kind": "clutter",
       "letters_kept": 1,
       "null_kept": [
        20,
        0
       ]
      }
     ]
    },
    {
     "text": "real Greek",
     "pages": 18,
     "seed": 12,
     "cer_matched_lines": {
      "map": {
       "cer_matched_lines": 0.138,
       "letters": 659
      },
      "mask": {
       "cer_matched_lines": 0.019,
       "letters": 260
      }
     },
     "cer_with_lm": {
      "map": 0.108,
      "mask": 0.008
     },
     "lines_found": "76 of 77",
     "letter_px_ratio_median": 0.981,
     "blank_pages": [
      {
       "kind": "texture",
       "letters_kept": 0,
       "null_kept": [
        0,
        0
       ]
      },
      {
       "kind": "clutter",
       "letters_kept": 0,
       "null_kept": [
        0,
        0
       ]
      },
      {
       "kind": "texture",
       "letters_kept": 10,
       "null_kept": [
        88,
        11
       ]
      },
      {
       "kind": "clutter",
       "letters_kept": 0,
       "null_kept": [
        0,
        0
       ]
      },
      {
       "kind": "texture",
       "letters_kept": 0,
       "null_kept": [
        44,
        0
       ]
      },
      {
       "kind": "clutter",
       "letters_kept": 1,
       "null_kept": [
        1,
        0
       ]
      }
     ]
    }
   ],
   "pherc172_drawings": {
    "data": "Bodillium/Herculaneum-Scroll-Labels commit 2266661c, PHerc. 172 hand-drawn ink labels (CC BY-NC-SA 4.0)",
    "reference": "the labeller's preliminary transcription, marked speculative by its author",
    "down": 4,
    "segments": 10,
    "split": "segments sorted by id; first five dev (checkpoint choice), last five test (report); rule fixed after the all-segment numbers were seen",
    "chosen_on_dev": 1000,
    "test_cer_ink_only": 0.76,
    "test_cer_with_lm": 0.67,
    "checkpoints": [
     {
      "step": 500,
      "note": "letter size from pitch only, before the scale search",
      "dev": {
       "cer_ink_only": 0.745,
       "ref_letters": 161,
       "letters_kept": 122,
       "null_letters_kept": [
        45,
        3
       ]
      },
      "test": {
       "cer_ink_only": 1.113,
       "ref_letters": 80,
       "letters_kept": 53,
       "null_letters_kept": [
        20,
        2
       ]
      },
      "all": {
       "cer_ink_only": 0.867,
       "ref_letters": 241,
       "letters_kept": 175,
       "null_letters_kept": [
        65,
        5
       ]
      }
     },
     {
      "step": 1000,
      "dev": {
       "cer_ink_only": 0.74,
       "ref_letters": 169,
       "letters_kept": 122,
       "null_letters_kept": [
        20,
        2
       ]
      },
      "test": {
       "cer_ink_only": 0.76,
       "ref_letters": 100,
       "letters_kept": 94,
       "null_letters_kept": [
        13,
        3
       ],
       "cer_with_lm": 0.67,
       "ref_letters_with_lm": 94
      },
      "all": {
       "cer_ink_only": 0.747,
       "ref_letters": 269,
       "letters_kept": 216,
       "null_letters_kept": [
        33,
        5
       ]
      }
     },
     {
      "step": 4000,
      "dev": {
       "cer_ink_only": 0.886,
       "ref_letters": 167,
       "letters_kept": 173,
       "null_letters_kept": [
        58,
        23
       ]
      },
      "test": {
       "cer_ink_only": 1.102,
       "ref_letters": 98,
       "letters_kept": 120,
       "null_letters_kept": [
        23,
        9
       ]
      },
      "all": {
       "cer_ink_only": 0.966,
       "ref_letters": 265,
       "letters_kept": 293,
       "null_letters_kept": [
        81,
        32
       ],
       "cer_with_lm": 0.851
      }
     }
    ]
   }
  },
  {
   "date": "2026-10-08",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "claim": "Model output on synthetic lines, lines composed from public papyrus letter photographs and public community drawings; not a reading, no hidden text",
   "check": "on CPU: real papyrus letterforms mixed into training, same network and schedule as the prototype",
   "model": {
    "format": "linenet-v1",
    "params": 593849,
    "channels": [
     32,
     64,
     96,
     128,
     192
    ],
    "dilations": [
     1,
     2
    ],
    "train_lines": 60000,
    "train_lines_real_letterforms": 24000,
    "steps": 4000,
    "batch": 48,
    "lr_max": 0.002,
    "device": "4 CPU threads",
    "seconds": 1528,
    "weights_in_git": false
   },
   "bank": {
    "source": "Hell-Char (github.com/ipavlopoulos/diachronic-greek-letterforms at aeef8466, CC BY 4.0)",
    "glyphs": 12761,
    "papyri": 144,
    "val_papyri": 21,
    "val_glyphs": 1904,
    "icdar_2023": "not included (Zenodo unreachable from this session)"
   },
   "validation": [
    {
     "step": 500,
     "val_cer": 0.373,
     "letters_per_blank_line": 0.65,
     "real_val_cer": 0.747
    },
    {
     "step": 1000,
     "val_cer": 0.202,
     "letters_per_blank_line": 0.67,
     "real_val_cer": 0.644
    },
    {
     "step": 1500,
     "val_cer": 0.18,
     "letters_per_blank_line": 0.67,
     "real_val_cer": 0.598
    },
    {
     "step": 2000,
     "val_cer": 0.158,
     "letters_per_blank_line": 0.75,
     "real_val_cer": 0.564
    },
    {
     "step": 2500,
     "val_cer": 0.154,
     "letters_per_blank_line": 0.65,
     "real_val_cer": 0.542
    },
    {
     "step": 3000,
     "val_cer": 0.138,
     "letters_per_blank_line": 0.71,
     "real_val_cer": 0.528
    },
    {
     "step": 3500,
     "val_cer": 0.13,
     "letters_per_blank_line": 0.81,
     "real_val_cer": 0.514
    },
    {
     "step": 4000,
     "val_cer": 0.128,
     "letters_per_blank_line": 0.75,
     "real_val_cer": 0.516
    }
   ],
   "real_letter_val_prototype": {
    "note": "same 600 real-letter val lines (5,839 letters, held-out papyri), greedy decode",
    "prototype_step_500": 0.852,
    "prototype_step_1000": 0.909,
    "prototype_step_4000": 0.821,
    "mixed_step_1000": 0.644,
    "mixed_step_3000": 0.527,
    "mixed_step_4000": 0.516
   },
   "p172_dev": [
    {
     "step": 500,
     "cer_ink_only": 0.982,
     "ref_letters": 167,
     "letters_kept": 158,
     "null_letters_kept": [
      42,
      26
     ]
    },
    {
     "step": 1000,
     "cer_ink_only": 0.732,
     "ref_letters": 164,
     "letters_kept": 152,
     "null_letters_kept": [
      42,
      9
     ]
    },
    {
     "step": 1500,
     "cer_ink_only": 0.695,
     "ref_letters": 167,
     "letters_kept": 141,
     "null_letters_kept": [
      43,
      15
     ]
    },
    {
     "step": 2000,
     "cer_ink_only": 0.579,
     "ref_letters": 164,
     "letters_kept": 169,
     "null_letters_kept": [
      44,
      6
     ]
    },
    {
     "step": 2500,
     "cer_ink_only": 0.591,
     "ref_letters": 164,
     "letters_kept": 172,
     "null_letters_kept": [
      43,
      6
     ]
    },
    {
     "step": 3000,
     "cer_ink_only": 0.549,
     "ref_letters": 164,
     "letters_kept": 165,
     "null_letters_kept": [
      40,
      10
     ]
    },
    {
     "step": 3500,
     "cer_ink_only": 0.573,
     "ref_letters": 164,
     "letters_kept": 163,
     "null_letters_kept": [
      40,
      9
     ]
    },
    {
     "step": 4000,
     "cer_ink_only": 0.622,
     "ref_letters": 164,
     "letters_kept": 159,
     "null_letters_kept": [
      36,
      15
     ]
    }
   ],
   "p172_prototype_dev": {
    "step_500": 0.745,
    "step_1000": 0.74,
    "step_4000": 0.886
   },
   "p172_test_chosen_on_dev": {
    "chosen_step": 3000,
    "cer_ink_only": 0.867,
    "cer_with_lm": 0.777,
    "ref_letters": 98,
    "letters_kept": 119,
    "null_letters_kept": [
     22,
     10
    ],
    "prototype": {
     "cer_ink_only": 0.76,
     "cer_with_lm": 0.67,
     "letters_kept": 94,
     "null_letters_kept": [
      13,
      3
     ]
    }
   },
   "verdict": "null on the gate: better on the dev half (0.549 against 0.740), worse on the test half (0.867 against 0.760); paired bootstrap over lines on the test half: +0.107, 95% interval -0.093 to +0.338, so the halves cannot separate the two models",
   "p172_paired_bootstrap": {
    "tool": "scripts/letters/compare_p172.py",
    "resamples": 20000,
    "diff": "mixed step 3000 minus prototype step 1000",
    "test_ink_only": {
     "diff": 0.107,
     "ci95_lines": [
      -0.093,
      0.338
     ],
     "ci95_segments": [
      -0.088,
      0.364
     ],
     "lines": 19,
     "share_mixed_worse": 0.839
    },
    "test_with_lm": {
     "diff": 0.106,
     "ci95_lines": [
      -0.118,
      0.368
     ],
     "ci95_segments": [
      -0.035,
      0.316
     ],
     "lines": 22
    },
    "dev_ink_only_selection_biased": {
     "diff": -0.191,
     "ci95_lines": [
      -0.293,
      -0.07
     ],
     "ci95_segments": [
      -0.214,
      0.005
     ],
     "lines": 25
    },
    "pooled_ink_only": {
     "diff": -0.079,
     "ci95_lines": [
      -0.191,
      0.049
     ],
     "ci95_segments": [
      -0.182,
      0.148
     ],
     "lines": 44
    }
   }
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "claim": "Model output on synthetic lines and public community drawings; not a reading, no hidden text",
   "check": "on Modal (one B200): full-scale synthetic training (letters-v1)",
   "run": {
    "gpu": "B200",
    "wall_seconds": 1670,
    "train_seconds": 472.5,
    "est_cost_usd": 2.4
   },
   "model": {
    "format": "linenet-v1",
    "params": 1483241,
    "channels": [
     48,
     96,
     160,
     192,
     256
    ],
    "dilations": [
     1,
     2,
     4
    ],
    "train_lines": 600000,
    "val_lines": 4000,
    "steps": 30000,
    "batch": 128,
    "lr_max": 0.002,
    "weights_in_git": false
   },
   "validation": [
    {
     "step": 2000,
     "val_cer": 0.1355,
     "letters_per_blank_line": 0.955
    },
    {
     "step": 4000,
     "val_cer": 0.1214,
     "letters_per_blank_line": 0.886
    },
    {
     "step": 6000,
     "val_cer": 0.0947,
     "letters_per_blank_line": 0.477
    },
    {
     "step": 8000,
     "val_cer": 0.089,
     "letters_per_blank_line": 0.659
    },
    {
     "step": 10000,
     "val_cer": 0.0885,
     "letters_per_blank_line": 0.318
    },
    {
     "step": 12000,
     "val_cer": 0.0815,
     "letters_per_blank_line": 0.477
    },
    {
     "step": 14000,
     "val_cer": 0.0778,
     "letters_per_blank_line": 0.432
    },
    {
     "step": 16000,
     "val_cer": 0.0789,
     "letters_per_blank_line": 0.409
    },
    {
     "step": 18000,
     "val_cer": 0.0774,
     "letters_per_blank_line": 0.318
    },
    {
     "step": 20000,
     "val_cer": 0.0767,
     "letters_per_blank_line": 0.455
    },
    {
     "step": 22000,
     "val_cer": 0.0756,
     "letters_per_blank_line": 0.295
    },
    {
     "step": 24000,
     "val_cer": 0.0749,
     "letters_per_blank_line": 0.386
    },
    {
     "step": 26000,
     "val_cer": 0.0751,
     "letters_per_blank_line": 0.386
    },
    {
     "step": 28000,
     "val_cer": 0.074,
     "letters_per_blank_line": 0.386
    },
    {
     "step": 30000,
     "val_cer": 0.0742,
     "letters_per_blank_line": 0.386
    }
   ],
   "p172_dev": [
    {
     "step": 2000,
     "cer_ink_only": 0.958,
     "ref_letters": 168,
     "letters_kept": 169,
     "null_letters_kept": [
      45,
      19
     ]
    },
    {
     "step": 4000,
     "cer_ink_only": 0.79,
     "ref_letters": 167,
     "letters_kept": 129,
     "null_letters_kept": [
      26,
      8
     ]
    },
    {
     "step": 6000,
     "cer_ink_only": 0.641,
     "ref_letters": 167,
     "letters_kept": 137,
     "null_letters_kept": [
      21,
      4
     ]
    },
    {
     "step": 8000,
     "cer_ink_only": 0.689,
     "ref_letters": 164,
     "letters_kept": 175,
     "null_letters_kept": [
      36,
      7
     ]
    },
    {
     "step": 10000,
     "cer_ink_only": 0.652,
     "ref_letters": 164,
     "letters_kept": 145,
     "null_letters_kept": [
      27,
      1
     ]
    },
    {
     "step": 12000,
     "cer_ink_only": 0.634,
     "ref_letters": 164,
     "letters_kept": 170,
     "null_letters_kept": [
      41,
      7
     ]
    },
    {
     "step": 14000,
     "cer_ink_only": 0.689,
     "ref_letters": 164,
     "letters_kept": 151,
     "null_letters_kept": [
      23,
      4
     ]
    },
    {
     "step": 16000,
     "cer_ink_only": 0.726,
     "ref_letters": 164,
     "letters_kept": 166,
     "null_letters_kept": [
      33,
      6
     ]
    },
    {
     "step": 18000,
     "cer_ink_only": 0.744,
     "ref_letters": 164,
     "letters_kept": 162,
     "null_letters_kept": [
      32,
      4
     ]
    },
    {
     "step": 20000,
     "cer_ink_only": 0.726,
     "ref_letters": 164,
     "letters_kept": 166,
     "null_letters_kept": [
      38,
      5
     ]
    },
    {
     "step": 22000,
     "cer_ink_only": 0.701,
     "ref_letters": 164,
     "letters_kept": 156,
     "null_letters_kept": [
      31,
      7
     ]
    },
    {
     "step": 24000,
     "cer_ink_only": 0.726,
     "ref_letters": 164,
     "letters_kept": 165,
     "null_letters_kept": [
      30,
      4
     ]
    },
    {
     "step": 26000,
     "cer_ink_only": 0.726,
     "ref_letters": 164,
     "letters_kept": 155,
     "null_letters_kept": [
      33,
      5
     ]
    },
    {
     "step": 28000,
     "cer_ink_only": 0.738,
     "ref_letters": 164,
     "letters_kept": 160,
     "null_letters_kept": [
      35,
      6
     ]
    },
    {
     "step": 30000,
     "cer_ink_only": 0.732,
     "ref_letters": 164,
     "letters_kept": 154,
     "null_letters_kept": [
      35,
      6
     ]
    }
   ],
   "p172_test_chosen_on_dev": {
    "chosen_step": 12000,
    "cer_ink_only": 0.933,
    "cer_with_lm": 0.831,
    "ref_letters": 90,
    "letters_kept": 111,
    "null_letters_kept": [
     17,
     6
    ],
    "verdicts": [
     "confident letters well above every null"
    ]
   },
   "spec_checks": {
    "1_training_sound": true,
    "2_synthetic_below_0.112": true,
    "3_test_below_0.760": false
   },
   "p172_paired_bootstrap": {
    "tool": "scripts/letters/compare_p172.py",
    "resamples": 20000,
    "diff": "letters-v1 step 12000 minus prototype step 1000, test half",
    "test_ink_only": {
     "diff": 0.173,
     "ci95_lines": [
      -0.046,
      0.444
     ],
     "ci95_segments_5_units": [
      0.06,
      0.332
     ],
     "lines": 21,
     "share_v1_worse": 0.938
    },
    "test_with_lm": {
     "diff": 0.161,
     "ci95_lines": [
      -0.069,
      0.458
     ],
     "ci95_segments_5_units": [
      0.062,
      0.342
     ],
     "lines": 22
    },
    "against_cpu_real_letterforms_step_3000_ink_only": {
     "diff": 0.066,
     "ci95_lines": [
      -0.221,
      0.36
     ],
     "lines": 20
    }
   },
   "verdict": "null on the gate: synthetic error 0.074 against the prototype's 0.112 and dev half 0.634 against 0.740, but test half 0.933 against 0.760 (paired bootstrap over lines +0.173, 95% interval -0.046 to +0.444); scale does not close the gap to drawn letters"
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "claim": "Model output on synthetic lines, public papyrus photographs and public community drawings; not a reading, no hidden text",
   "check": "on Modal (one B200): real letterforms (letters-v2-real)",
   "run": {
    "gpu": "B200",
    "wall_seconds": 1460,
    "train_seconds": 468.9,
    "est_cost_usd": 2.24
   },
   "bank": {
    "glyphs": 43806,
    "papyri": 296,
    "papyri_note": "144 Hell-Char papyri plus 152 ICDAR photographs, each photograph counted as its own id; several photographs may show one papyrus",
    "per_source": {
     "hellchar": 12761,
     "icdar": 31045
    },
    "icdar_included": true,
    "icdar_skipped": 2,
    "icdar_skipped_first": [
     "P.Corn.Inv.MSS.A.101.XIII.jpg: OSError('image file is truncated (3 bytes not processed)')",
     "PSI_XIII_1298_15a_r_1.jpg: OSError('image file is truncated (6 bytes not processed)')"
    ],
    "icdar_passed_over": {
     "boxes_not_a_letter": 1,
     "photographs_not_found": 1,
     "boxes_too_small": 846,
     "json_unreadable": 3
    },
    "val_papyri": 21,
    "val_glyphs": 1904,
    "dropped": {
     "flat": 189,
     "size": 130,
     "texture": 3508
    }
   },
   "model": {
    "format": "linenet-v1",
    "channels": [
     48,
     96,
     160,
     192,
     256
    ],
    "dilations": [
     1,
     2,
     4
    ],
    "train_lines": 600000,
    "p_bank": 0.5,
    "val_lines": 4000,
    "val_real_lines": 2000,
    "steps": 30000,
    "batch": 128,
    "lr_max": 0.002,
    "weights_in_git": false,
    "params": 1483241
   },
   "validation": [
    {
     "step": 2000,
     "val_cer": 0.191,
     "letters_per_blank_line": 0.409,
     "real_val_cer": 0.6213
    },
    {
     "step": 4000,
     "val_cer": 0.1246,
     "letters_per_blank_line": 0.477,
     "real_val_cer": 0.5267
    },
    {
     "step": 6000,
     "val_cer": 0.1061,
     "letters_per_blank_line": 0.864,
     "real_val_cer": 0.5003
    },
    {
     "step": 8000,
     "val_cer": 0.1098,
     "letters_per_blank_line": 0.5,
     "real_val_cer": 0.4947
    },
    {
     "step": 10000,
     "val_cer": 0.0899,
     "letters_per_blank_line": 0.636,
     "real_val_cer": 0.478
    },
    {
     "step": 12000,
     "val_cer": 0.1027,
     "letters_per_blank_line": 0.636,
     "real_val_cer": 0.4869
    },
    {
     "step": 14000,
     "val_cer": 0.0913,
     "letters_per_blank_line": 0.591,
     "real_val_cer": 0.4774
    },
    {
     "step": 16000,
     "val_cer": 0.0888,
     "letters_per_blank_line": 0.409,
     "real_val_cer": 0.4643
    },
    {
     "step": 18000,
     "val_cer": 0.0858,
     "letters_per_blank_line": 0.409,
     "real_val_cer": 0.4575
    },
    {
     "step": 20000,
     "val_cer": 0.0856,
     "letters_per_blank_line": 0.477,
     "real_val_cer": 0.4524
    },
    {
     "step": 22000,
     "val_cer": 0.0869,
     "letters_per_blank_line": 0.5,
     "real_val_cer": 0.4541
    },
    {
     "step": 24000,
     "val_cer": 0.0858,
     "letters_per_blank_line": 0.409,
     "real_val_cer": 0.4434
    },
    {
     "step": 26000,
     "val_cer": 0.0849,
     "letters_per_blank_line": 0.386,
     "real_val_cer": 0.4456
    },
    {
     "step": 28000,
     "val_cer": 0.084,
     "letters_per_blank_line": 0.364,
     "real_val_cer": 0.4473
    },
    {
     "step": 30000,
     "val_cer": 0.0842,
     "letters_per_blank_line": 0.364,
     "real_val_cer": 0.4444
    }
   ],
   "p172_dev": {
    "checkpoints": 15,
    "cer_ink_only_range": [
     0.482,
     0.599
    ],
    "chosen_step": 30000,
    "chosen_cer_ink_only": 0.482,
    "per_step": [
     {
      "step": 2000,
      "cer_ink_only": 0.512,
      "ref_letters": 164,
      "letters_kept": 162,
      "null_letters_kept": [
       30,
       11
      ]
     },
     {
      "step": 4000,
      "cer_ink_only": 0.599,
      "ref_letters": 167,
      "letters_kept": 178,
      "null_letters_kept": [
       18,
       25
      ]
     },
     {
      "step": 6000,
      "cer_ink_only": 0.561,
      "ref_letters": 164,
      "letters_kept": 188,
      "null_letters_kept": [
       37,
       12
      ]
     },
     {
      "step": 8000,
      "cer_ink_only": 0.549,
      "ref_letters": 164,
      "letters_kept": 176,
      "null_letters_kept": [
       32,
       11
      ]
     },
     {
      "step": 10000,
      "cer_ink_only": 0.549,
      "ref_letters": 164,
      "letters_kept": 183,
      "null_letters_kept": [
       32,
       17
      ]
     },
     {
      "step": 12000,
      "cer_ink_only": 0.555,
      "ref_letters": 164,
      "letters_kept": 184,
      "null_letters_kept": [
       38,
       12
      ]
     },
     {
      "step": 14000,
      "cer_ink_only": 0.524,
      "ref_letters": 164,
      "letters_kept": 184,
      "null_letters_kept": [
       33,
       14
      ]
     },
     {
      "step": 16000,
      "cer_ink_only": 0.543,
      "ref_letters": 164,
      "letters_kept": 180,
      "null_letters_kept": [
       35,
       14
      ]
     },
     {
      "step": 18000,
      "cer_ink_only": 0.518,
      "ref_letters": 164,
      "letters_kept": 175,
      "null_letters_kept": [
       33,
       17
      ]
     },
     {
      "step": 20000,
      "cer_ink_only": 0.5,
      "ref_letters": 164,
      "letters_kept": 179,
      "null_letters_kept": [
       35,
       15
      ]
     },
     {
      "step": 22000,
      "cer_ink_only": 0.533,
      "ref_letters": 167,
      "letters_kept": 171,
      "null_letters_kept": [
       36,
       27
      ]
     },
     {
      "step": 24000,
      "cer_ink_only": 0.494,
      "ref_letters": 164,
      "letters_kept": 173,
      "null_letters_kept": [
       33,
       18
      ]
     },
     {
      "step": 26000,
      "cer_ink_only": 0.5,
      "ref_letters": 164,
      "letters_kept": 169,
      "null_letters_kept": [
       34,
       17
      ]
     },
     {
      "step": 28000,
      "cer_ink_only": 0.488,
      "ref_letters": 164,
      "letters_kept": 169,
      "null_letters_kept": [
       34,
       17
      ]
     },
     {
      "step": 30000,
      "cer_ink_only": 0.482,
      "ref_letters": 164,
      "letters_kept": 170,
      "null_letters_kept": [
       34,
       17
      ]
     }
    ]
   },
   "p172_test_chosen_on_dev": {
    "chosen_step": 30000,
    "cer_ink_only": 0.806,
    "cer_with_lm": 0.761,
    "ref_letters": 93,
    "ref_letters_with_lm": 88,
    "letters_kept": 121,
    "null_letters_kept": [
     19,
     9
    ],
    "verdicts": [
     "confident letters well above every null"
    ]
   },
   "spec_checks": {
    "1_bank_at_least_12000_and_icdar_reported": true,
    "1_params_1483241_and_loss_falls": true,
    "1_no_error": true,
    "2_real_val_cer_falls": true,
    "3_test_below_0.760": false,
    "beats_v1_0.933": true
   },
   "p172_paired_bootstrap": {
    "tool": "scripts/letters/compare_p172.py",
    "resamples": 20000,
    "scorer": "as recorded (columns off), test half",
    "v2_minus_v1": {
     "ink_only": {
      "diff": -0.127,
      "ci95_lines": [
       -0.482,
       0.199
      ],
      "lines": 19,
      "share_worse": 0.22
     },
     "with_lm": {
      "diff": -0.07,
      "ci95_lines": [
       -0.323,
       0.167
      ],
      "lines": 20
     }
    },
    "v2_minus_prototype": {
     "ink_only": {
      "diff": 0.046,
      "ci95_lines": [
       -0.187,
       0.283
      ],
      "lines": 21,
      "share_worse": 0.666
     },
     "with_lm": {
      "diff": 0.091,
      "ci95_lines": [
       -0.125,
       0.357
      ],
      "lines": 20
     }
    },
    "v2_minus_cpu_real_letterforms": {
     "ink_only": {
      "diff": -0.061,
      "ci95_lines": [
       -0.298,
       0.146
      ],
      "lines": 19
     },
     "with_lm": {
      "diff": -0.015,
      "ci95_lines": [
       -0.199,
       0.178
      ],
      "lines": 21
     }
    }
   },
   "post_hoc_layout_split": {
    "note": "Interpretation, found after the scores were seen: test segments 20241113070770, 20241113080880 and 20241113090990 are two-column strips read as single lines across both columns; errors over matched reference letters, ink only",
    "single_column_test_segments": [
     "20241108115232",
     "20241108120732"
    ],
    "single_column": {
     "prototype": [
      34,
      54
     ],
     "cpu_real_letterforms_step3000": [
      28,
      52
     ],
     "letters_v1": [
      40,
      52
     ],
     "letters_v2_real": [
      25,
      52
     ]
    },
    "two_column_strips": {
     "prototype": [
      42,
      46
     ],
     "cpu_real_letterforms_step3000": [
      57,
      46
     ],
     "letters_v1": [
      44,
      38
     ],
     "letters_v2_real": [
      50,
      41
     ]
    }
   },
   "verdict": "Gate failed: 0.806 on the test half against the prototype's 0.760, null control holding; better than letters-v1 (0.933). Post hoc, the test half's prototype lead came from two-column strips read as one line; the 2026-10-09 column re-score (next entry) reads the test half at 0.426 against 0.625; next is a fresh test set"
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "claim": "Model output on public community drawings and synthetic pages; not a reading, no hidden text",
   "check": "diagnostic re-score with the reader's column step (kit/letterread.py find_columns, analyze(columns='auto')); scorer unchanged; checkpoints already chosen on the dev half; rule written before the re-score, gutter threshold set after the test strips had been seen, so not a gate",
   "p172": {
    "prototype_step1000": {
     "off": {
      "dev_ink_only": 0.74,
      "dev_with_lm": 0.707,
      "test_ink_only": 0.76,
      "test_with_lm": 0.67,
      "test_ref_letters": 100
     },
     "auto": {
      "dev_ink_only": 0.639,
      "dev_with_lm": 0.686,
      "test_ink_only": 0.625,
      "test_with_lm": 0.622,
      "test_ref_letters": 120
     }
    },
    "cpu_real_letterforms_step3000": {
     "off": {
      "dev_ink_only": 0.549,
      "dev_with_lm": 0.5,
      "test_ink_only": 0.867,
      "test_with_lm": 0.777,
      "test_ref_letters": 98
     },
     "auto": {
      "dev_ink_only": 0.479,
      "dev_with_lm": 0.462,
      "test_ink_only": 0.525,
      "test_with_lm": 0.508,
      "test_ref_letters": 122
     }
    },
    "letters_v1_step12000": {
     "off": {
      "dev_ink_only": 0.634,
      "dev_with_lm": 0.585,
      "test_ink_only": 0.933,
      "test_with_lm": 0.831,
      "test_ref_letters": 90
     },
     "auto": {
      "dev_ink_only": 0.612,
      "dev_with_lm": 0.594,
      "test_ink_only": 0.739,
      "test_with_lm": 0.631,
      "test_ref_letters": 111
     }
    },
    "letters_v2_real_step30000": {
     "off": {
      "dev_ink_only": 0.482,
      "dev_with_lm": 0.409,
      "test_ink_only": 0.806,
      "test_with_lm": 0.761,
      "test_ref_letters": 93
     },
     "auto": {
      "dev_ink_only": 0.471,
      "dev_with_lm": 0.4,
      "test_ink_only": 0.426,
      "test_with_lm": 0.451,
      "test_ref_letters": 122
     }
    },
    "off_reproduces_recorded": "all four, including both Modal runs' dev and test numbers on CPU"
   },
   "p172_paired_bootstrap_auto": {
    "tool": "scripts/letters/compare_p172.py",
    "resamples": 20000,
    "both_with_columns": true,
    "v2_minus_prototype": {
     "test_ink_only": [
      -0.199,
      -0.337,
      -0.042
     ],
     "test_with_lm": [
      -0.171,
      -0.329,
      0.007
     ],
     "dev_ink_only": [
      -0.168,
      -0.282,
      -0.034
     ],
     "dev_with_lm": [
      -0.286,
      -0.407,
      -0.148
     ],
     "test_ink_only_segment_level_ci95": [
      -0.456,
      -0.019
     ]
    },
    "v2_minus_v1": {
     "test_ink_only": [
      -0.313,
      -0.558,
      -0.105
     ],
     "test_with_lm": [
      -0.18,
      -0.322,
      -0.058
     ],
     "dev_ink_only": [
      -0.141,
      -0.252,
      -0.012
     ],
     "dev_with_lm": [
      -0.194,
      -0.306,
      -0.072
     ]
    },
    "v2_minus_cpu_real_letterforms": {
     "test_ink_only": [
      -0.098,
      -0.204,
      0.009
     ],
     "test_with_lm": [
      -0.057,
      -0.146,
      0.038
     ],
     "dev_ink_only": [
      -0.009,
      -0.109,
      0.09
     ],
     "dev_with_lm": [
      -0.062,
      -0.163,
      0.04
     ]
    },
    "cpu_real_letterforms_minus_prototype": {
     "test_ink_only": [
      -0.1,
      -0.204,
      0.022
     ],
     "test_with_lm": [
      -0.114,
      -0.237,
      0.034
     ],
     "dev_ink_only": [
      -0.16,
      -0.267,
      -0.043
     ],
     "dev_with_lm": [
      -0.225,
      -0.335,
      -0.105
     ]
    },
    "format": "[difference, ci95 low, ci95 high] over lines (22 to 25 per comparison); dev half chose every checkpoint"
   },
   "column_step_effect_test_ink_only": {
    "prototype": [
     -0.135,
     -0.265,
     -0.009
    ],
    "cpu_real_letterforms": [
     -0.343,
     -0.654,
     -0.069
    ],
    "letters_v1": [
     -0.195,
     -0.49,
     0.029
    ],
    "letters_v2_real": [
     -0.38,
     -0.675,
     -0.113
    ]
   },
   "synthetic_pages_seed11_prototype": {
    "text_pages_off_and_auto_identical": true,
    "map": 0.278,
    "mask": 0.064,
    "lines_found": "106 of 106",
    "blank_pages_letters_kept": [
     0,
     0,
     0,
     0,
     0,
     0
    ],
    "blank_pages_split_by_column_step": "2 of 6 (texture page 10 to 84 lines read, null letters 3 to 37; clutter page 16 to 28 lines)"
   },
   "find_columns_false_splits_200_synthetic_pages": {
    "split": 6,
    "note": "find_columns alone at the first letter-size estimate, no scale search; 3 at ragged line ends, 3 with the estimate several times too small"
   },
   "verdict": "Model output: with columns read separately, both PHerc. 172 halves put the two real-letterform models ahead of the two synthetic-only ones, and letters-v2-real has the lowest error (test 0.426 against the prototype's 0.625 and letters-v1's 0.739; intervals exclude zero against both; against the CPU real-letterform run neither half resolves the difference). Interpretation: real letterforms help at full scale where more synthetic data did not, and the dev-test disagreement came mostly from lines read across two columns. Diagnostic only; the next claim needs a fresh test set"
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "prereg": "rule fixed in writing on 2026-10-09 before the compared readers ran (preregistration kept in the private workbench; summarized in the source log)",
   "claim": "Model output on public official ink-label drawings; references are unanimous machine transcriptions of the drawings, not a papyrologist's reading; not an ink-map test",
   "check": "fresh preregistered letter-recognition test: PHerc. 1667 official ink labels (w013, w018, w023, w028, w029, w031), 91 groups by a model-free rule, 87 scored (122 letters) where three blind transcriptions agree; each group read with its line centre and letter height given; four checkpoints chosen earlier on the PHerc. 172 dev half; paired bootstrap over groups, 20,000 resamples",
   "data": {
    "groups": 91,
    "scored_groups": 87,
    "scored_letters": 122,
    "groups_json_sha256": "1104ba0b5260a188ea44b4aa7286ef7e482c16213ce1b0e777a9c54b4e7b30d5",
    "refs_sha256": "0347c8f58d9453d0ba3da2ac255c8013ea2763146ffcd48a44f64fc1c98cd520"
   },
   "deviation_1": {
    "what": "first read void: line centres stored in segment rows, so most groups were read outside their crop; fixed before the corrected read, crops and references byte-identical",
    "void_first_read_ink_only": {
     "prototype_step1000": 0.967,
     "real_letterforms_cpu_step3000": 0.91,
     "letters_v1_step12000": 0.926,
     "letters_v2_real_step30000": 0.885
    }
   },
   "reads": {
    "prototype_step1000": {
     "ink_only": 0.402,
     "ink_only_errors": 49,
     "with_lm": 0.492,
     "cer_permuted_refs_ink_only": 1.09,
     "letters_kept": 77,
     "null_letters_kept": [
      7,
      0
     ],
     "controls_pass": true
    },
    "real_letterforms_cpu_step3000": {
     "ink_only": 0.033,
     "ink_only_errors": 4,
     "with_lm": 0.025,
     "cer_permuted_refs_ink_only": 1.164,
     "letters_kept": 114,
     "null_letters_kept": [
      33,
      2
     ],
     "controls_pass": true
    },
    "letters_v1_step12000": {
     "ink_only": 0.32,
     "ink_only_errors": 39,
     "with_lm": 0.311,
     "cer_permuted_refs_ink_only": 1.205,
     "letters_kept": 105,
     "null_letters_kept": [
      17,
      1
     ],
     "controls_pass": true
    },
    "letters_v2_real_step30000": {
     "ink_only": 0.025,
     "ink_only_errors": 3,
     "with_lm": 0.041,
     "cer_permuted_refs_ink_only": 1.172,
     "letters_kept": 119,
     "null_letters_kept": [
      22,
      4
     ],
     "controls_pass": true
    }
   },
   "paired_bootstrap": {
    "H1_v2_minus_prototype": {
     "ink_only_groups": {
      "diff": -0.377,
      "ci95": [
       -0.471,
       -0.292
      ]
     },
     "with_lm_groups": {
      "diff": -0.451,
      "ci95": [
       -0.546,
       -0.362
      ]
     },
     "ink_only_segments": {
      "diff": -0.377,
      "ci95": [
       -0.433,
       -0.301
      ]
     }
    },
    "H2_v2_minus_v1": {
     "ink_only_groups": {
      "diff": -0.295,
      "ci95": [
       -0.38,
       -0.213
      ]
     },
     "with_lm_groups": {
      "diff": -0.27,
      "ci95": [
       -0.365,
       -0.179
      ]
     },
     "ink_only_segments": {
      "diff": -0.295,
      "ci95": [
       -0.354,
       -0.253
      ]
     }
    },
    "cpu_real_minus_prototype": {
     "ink_only_groups": {
      "diff": -0.369,
      "ci95": [
       -0.463,
       -0.28
      ]
     },
     "with_lm_groups": {
      "diff": -0.467,
      "ci95": [
       -0.562,
       -0.376
      ]
     },
     "ink_only_segments": {
      "diff": -0.369,
      "ci95": [
       -0.41,
       -0.298
      ]
     }
    },
    "v2_minus_cpu_real": {
     "ink_only_groups": {
      "diff": -0.008,
      "ci95": [
       -0.046,
       0.025
      ]
     },
     "with_lm_groups": {
      "diff": 0.016,
      "ci95": [
       -0.016,
       0.048
      ]
     },
     "ink_only_segments": {
      "diff": -0.008,
      "ci95": [
       -0.033,
       0.0
      ]
     }
    },
    "v1_minus_prototype": {
     "ink_only_groups": {
      "diff": -0.082,
      "ci95": [
       -0.177,
       0.008
      ]
     },
     "with_lm_groups": {
      "diff": -0.18,
      "ci95": [
       -0.283,
       -0.084
      ]
     },
     "ink_only_segments": {
      "diff": -0.082,
      "ci95": [
       -0.103,
       -0.031
      ]
     }
    }
   },
   "verdict": "H1 confirmed (letters-v2-real minus prototype -0.377, -0.471 to -0.292) and H2 confirmed (minus letters-v1 -0.295, -0.380 to -0.213); all four readers pass both controls. The two real-letterform readers are near the ceiling of this set (3 and 4 errors in 122) and are not separated. Letter recognition only, with the line given"
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "prereg": "rule fixed in writing on 2026-10-09 before the compared readers ran (preregistration kept in the private workbench; summarized in the source log)",
   "claim": "Model output on public PHerc. 1667 2.4 um ink-detection maps (a published scroll); references are the drawing test's unanimous machine transcriptions; line and letter size given",
   "check": "the drawing test's 87 groups cut from the same places of each segment's 2.4 um ink map (same grid as the labels; alignment checked), read with the same checkpoints and settings; primary measure free-ends error (letters before and after the reference not charged), paired bootstrap over groups",
   "maps": "PHerc1667/segments/<segment>/ink-detection/...2.399um...volume-20251217075048-20260417190342-new_canon_autoresearch_recipe-tile256-stride128.tif",
   "alignment": {
    "map_under_labels": [
     0.623,
     0.806
    ],
    "map_elsewhere_in_windows": [
     0.194,
     0.277
    ],
    "best_shift_grid_px_max": 1
   },
   "map_windows_manifest_sha256": "f2c69ba7705685f5f9883de2b8c6ceccce36f1bbc438a1750667a044b7039b3a",
   "reads": {
    "prototype_step1000": {
     "free_ends_ink_only": 0.68,
     "free_ends_with_lm": 0.738,
     "plain_ink_only": 0.795,
     "plain_with_lm": 0.779,
     "free_ends_permuted_refs_ink_only": 0.934,
     "groups_fully_right_free_ends": 15,
     "letters_kept": 59,
     "null_letters_kept": [
      6,
      0
     ],
     "controls_pass": true
    },
    "real_letterforms_cpu_step3000": {
     "free_ends_ink_only": 0.344,
     "free_ends_with_lm": 0.328,
     "plain_ink_only": 1.0,
     "plain_with_lm": 0.73,
     "free_ends_permuted_refs_ink_only": 0.885,
     "groups_fully_right_free_ends": 50,
     "letters_kept": 126,
     "null_letters_kept": [
      27,
      7
     ],
     "controls_pass": true
    },
    "letters_v1_step12000": {
     "free_ends_ink_only": 0.705,
     "free_ends_with_lm": 0.73,
     "plain_ink_only": 0.959,
     "plain_with_lm": 0.844,
     "free_ends_permuted_refs_ink_only": 0.926,
     "groups_fully_right_free_ends": 14,
     "letters_kept": 74,
     "null_letters_kept": [
      8,
      4
     ],
     "controls_pass": true
    },
    "letters_v2_real_step30000": {
     "free_ends_ink_only": 0.369,
     "free_ends_with_lm": 0.369,
     "plain_ink_only": 0.721,
     "plain_with_lm": 0.541,
     "free_ends_permuted_refs_ink_only": 0.91,
     "groups_fully_right_free_ends": 45,
     "letters_kept": 107,
     "null_letters_kept": [
      12,
      8
     ],
     "controls_pass": true
    }
   },
   "paired_bootstrap": {
    "M1_v2_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.311,
      "ci95": [
       -0.415,
       -0.21
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.369,
      "ci95": [
       -0.476,
       -0.26
      ]
     },
     "plain_ink_only": {
      "diff": -0.074,
      "ci95": [
       -0.217,
       0.073
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.311,
      "ci95": [
       -0.45,
       -0.222
      ]
     }
    },
    "M2_v2_minus_v1": {
     "free_ends_ink_only": {
      "diff": -0.336,
      "ci95": [
       -0.43,
       -0.243
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.361,
      "ci95": [
       -0.459,
       -0.263
      ]
     },
     "plain_ink_only": {
      "diff": -0.238,
      "ci95": [
       -0.371,
       -0.096
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.336,
      "ci95": [
       -0.374,
       -0.262
      ]
     }
    },
    "cpu_real_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.336,
      "ci95": [
       -0.427,
       -0.246
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.41,
      "ci95": [
       -0.5,
       -0.32
      ]
     },
     "plain_ink_only": {
      "diff": 0.205,
      "ci95": [
       0.055,
       0.36
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.336,
      "ci95": [
       -0.45,
       -0.274
      ]
     }
    },
    "v2_minus_cpu_real": {
     "free_ends_ink_only": {
      "diff": 0.025,
      "ci95": [
       -0.06,
       0.107
      ]
     },
     "free_ends_with_lm": {
      "diff": 0.041,
      "ci95": [
       -0.042,
       0.123
      ]
     },
     "plain_ink_only": {
      "diff": -0.279,
      "ci95": [
       -0.426,
       -0.139
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": 0.025,
      "ci95": [
       -0.034,
       0.068
      ]
     }
    },
    "v1_minus_prototype": {
     "free_ends_ink_only": {
      "diff": 0.025,
      "ci95": [
       -0.044,
       0.094
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.008,
      "ci95": [
       -0.087,
       0.067
      ]
     },
     "plain_ink_only": {
      "diff": 0.164,
      "ci95": [
       0.05,
       0.279
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": 0.025,
      "ci95": [
       -0.1,
       0.099
      ]
     }
    }
   },
   "verdict": "M1 confirmed (letters-v2-real minus prototype, free-ends -0.311, -0.415 to -0.210) and M2 confirmed (minus letters-v1 -0.336, -0.430 to -0.243); all four readers pass both controls. The real-letterform readers find about two of three reference letters on the map with the line given (free-ends 0.344 CPU, 0.369 letters-v2-real; not separated)"
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "prereg": "rule fixed in writing on 2026-10-09 before the compared readers ran (preregistration kept in the private workbench; summarized in the source log)",
   "claim": "Model output on public PHerc. 1667 2.4 um ink-detection maps (a published scroll); references are the drawing test's unanimous machine transcriptions; the reader finds the lines itself",
   "check": "the map test's 87 groups in windows two letter heights taller above and below (about three lines each), read with lines=None; a group is scored on the found line nearest its true centre within 0.5 letter heights, else every reference letter is an error; primary read letter height given, secondary fully automatic (scale search); primary measure free-ends error, paired bootstrap over groups",
   "wide_windows_manifest_sha256": "d87c368a4045d234b859753f778a5f567c62e6fadaf4e9454e22ff2dbcd73646",
   "missed_lines_primary": [
    "w023-g005",
    "w029-g000",
    "w031-g008",
    "w031-g010"
   ],
   "reads_letter_height_given": {
    "prototype_step1000": {
     "free_ends_ink_only": 0.754,
     "free_ends_with_lm": 0.811,
     "best_line_free_ends_ink_only": 0.738,
     "plain_ink_only": 0.844,
     "plain_with_lm": 0.836,
     "free_ends_permuted_refs_ink_only": 0.934,
     "groups_fully_right_free_ends": 10,
     "line_hits": 83,
     "lines_per_window": 2.98,
     "letters_kept": 130,
     "null_letters_kept": [
      8,
      7
     ],
     "controls_pass": true
    },
    "real_letterforms_cpu_step3000": {
     "free_ends_ink_only": 0.344,
     "free_ends_with_lm": 0.361,
     "best_line_free_ends_ink_only": 0.287,
     "plain_ink_only": 1.016,
     "plain_with_lm": 0.811,
     "free_ends_permuted_refs_ink_only": 0.861,
     "groups_fully_right_free_ends": 49,
     "line_hits": 83,
     "lines_per_window": 2.98,
     "letters_kept": 302,
     "null_letters_kept": [
      54,
      39
     ],
     "controls_pass": true
    },
    "letters_v1_step12000": {
     "free_ends_ink_only": 0.746,
     "free_ends_with_lm": 0.762,
     "best_line_free_ends_ink_only": 0.738,
     "plain_ink_only": 0.959,
     "plain_with_lm": 0.869,
     "free_ends_permuted_refs_ink_only": 0.951,
     "groups_fully_right_free_ends": 10,
     "line_hits": 83,
     "lines_per_window": 2.98,
     "letters_kept": 161,
     "null_letters_kept": [
      23,
      17
     ],
     "controls_pass": true
    },
    "letters_v2_real_step30000": {
     "free_ends_ink_only": 0.402,
     "free_ends_with_lm": 0.459,
     "best_line_free_ends_ink_only": 0.369,
     "plain_ink_only": 0.754,
     "plain_with_lm": 0.648,
     "free_ends_permuted_refs_ink_only": 0.918,
     "groups_fully_right_free_ends": 44,
     "line_hits": 83,
     "lines_per_window": 2.98,
     "letters_kept": 279,
     "null_letters_kept": [
      27,
      17
     ],
     "controls_pass": true
    }
   },
   "reads_fully_automatic": {
    "prototype_step1000": {
     "free_ends_ink_only": 0.746,
     "free_ends_with_lm": 0.795,
     "best_line_free_ends_ink_only": 0.738,
     "plain_ink_only": 0.959,
     "plain_with_lm": 0.869,
     "free_ends_permuted_refs_ink_only": 0.902,
     "groups_fully_right_free_ends": 12,
     "line_hits": 83,
     "lines_per_window": 3.02,
     "letters_kept": 228,
     "null_letters_kept": [
      13,
      2
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "median": 0.81,
      "range": [
       0.43,
       1.96
      ]
     }
    },
    "real_letterforms_cpu_step3000": {
     "free_ends_ink_only": 0.451,
     "free_ends_with_lm": 0.475,
     "best_line_free_ends_ink_only": 0.385,
     "plain_ink_only": 1.205,
     "plain_with_lm": 0.926,
     "free_ends_permuted_refs_ink_only": 0.861,
     "groups_fully_right_free_ends": 38,
     "line_hits": 84,
     "lines_per_window": 3.01,
     "letters_kept": 406,
     "null_letters_kept": [
      41,
      25
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "median": 0.9,
      "range": [
       0.39,
       2.06
      ]
     }
    },
    "letters_v1_step12000": {
     "free_ends_ink_only": 0.697,
     "free_ends_with_lm": 0.73,
     "best_line_free_ends_ink_only": 0.656,
     "plain_ink_only": 1.016,
     "plain_with_lm": 0.934,
     "free_ends_permuted_refs_ink_only": 0.926,
     "groups_fully_right_free_ends": 16,
     "line_hits": 84,
     "lines_per_window": 3.05,
     "letters_kept": 270,
     "null_letters_kept": [
      26,
      9
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "median": 0.89,
      "range": [
       0.42,
       1.8
      ]
     }
    },
    "letters_v2_real_step30000": {
     "free_ends_ink_only": 0.393,
     "free_ends_with_lm": 0.402,
     "best_line_free_ends_ink_only": 0.361,
     "plain_ink_only": 0.787,
     "plain_with_lm": 0.648,
     "free_ends_permuted_refs_ink_only": 0.918,
     "groups_fully_right_free_ends": 44,
     "line_hits": 84,
     "lines_per_window": 3.05,
     "letters_kept": 373,
     "null_letters_kept": [
      30,
      17
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "median": 0.92,
      "range": [
       0.42,
       1.71
      ]
     }
    }
   },
   "paired_bootstrap_letter_height_given": {
    "L1_v2_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.352,
      "ci95": [
       -0.459,
       -0.248
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.352,
      "ci95": [
       -0.458,
       -0.246
      ]
     },
     "best_line_free_ends_ink_only": {
      "diff": -0.369,
      "ci95": [
       -0.476,
       -0.266
      ]
     },
     "plain_ink_only": {
      "diff": -0.09,
      "ci95": [
       -0.243,
       0.068
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.352,
      "ci95": [
       -0.444,
       -0.241
      ]
     }
    },
    "L2_v2_minus_v1": {
     "free_ends_ink_only": {
      "diff": -0.344,
      "ci95": [
       -0.439,
       -0.252
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.303,
      "ci95": [
       -0.395,
       -0.212
      ]
     },
     "best_line_free_ends_ink_only": {
      "diff": -0.369,
      "ci95": [
       -0.467,
       -0.273
      ]
     },
     "plain_ink_only": {
      "diff": -0.205,
      "ci95": [
       -0.328,
       -0.073
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.344,
      "ci95": [
       -0.417,
       -0.191
      ]
     }
    },
    "cpu_real_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.41,
      "ci95": [
       -0.5,
       -0.319
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.451,
      "ci95": [
       -0.54,
       -0.362
      ]
     },
     "best_line_free_ends_ink_only": {
      "diff": -0.451,
      "ci95": [
       -0.548,
       -0.358
      ]
     },
     "plain_ink_only": {
      "diff": 0.172,
      "ci95": [
       0.032,
       0.324
      ]
     }
    },
    "v2_minus_cpu_real": {
     "free_ends_ink_only": {
      "diff": 0.057,
      "ci95": [
       -0.026,
       0.142
      ]
     },
     "free_ends_with_lm": {
      "diff": 0.098,
      "ci95": [
       0.017,
       0.181
      ]
     },
     "best_line_free_ends_ink_only": {
      "diff": 0.082,
      "ci95": [
       -0.008,
       0.171
      ]
     },
     "plain_ink_only": {
      "diff": -0.262,
      "ci95": [
       -0.388,
       -0.142
      ]
     }
    },
    "v1_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.008,
      "ci95": [
       -0.089,
       0.07
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.049,
      "ci95": [
       -0.128,
       0.027
      ]
     },
     "best_line_free_ends_ink_only": {
      "diff": 0.0,
      "ci95": [
       -0.076,
       0.074
      ]
     },
     "plain_ink_only": {
      "diff": 0.115,
      "ci95": [
       0.0,
       0.23
      ]
     }
    }
   },
   "cost_of_finding_the_line_free_ends_ink_only": {
    "prototype_step1000": {
     "diff": 0.074,
     "ci95": [
      0.008,
      0.143
     ]
    },
    "real_letterforms_cpu_step3000": {
     "diff": 0.0,
     "ci95": [
      -0.073,
      0.07
     ]
    },
    "letters_v1_step12000": {
     "diff": 0.041,
     "ci95": [
      -0.007,
      0.09
     ]
    },
    "letters_v2_real_step30000": {
     "diff": 0.033,
     "ci95": [
      -0.031,
      0.096
     ]
    }
   },
   "paired_bootstrap_fully_automatic": {
    "v2_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.352,
      "ci95": [
       -0.448,
       -0.256
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.393,
      "ci95": [
       -0.487,
       -0.297
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.352,
      "ci95": [
       -0.421,
       -0.276
      ]
     }
    },
    "v2_minus_v1": {
     "free_ends_ink_only": {
      "diff": -0.303,
      "ci95": [
       -0.41,
       -0.194
      ]
     },
     "free_ends_with_lm": {
      "diff": -0.328,
      "ci95": [
       -0.429,
       -0.225
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.303,
      "ci95": [
       -0.345,
       -0.261
      ]
     }
    },
    "cpu_real_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.295,
      "ci95": [
       -0.393,
       -0.193
      ]
     }
    },
    "v2_minus_cpu_real": {
     "free_ends_ink_only": {
      "diff": -0.057,
      "ci95": [
       -0.157,
       0.038
      ]
     }
    },
    "v1_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.049,
      "ci95": [
       -0.138,
       0.036
      ]
     }
    }
   },
   "cost_of_scale_search_free_ends_ink_only": {
    "prototype_step1000": {
     "diff": -0.008,
     "ci95": [
      -0.089,
      0.068
     ]
    },
    "real_letterforms_cpu_step3000": {
     "diff": 0.107,
     "ci95": [
      0.032,
      0.187
     ]
    },
    "letters_v1_step12000": {
     "diff": -0.049,
     "ci95": [
      -0.111,
      0.008
     ]
    },
    "letters_v2_real_step30000": {
     "diff": -0.008,
     "ci95": [
      -0.085,
      0.068
     ]
    }
   },
   "verdict": "L1 and L2 confirmed: with the lines found by the reader, letters-v2-real has lower free-ends error than the prototype and than letters-v1 (Model output). Finding the line costs the real-letterform readers little on these windows, and the fully automatic read keeps the ranking; windows are centred on known letter groups, so finding where text is in a whole column is not yet tested (Interpretation)."
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "prereg": "rule fixed in writing on 2026-10-09 before the compared readers ran (preregistration kept in the private workbench; summarized in the source log)",
   "claim": "Model output on public PHerc. 1667 2.4 um ink-detection maps (a published scroll); references are the drawing test's unanimous machine transcriptions; the reader reads each segment's whole labelled region and finds columns, lines and letter positions itself",
   "check": "each segment's labelled frame plus one letter height read once (columns auto, lines found); each of the 87 groups scored on the letters whose boxes centre in its window span on the found line nearest its line (within 0.5 letter heights after undoing the column's deskew), else every reference letter is an error; ink only; primary read letter height given, secondary fully automatic; primary measure free-ends error, paired bootstrap over groups",
   "region_files_manifest_sha256": "6f61867f4ffc5f3efccdfa74a8d408eb14eab4cbf53c2b5387a250285085c0c6",
   "missed_lines_primary": [
    "w028-g000",
    "w028-g001",
    "w029-g003",
    "w031-g002"
   ],
   "reads_letter_height_given": {
    "prototype_step1000": {
     "free_ends_ink_only": 0.754,
     "plain_ink_only": 0.984,
     "best_candidate_line_free_ends_ink_only": 0.639,
     "best_candidate_line_permuted_refs": 0.746,
     "free_ends_permuted_refs_ink_only": 0.926,
     "groups_fully_right_free_ends": 11,
     "line_hits": 83,
     "lines_read": 76,
     "columns_per_segment": {
      "w013": 1,
      "w018": 3,
      "w023": 2,
      "w028": 1,
      "w029": 1,
      "w031": 1
     },
     "letters_kept": 249,
     "null_letters_kept": [
      24,
      7
     ],
     "controls_pass": true
    },
    "real_letterforms_cpu_step3000": {
     "free_ends_ink_only": 0.402,
     "plain_ink_only": 1.443,
     "best_candidate_line_free_ends_ink_only": 0.279,
     "best_candidate_line_permuted_refs": 0.451,
     "free_ends_permuted_refs_ink_only": 0.852,
     "groups_fully_right_free_ends": 43,
     "line_hits": 83,
     "lines_read": 76,
     "columns_per_segment": {
      "w013": 1,
      "w018": 3,
      "w023": 2,
      "w028": 1,
      "w029": 1,
      "w031": 1
     },
     "letters_kept": 550,
     "null_letters_kept": [
      104,
      63
     ],
     "controls_pass": true
    },
    "letters_v1_step12000": {
     "free_ends_ink_only": 0.68,
     "plain_ink_only": 1.484,
     "best_candidate_line_free_ends_ink_only": 0.508,
     "best_candidate_line_permuted_refs": 0.582,
     "free_ends_permuted_refs_ink_only": 0.893,
     "groups_fully_right_free_ends": 16,
     "line_hits": 83,
     "lines_read": 76,
     "columns_per_segment": {
      "w013": 1,
      "w018": 3,
      "w023": 2,
      "w028": 1,
      "w029": 1,
      "w031": 1
     },
     "letters_kept": 465,
     "null_letters_kept": [
      84,
      46
     ],
     "controls_pass": true
    },
    "letters_v2_real_step30000": {
     "free_ends_ink_only": 0.328,
     "plain_ink_only": 1.451,
     "best_candidate_line_free_ends_ink_only": 0.189,
     "best_candidate_line_permuted_refs": 0.426,
     "free_ends_permuted_refs_ink_only": 0.828,
     "groups_fully_right_free_ends": 51,
     "line_hits": 83,
     "lines_read": 76,
     "columns_per_segment": {
      "w013": 1,
      "w018": 3,
      "w023": 2,
      "w028": 1,
      "w029": 1,
      "w031": 1
     },
     "letters_kept": 585,
     "null_letters_kept": [
      111,
      57
     ],
     "controls_pass": true
    }
   },
   "reads_fully_automatic": {
    "prototype_step1000": {
     "free_ends_ink_only": 0.803,
     "plain_ink_only": 1.156,
     "best_candidate_line_free_ends_ink_only": 0.672,
     "best_candidate_line_permuted_refs": 0.713,
     "free_ends_permuted_refs_ink_only": 0.934,
     "groups_fully_right_free_ends": 8,
     "line_hits": 83,
     "lines_read": 69,
     "columns_per_segment": {
      "w013": 1,
      "w018": 2,
      "w023": 2,
      "w028": 1,
      "w029": 1,
      "w031": 1
     },
     "letters_kept": 342,
     "null_letters_kept": [
      41,
      22
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "w013": 1.18,
      "w018": 0.68,
      "w023": 0.58,
      "w028": 0.5,
      "w029": 0.73,
      "w031": 0.62
     }
    },
    "real_letterforms_cpu_step3000": {
     "free_ends_ink_only": 0.443,
     "plain_ink_only": 1.459,
     "best_candidate_line_free_ends_ink_only": 0.254,
     "best_candidate_line_permuted_refs": 0.451,
     "free_ends_permuted_refs_ink_only": 0.852,
     "groups_fully_right_free_ends": 43,
     "line_hits": 82,
     "lines_read": 65,
     "columns_per_segment": {
      "w013": 1,
      "w018": 1,
      "w023": 2,
      "w028": 1,
      "w029": 2,
      "w031": 1
     },
     "letters_kept": 524,
     "null_letters_kept": [
      105,
      63
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "w013": 1.18,
      "w018": 1.36,
      "w023": 0.74,
      "w028": 0.8,
      "w029": 0.92,
      "w031": 1.24
     }
    },
    "letters_v1_step12000": {
     "free_ends_ink_only": 0.59,
     "plain_ink_only": 1.426,
     "best_candidate_line_free_ends_ink_only": 0.459,
     "best_candidate_line_permuted_refs": 0.615,
     "free_ends_permuted_refs_ink_only": 0.885,
     "groups_fully_right_free_ends": 23,
     "line_hits": 83,
     "lines_read": 74,
     "columns_per_segment": {
      "w013": 1,
      "w018": 2,
      "w023": 2,
      "w028": 1,
      "w029": 2,
      "w031": 1
     },
     "letters_kept": 566,
     "null_letters_kept": [
      120,
      66
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "w013": 1.18,
      "w018": 1.09,
      "w023": 0.58,
      "w028": 0.99,
      "w029": 0.92,
      "w031": 0.99
     }
    },
    "letters_v2_real_step30000": {
     "free_ends_ink_only": 0.336,
     "plain_ink_only": 1.279,
     "best_candidate_line_free_ends_ink_only": 0.197,
     "best_candidate_line_permuted_refs": 0.451,
     "free_ends_permuted_refs_ink_only": 0.861,
     "groups_fully_right_free_ends": 52,
     "line_hits": 82,
     "lines_read": 57,
     "columns_per_segment": {
      "w013": 1,
      "w018": 1,
      "w023": 2,
      "w028": 1,
      "w029": 1,
      "w031": 1
     },
     "letters_kept": 557,
     "null_letters_kept": [
      107,
      36
     ],
     "controls_pass": true,
     "chosen_over_drawn_letter_size": {
      "w013": 1.18,
      "w018": 1.36,
      "w023": 0.92,
      "w028": 1.24,
      "w029": 1.15,
      "w031": 0.62
     }
    }
   },
   "paired_bootstrap_letter_height_given": {
    "W1_v2_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.426,
      "ci95": [
       -0.529,
       -0.325
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.426,
      "ci95": [
       -0.47,
       -0.333
      ]
     },
     "plain_ink_only": {
      "diff": 0.467,
      "ci95": [
       0.299,
       0.664
      ]
     }
    },
    "W2_v2_minus_v1": {
     "free_ends_ink_only": {
      "diff": -0.352,
      "ci95": [
       -0.452,
       -0.25
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.352,
      "ci95": [
       -0.393,
       -0.295
      ]
     },
     "plain_ink_only": {
      "diff": -0.033,
      "ci95": [
       -0.206,
       0.15
      ]
     }
    },
    "cpu_real_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.352,
      "ci95": [
       -0.443,
       -0.266
      ]
     },
     "plain_ink_only": {
      "diff": 0.459,
      "ci95": [
       0.31,
       0.634
      ]
     }
    },
    "v2_minus_cpu_real": {
     "free_ends_ink_only": {
      "diff": -0.074,
      "ci95": [
       -0.153,
       0.008
      ]
     },
     "plain_ink_only": {
      "diff": 0.008,
      "ci95": [
       -0.124,
       0.137
      ]
     }
    },
    "v1_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.074,
      "ci95": [
       -0.154,
       0.0
      ]
     },
     "plain_ink_only": {
      "diff": 0.5,
      "ci95": [
       0.333,
       0.684
      ]
     }
    }
   },
   "region_minus_centred_window_free_ends_ink_only": {
    "prototype_step1000": {
     "diff": 0.0,
     "ci95": [
      -0.065,
      0.063
     ]
    },
    "real_letterforms_cpu_step3000": {
     "diff": 0.057,
     "ci95": [
      -0.009,
      0.128
     ]
    },
    "letters_v1_step12000": {
     "diff": -0.066,
     "ci95": [
      -0.136,
      0.007
     ]
    },
    "letters_v2_real_step30000": {
     "diff": -0.074,
     "ci95": [
      -0.146,
      -0.007
     ]
    }
   },
   "paired_bootstrap_fully_automatic": {
    "v2_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.467,
      "ci95": [
       -0.562,
       -0.372
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.467,
      "ci95": [
       -0.574,
       -0.292
      ]
     }
    },
    "v2_minus_v1": {
     "free_ends_ink_only": {
      "diff": -0.254,
      "ci95": [
       -0.354,
       -0.154
      ]
     },
     "free_ends_ink_only_segments": {
      "diff": -0.254,
      "ci95": [
       -0.304,
       -0.19
      ]
     }
    },
    "cpu_real_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.361,
      "ci95": [
       -0.463,
       -0.259
      ]
     }
    },
    "v2_minus_cpu_real": {
     "free_ends_ink_only": {
      "diff": -0.107,
      "ci95": [
       -0.192,
       -0.024
      ]
     }
    },
    "v1_minus_prototype": {
     "free_ends_ink_only": {
      "diff": -0.213,
      "ci95": [
       -0.305,
       -0.119
      ]
     }
    }
   },
   "verdict": "W1 and W2 confirmed: reading whole map regions with columns, lines and letter positions found by the reader, letters-v2-real has lower free-ends error than the prototype and than letters-v1 (Model output). It does no worse than in centred windows and is the best reader with nothing given (0.336, 52 of 87 groups fully right); the regions are the boxes around labelled letters, so where text is was partly given (Interpretation)."
  },
  {
   "date": "2026-10-09",
   "source": "docs/logs/2026-10-14-letter-reader.md",
   "prereg": "rule fixed in writing on 2026-10-09 before the compared readers ran (preregistration kept in the private workbench; summarized in the source log)",
   "claim": "Model output on public PHerc. 1667 2.4 um ink-detection maps of the 13 segments with no ink labels (w011, w012, w030, w032 to w041; a published scroll); the reference is the published transcription of columns 5 to 22 (arXiv 2606.29085, Methods), parsed by scripts/letters/p1667_text.py from the PDF's word boxes",
   "check": "each segment's whole map read once (columns auto, lines found, ink only); every read line aligned to every published line (local alignment: match +1, mismatch -1, gap -1, '?' 0); decisive statistic: ordered pairs, consecutive read lines of one reader column whose unique best published lines score at least 5 and are 1 or 2 lines apart in one column, in order; U1 needs at least 3 pairs, more than every one of 297 decoy texts (words, corpus, shuffled; shifted reported only) and a passing reader control; fit (sum of best scores) reported only",
   "reference_sha256": "f936e457eb047f27014e642f50b750b6ee48c825b9eb372fda324449eec0887c",
   "reference": {
    "columns": 18,
    "lines": 177,
    "visible_letters": 1752,
    "unknown_letters": 586
   },
   "decoys_sha256": "a8bda3815e043f89ab59e1a09bdfd85fa192b605d37c3dc447b204bbf23b145b",
   "cut_maps_manifest_sha256": "632419002cfde2c8bea929707b07246e968b9bf87f29a3d1d33bfe131be9efa8",
   "scorer_check_labelled_regions": {
    "prototype_step1000": {
     "lines": 76,
     "letters_read": 352,
     "letters_kept": 249,
     "null_letters_kept": [
      24,
      7
     ],
     "reader_control_passes": true,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 153,
     "fit_p": {
      "shifted": 0.92,
      "words": 0.41,
      "corpus": 0.06,
      "shuffled": 0.16
     },
     "summed_excess_over_shifted": -2.6,
     "letter_px_given_range": [
      74.0,
      103.5
     ]
    },
    "real_letterforms_cpu_step3000": {
     "lines": 76,
     "letters_read": 1154,
     "letters_kept": 550,
     "null_letters_kept": [
      104,
      63
     ],
     "reader_control_passes": true,
     "pairs": 5,
     "pairs_decoy_max": 1,
     "pairs_shifted_max": 5,
     "fit_T": 300,
     "fit_p": {
      "shifted": 0.01,
      "words": 0.01,
      "corpus": 0.01,
      "shuffled": 0.01
     },
     "summed_excess_over_shifted": 27.2,
     "letter_px_given_range": [
      74.0,
      103.5
     ]
    },
    "letters_v1_step12000": {
     "lines": 76,
     "letters_read": 854,
     "letters_kept": 465,
     "null_letters_kept": [
      84,
      46
     ],
     "reader_control_passes": true,
     "pairs": 0,
     "pairs_decoy_max": 1,
     "pairs_shifted_max": 1,
     "fit_T": 202,
     "fit_p": {
      "shifted": 0.9,
      "words": 0.09,
      "corpus": 0.05,
      "shuffled": 0.29
     },
     "summed_excess_over_shifted": -3.2,
     "letter_px_given_range": [
      74.0,
      103.5
     ]
    },
    "letters_v2_real_step30000": {
     "lines": 76,
     "letters_read": 1269,
     "letters_kept": 585,
     "null_letters_kept": [
      111,
      57
     ],
     "reader_control_passes": true,
     "pairs": 11,
     "pairs_decoy_max": 1,
     "pairs_shifted_max": 10,
     "fit_T": 329,
     "fit_p": {
      "shifted": 0.01,
      "words": 0.01,
      "corpus": 0.01,
      "shuffled": 0.01
     },
     "summed_excess_over_shifted": 31.6,
     "letter_px_given_range": [
      74.0,
      103.5
     ]
    }
   },
   "reads_letter_height_given": {
    "prototype_step1000": {
     "lines": 107,
     "letters_read": 41,
     "letters_kept": 27,
     "null_letters_kept": [
      3,
      1
     ],
     "reader_control_passes": true,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 33,
     "fit_p": {
      "shifted": 0.99,
      "words": 0.95,
      "corpus": 0.71,
      "shuffled": 0.84
     },
     "summed_excess_over_shifted": -0.0,
     "per_segment": {
      "w011": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w012": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w030": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.03
      },
      "w032": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w033": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.01
      },
      "w034": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w035": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w036": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.01
      },
      "w037": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w038": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w039": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w040": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w041": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      }
     }
    },
    "real_letterforms_cpu_step3000": {
     "lines": 107,
     "letters_read": 473,
     "letters_kept": 184,
     "null_letters_kept": [
      90,
      45
     ],
     "reader_control_passes": false,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 169,
     "fit_p": {
      "shifted": 1.0,
      "words": 0.87,
      "corpus": 0.42,
      "shuffled": 1.0
     },
     "summed_excess_over_shifted": -3.6,
     "per_segment": {
      "w011": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.02
      },
      "w012": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.01
      },
      "w030": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.586
      },
      "w032": {
       "pairs": 0,
       "summed_excess_over_shifted": -2.121
      },
      "w033": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.141
      },
      "w034": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.727
      },
      "w035": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.707
      },
      "w036": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.152
      },
      "w037": {
       "pairs": 0,
       "summed_excess_over_shifted": 1.212
      },
      "w038": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.808
      },
      "w039": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w040": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.131
      },
      "w041": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      }
     }
    },
    "letters_v1_step12000": {
     "lines": 107,
     "letters_read": 238,
     "letters_kept": 129,
     "null_letters_kept": [
      53,
      84
     ],
     "reader_control_passes": false,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 99,
     "fit_p": {
      "shifted": 0.98,
      "words": 0.99,
      "corpus": 0.8,
      "shuffled": 1.0
     },
     "summed_excess_over_shifted": -2.4,
     "per_segment": {
      "w011": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.929
      },
      "w012": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.061
      },
      "w030": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.303
      },
      "w032": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.101
      },
      "w033": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.848
      },
      "w034": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.03
      },
      "w035": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.172
      },
      "w036": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.939
      },
      "w037": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w038": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w039": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w040": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w041": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      }
     }
    },
    "letters_v2_real_step30000": {
     "lines": 107,
     "letters_read": 613,
     "letters_kept": 196,
     "null_letters_kept": [
      60,
      59
     ],
     "reader_control_passes": true,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 182,
     "fit_p": {
      "shifted": 0.68,
      "words": 0.86,
      "corpus": 0.82,
      "shuffled": 0.83
     },
     "summed_excess_over_shifted": -0.3,
     "per_segment": {
      "w011": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w012": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.061
      },
      "w030": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.444
      },
      "w032": {
       "pairs": 0,
       "summed_excess_over_shifted": -1.02
      },
      "w033": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.333
      },
      "w034": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.141
      },
      "w035": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.444
      },
      "w036": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.02
      },
      "w037": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.101
      },
      "w038": {
       "pairs": 0,
       "summed_excess_over_shifted": -0.869
      },
      "w039": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.04
      },
      "w040": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.0
      },
      "w041": {
       "pairs": 0,
       "summed_excess_over_shifted": 0.01
      }
     }
    }
   },
   "reads_fully_automatic": {
    "prototype_step1000": {
     "lines": 111,
     "letters_read": 95,
     "letters_kept": 61,
     "null_letters_kept": [
      6,
      7
     ],
     "reader_control_passes": true,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 62,
     "fit_p": {
      "shifted": 0.96,
      "words": 0.3,
      "corpus": 0.28,
      "shuffled": 0.67
     },
     "summed_excess_over_shifted": -1.0,
     "letter_px_chosen_range": [
      32.6,
      615.2
     ]
    },
    "real_letterforms_cpu_step3000": {
     "lines": 110,
     "letters_read": 443,
     "letters_kept": 196,
     "null_letters_kept": [
      70,
      70
     ],
     "reader_control_passes": false,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 163,
     "fit_p": {
      "shifted": 0.78,
      "words": 0.74,
      "corpus": 0.8,
      "shuffled": 1.0
     },
     "summed_excess_over_shifted": -1.1,
     "letter_px_chosen_range": [
      40.7,
      615.2
     ]
    },
    "letters_v1_step12000": {
     "lines": 83,
     "letters_read": 220,
     "letters_kept": 135,
     "null_letters_kept": [
      55,
      71
     ],
     "reader_control_passes": false,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 97,
     "fit_p": {
      "shifted": 0.92,
      "words": 0.95,
      "corpus": 0.67,
      "shuffled": 0.95
     },
     "summed_excess_over_shifted": -1.1,
     "letter_px_chosen_range": [
      41.9,
      615.2
     ]
    },
    "letters_v2_real_step30000": {
     "lines": 111,
     "letters_read": 538,
     "letters_kept": 187,
     "null_letters_kept": [
      57,
      58
     ],
     "reader_control_passes": true,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 171,
     "fit_p": {
      "shifted": 0.93,
      "words": 0.83,
      "corpus": 0.66,
      "shuffled": 0.98
     },
     "summed_excess_over_shifted": -2.4,
     "letter_px_chosen_range": [
      32.1,
      615.2
     ]
    }
   },
   "paired_bootstrap_letter_height_given": {
    "U2_v2_minus_prototype": {
     "pairs": {
      "diff": 0.0,
      "ci95": [
       0.0,
       0.0
      ]
     },
     "summed_excess": {
      "diff": -0.3,
      "ci95": [
       -3.6,
       2.4
      ]
     }
    },
    "U3_v2_minus_v1": {
     "pairs": {
      "diff": 0.0,
      "ci95": [
       0.0,
       0.0
      ]
     },
     "summed_excess": {
      "diff": 2.0,
      "ci95": [
       -2.1,
       6.1
      ]
     },
     "counts": false,
     "why": "letters-v1 fails the reader control"
    }
   },
   "post_hoc_not_decisive": {
    "letters_v2_real_mirrored_maps": {
     "lines": 107,
     "letters_read": 566,
     "letters_kept": 149,
     "null_letters_kept": [
      60,
      61
     ],
     "reader_control_passes": false,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 172,
     "fit_p": {
      "shifted": 0.99,
      "words": 0.84,
      "corpus": 0.63,
      "shuffled": 1.0
     },
     "summed_excess_over_shifted": -3.8
    },
    "letters_v2_real_maps_turned_180": {
     "lines": 110,
     "letters_read": 598,
     "letters_kept": 168,
     "null_letters_kept": [
      62,
      71
     ],
     "reader_control_passes": false,
     "pairs": 0,
     "pairs_decoy_max": 0,
     "pairs_shifted_max": 0,
     "fit_T": 175,
     "fit_p": {
      "shifted": 0.97,
      "words": 0.92,
      "corpus": 0.9,
      "shuffled": 1.0
     },
     "summed_excess_over_shifted": -3.0
    },
    "letters_v2_real_labelled_segments_whole": {
     "lines": 88,
     "letters_read": 1704,
     "letters_kept": 782,
     "null_letters_kept": [
      159,
      105
     ],
     "reader_control_passes": true,
     "pairs": 7,
     "pairs_decoy_max": 1,
     "pairs_shifted_max": 6,
     "fit_T": 342,
     "fit_p": {
      "shifted": 0.01,
      "words": 0.01,
      "corpus": 0.01,
      "shuffled": 0.01
     },
     "summed_excess_over_shifted": 31.9,
     "pairs_by_segment": {
      "w013": 5,
      "w018": 2,
      "w023": 0,
      "w028": 0,
      "w029": 0,
      "w031": 0
     }
    },
    "papyrus_grid_px_millions": {
     "unlabelled_13": 25.1,
     "labelled_6": 50.5,
     "w013": 11.6,
     "w018": 14.2,
     "w023": 14.1,
     "w028_w029_w031": [
      3.7,
      3.7,
      3.2
     ],
     "unlabelled_range": [
      0.5,
      3.6
     ]
    },
    "letters_v2_real_region_test_free_ends_error_by_segment": {
     "w013": [
      3,
      18,
      0.167
     ],
     "w018": [
      13,
      57,
      0.228
     ],
     "w023": [
      10,
      18,
      0.556
     ],
     "w028": [
      3,
      6,
      0.5
     ],
     "w029": [
      5,
      9,
      0.556
     ],
     "w031": [
      6,
      14,
      0.429
     ],
     "format": "errors, reference letters, rate; region test, letter height given per segment"
    },
    "orientations_read": "as stored, mirrored left to right, turned 180 degrees; the vertical flip was not read"
   },
   "verdict": "U1 not shown: letters-v2-real passes the reader control but makes 0 ordered pairs on the published text (no decoy makes any) and no read line scores 5 or more on any published line; U2 not resolved (0 against 0); U3 does not count, since letters-v1 fails the reader control (Model output). Post hoc, the same chain read over the six labelled segments' whole maps makes 7 pairs against a decoy maximum of 1, all on w013 and w018, where letters-v2-real's letter error is lowest (0.17 and 0.23 in the region test), and none on w023 (as large as w018), w028, w029 and w031, where it errs on about one letter in two; so the test could pass only if the reader's error on the unlabelled maps were that low, it had little power, and it cannot tell whether maps of unlabelled text are harder to read (Interpretation)."
  },
  {
   "date": "2026-10-14",
   "source": "docs/letter-reader.md (sections The control map and Why these rules)",
   "claim": "Model output on a published scroll's public map (PHerc. 1667 w013, averaged 16 by 16 to about 38 um) and on synthetic pages with no letters; not a reading, no hidden text",
   "check": "the reader's control status (kit/letterread.py control_status: floor MIN_LETTERS 10, factor STRONG_FACTOR 3 over the largest null and the control map) and the control map read on the image's lines",
   "model": "letters-v2-real step 30000 (SHA-256 44946dad...)",
   "w013_whole_map": {
    "letter_px": 86,
    "lines": 13,
    "letters_kept": 122,
    "null_letters_kept": [
     43,
     33,
     27,
     28
    ],
    "status": "weak",
    "ratio": 2.84,
    "chance_per_line": 2.52,
    "kept_on_lines_0_to_9": 110,
    "control_map_same_map": {
     "letters_kept": 122,
     "status": "none"
    },
    "control_map_tile_shuffled_copy": {
     "letters_kept": 40,
     "ratio": 3.05,
     "status": "weak"
    },
    "unchanged": "letters, boxes, confidences and control counts identical to the read before the status was added"
   },
   "w013_windows_520_by_1000_px": {
    "read": "letter_px 86, four nulls, floor from each window's own nulls",
    "as_stored": {
     "windows": 20,
     "old_verdict_passed": 12,
     "strong": 4,
     "too_few": 8,
     "none": 8,
     "strong_letters_kept": "10 to 16",
     "too_few_letters_kept": "1 to 7"
    },
    "tile_shuffled_copies": {
     "copies": 3,
     "windows": 60,
     "old_verdict_passed": 17,
     "old_well_above": 2,
     "letters_kept_where_old_passed": "1 to 5",
     "max_letters_kept": 5,
     "too_few": 17,
     "none": 43
    }
   },
   "synthetic_blank_pages": {
    "generator": "scripts/letters/eval_pages.py blank_page, letter height 24 px, seeds 21 and 22",
    "pages": 24,
    "reads": 48,
    "reads_per_page": "size searched, and 24 px given; four nulls",
    "none": 47,
    "weak": 1,
    "weak_read": "texture page with the size searched: 134 kept against a largest null of 126"
   },
   "reproduce": "python scripts/letters/eval_controls.py windows w013-b16.tif --weights W.npz (defaults 86 px, 520 by 1000 px, 3 copies); python scripts/letters/eval_controls.py blanks --weights W.npz"
  },
  {
   "date": "2026-10-14",
   "source": "docs/letter-reader.md (sections Reading a photograph and Tilted, turned and mirrored text)",
   "claim": "Model output on synthetic photograph-like pages and on two public-domain photographs of published papyri; not a reading. No character error was measured on a real photograph, so letter accuracy on whole photographs is untested",
   "check": "ink polarity (--polarity bright, dark and auto, auto choosing by a quick read both ways: up to 8 lines, 2 nulls each, the larger margin of letters kept over the largest null, bright on a tie), the polarity hint after a failed read, the deskew fill for dark ink, and the quarter-turn orientation read",
   "model": "letters-v2-real step 30000 (SHA-256 44946dad...)",
   "reproduce": "python scripts/letters/eval_photos.py --weights W.npz (defaults: 12 pages, 12 blank pages of each kind, seed 31, four nulls); 9 min 7 s on 2 CPU threads",
   "synthetic_photo_pages": {
    "pages": 12,
    "letters_per_page_px": "18 to 30",
    "lines_per_page": "4 to 7",
    "turned_deg": "2 to 5 either way",
    "read": "true letter height given, four nulls, 8-bit grey, dark ink on a light ground",
    "bright": {
     "status": {
      "none": 11,
      "too_few": 1
     },
     "cer": 0.895,
     "hints": 12,
     "deskew_within_1_deg": 1,
     "deskew_at_limit": 11
    },
    "dark": {
     "status": {
      "strong": 11,
      "weak": 1
     },
     "cer": 0.055,
     "letters": 732,
     "deskew_within_1_deg": 8,
     "deskew_at_limit": 0
    },
    "auto": {
     "chose_dark": 12,
     "same_reads_as_dark": true
    },
    "dark_rows_sheared_in_filled_with_0": {
     "status": {
      "strong": 12
     },
     "cer": 0.076,
     "deskew_within_1_deg": 2,
     "deskew_at_limit": 10
    },
    "quarter_turn_read_with_auto": {
     "best_rot90": 10,
     "quick_read_chose_bright": 3,
     "quick_read_with_rotate_90_chose_dark": 12
    },
    "white_mount_250_of_255_60_percent": {
     "auto_chose_dark": 12,
     "grey_level_guess_dark": 12
    },
    "dark_mount_20_of_255_60_percent": {
     "auto_chose_dark": 8,
     "dark_quick_read_letters_kept": 0,
     "bright_quick_read_letters_kept": "2 to 17",
     "bright_quick_read_largest_null": "5 to 16",
     "grey_level_guess_dark": 0,
     "note": "neither polarity reads; crop to the papyrus"
    }
   },
   "blank_pages": {
    "photo_like": {
     "pages": 12,
     "auto_none": 12,
     "dark_none": 12
    },
    "ink_map_like": {
     "pages": 12,
     "auto_none": 12,
     "bright_none": 12
    },
    "read": "24 px given, four nulls"
   },
   "photographs": {
    "source": "Wikimedia Commons, marked public domain, Princeton University Library; downloaded 2026-10-14, not committed",
    "P.Oxy. 551": {
     "url": "https://upload.wikimedia.org/wikipedia/commons/b/bf/Papyrus_Oxyrhynchus_551_-_Princeton_University_Library%2C_AM_4405_-_Homer%2C_Iliad_XIV%2C227%E2%80%93253%2C256%E2%80%93263.jpg",
     "sha256": "feac2ca61462d85f36d094e78593493bfd85f7fefc1d22ff2218475dbdc826b5",
     "size_px": [
      1939,
      2148
     ],
     "letter_px": 22,
     "bright": {
      "status": "none",
      "letters_kept": 4,
      "null_letters_kept": [
       45,
       10,
       38,
       24
      ],
      "hint": true,
      "quick_read_dark": [
       155,
       [
        41,
        18
       ]
      ],
      "quick_read_bright": [
       1,
       [
        11,
        8
       ]
      ]
     },
     "dark": {
      "status": "strong",
      "letters_kept": 394,
      "null_letters_kept": [
       105,
       19,
       122,
       33
      ],
      "ratio": 3.23,
      "lines": 20,
      "columns": 1,
      "orientation": {
       "as_is": 394,
       "mirrored": 294,
       "rot180": 307,
       "rot90": 113,
       "rot270": 122
      }
     },
     "auto": {
      "chose": "dark",
      "same_read_as_dark": true
     },
     "no_letter_size": "stops: no regular line spacing in either polarity; a crop of the written rows estimates 23.3 px"
    },
    "P.Oxy. 1183": {
     "url": "https://upload.wikimedia.org/wikipedia/commons/6/6c/Papyrus_Oxyrhynchus_1183_-_Princeton_University_Library%2C_AM_4097_-_Isocrates%2C_Trapeziticus_44%E2%80%9348.jpg",
     "sha256": "d35d4a1a430caf596b53954728fd40f9193bf9237295544b20a1bbcdb47de27d",
     "size_px": [
      2571,
      1750
     ],
     "letter_px": 28,
     "bright": {
      "status": "none",
      "letters_kept": 3,
      "null_letters_kept": [
       21,
       17,
       29,
       15
      ],
      "hint": true,
      "quick_read_dark": [
       81,
       [
        18,
        8
       ]
      ],
      "quick_read_bright": [
       1,
       [
        10,
        8
       ]
      ]
     },
     "dark": {
      "status": "weak",
      "letters_kept": 188,
      "null_letters_kept": [
       76,
       49,
       84,
       38
      ],
      "ratio": 2.24,
      "lines": 26,
      "columns": 1,
      "orientation": {
       "as_is": 188,
       "mirrored": 160,
       "rot180": 181,
       "rot90": 98,
       "rot270": 127
      }
     },
     "auto": {
      "chose": "dark",
      "same_read_as_dark": true
     },
     "no_letter_size": "stops: no regular line spacing in either polarity"
    },
    "auto_extra_time_s": "17 to 28 on reads of about a minute (4-core container, 2 threads, other work running)"
   },
   "w013_regression": "PHerc. 1667 w013 averaged 16 by 16, --letter-px 86: letters, boxes, confidences and control counts identical to the read before this change; orientation as_is 122, mirrored 95, rot180 104 unchanged, with rot90 37 and rot270 30 added"
  },
  {
   "date": "2026-10-14",
   "source": "docs/letter-reader.md (sections Exports: CSV and PAGE XML, and Fine-tune on your own lines)",
   "claim": "Software checks, and Model output on synthetic pages; not a reading. The fine-tune run checks that the chain works; it measures nothing about fine-tuning on real lines",
   "check": "letterread --csv and --page-xml exports: PAGE 2019-07-15 files validated against PRImA's schema with lxml, glyph coordinates in the image file's pixels; kit letterlines build and score, and scripts/letters/finetune_maps.py, run end to end on synthetic pages",
   "model": "letters-v2-real step 30000 (SHA-256 44946dad...)",
   "page_xml_schema_validation": {
    "schema": "http://schema.primaresearch.org/PAGE/gts/pagecontent/2019-07-15/pagecontent.xsd",
    "valid": [
     "synthetic level page",
     "synthetic two-column page",
     "synthetic page read with a quarter turn",
     "synthetic page read mirrored and turned 270 degrees",
     "synthetic page with no lines",
     "command line read with --crop and block-averaging by 2",
     "PHerc. 1667 w013 averaged 16 by 16, --letter-px 86 (13 lines, 262 glyphs, CSV 262 rows)",
     "PHerc. 1667 w013 whole 2.4 um map, --crop 2000 30000 1000 60000, --letter-mm 3.3 --um-per-px 2.399, block-averaged by 42"
    ],
    "w013_crop_glyphs_at_box_plus_crop_corner": "195 of 195",
    "w013_crop_page_size_px": [
     79600,
     41600
    ],
    "w013_regression": "with --csv and --page-xml, letters, boxes, confidences and control counts identical to the read before this change"
   },
   "finetune_chain_synthetic": {
    "machine": "4-core cloud container",
    "pages": "12 synthetic map-like pages, letters 40 px, 6 level lines of 18 to 26 letters; pages 0 to 8 to train (54 lines), 9 to 11 set apart (18 lines)",
    "build_s": 2.4,
    "gen_dataset_12000_lines": "6 min 43 s with 4 processes",
    "finetune_s_per_step_4_threads": 3.1,
    "finetune_steps": 50,
    "score_s_18_lines_2_models": 3.5,
    "cer_set_apart_before": 0.191,
    "cer_set_apart_after": 0.186,
    "diff": -0.005,
    "ci95_bootstrap_over_lines": [
     -0.028,
     0.019
    ]
   },
   "reproduce": "python -m kit letterread MAP.tif --weights W.npz --letter-px 86 --csv OUT.csv --page-xml OUT.xml; python -m kit letterlines build RECORDS.json --split train --out TRAIN.npz; python scripts/letters/finetune_maps.py train TRAIN.npz SHARDS --weights W.npz --out OUT --steps 50; python -m kit letterlines score TEST.npz --weights W.npz --weights OUT/linenet.npz"
  },
  {
   "date": "2026-10-14",
   "source": "docs/letter-reader.md (sections Install, and Check the install: the offline demo)",
   "claim": "Software checks, and Model output on synthetic images; not a reading. A pass shows that an install reads clean letters in the hand it was trained on and that its controls stop noise, nothing about reading a scroll",
   "check": "kit letterdemo (scrolls-letterdemo): four lines of the 24 Greek capitals in seeded orders, 32 px tall, drawn in numpy with the skeleton hand of scripts/letters/letterdata.py on a noisy ground, and the same ground with no letters, read at 32 px with 4 nulls; pass when the text page is strong, its character error is at most 0.25 and the noise image is none or too_few. Package scrolls-workbench (pyproject.toml) built as a wheel and installed into fresh virtual environments outside the clone",
   "model": "letters-v2-real step 30000 (SHA-256 44946dad...); random weights from tests/test_letterread.random_weights as the known-bad model",
   "demo_released_weights": {
    "text_page": {
     "status": "strong",
     "letters_kept": 96,
     "letters_drawn": 96,
     "null_letters_kept": [
      7,
      0,
      13,
      2
     ],
     "ratio": 7.38,
     "character_error": 0.0,
     "lines_found": 4
    },
    "noise_image": {
     "status": "none",
     "letters_kept": 0,
     "null_letters_kept": [
      2,
      0,
      2,
      1
     ]
    },
    "verdict": "DEMO PASS, exit code 0",
    "same_on": [
     "Python 3.10.20, numpy 2.2.6 (wheel installed in a fresh virtual environment, weights downloaded from the GitHub release)",
     "Python 3.13.16, numpy 2.5.3"
    ],
    "seconds_4_core_container": [
     5.7,
     11.8
    ]
   },
   "demo_other_seeds": {
    "seeds": 12,
    "passed": 12,
    "character_error": 0.0,
    "text_letters_kept": [
     95,
     96
    ],
    "text_ratio_range": [
     6.86,
     10.67
    ],
    "noise_letters_kept": 0
   },
   "demo_random_weights": {
    "seeds": 3,
    "passed": 0,
    "text_status": [
     "weak",
     "none",
     "none"
    ],
    "character_error": [
     0.906,
     3.865,
     2.865
    ],
    "verdict": "DEMO FAIL, exit code 1"
   },
   "package": {
    "wheel": "scrolls_workbench-0.1.0-py3-none-any.whl, kit with kit/data/*.json",
    "fresh_venv_outside_clone": "scrolls-letterread on the demo page PNG: 96 letters kept, strong; scrolls-workbench prizes and plan; with no extra, scrolls-letterdemo stops with DEMO ERROR and exit code 2",
    "editable_install": "pip install -e . works"
   },
   "w013_regression": "letters, boxes, confidences and control counts identical to the read before this change",
   "reproduce": "python -m pip install \"scrolls-workbench[reader] @ git+https://github.com/ChaseHendrick/Scrolls\"; scrolls-letterdemo --out DIR; scrolls-letterdemo --weights RANDOM.npz (exit 1)"
  },
  {
   "date": "2026-10-14",
   "source": "https://github.com/ChaseHendrick/Scrolls/releases/tag/letters-v2-ccby (release notes) and docs/letter-reader.md (Model card: letters-v2-ccby, Benchmarks: letters-v2-ccby against letters-v2-real)",
   "claim": "Model output on public data (PHerc. 1667 official ink labels and 2.4 um ink maps, PHerc. 172 community drawings, synthetic pages, Hell-Char validation lines); not a reading, no hidden text",
   "check": "letters-v2-ccby, the letter reader's default weights from 14 October 2026 (CC BY 4.0), against letters-v2-real (CC BY-NC 4.0) on the same inputs: CPU, numpy reader, public code at commit 7dc92e8",
   "model": {
    "letters-v2-ccby": {
     "file": "letters-v2-ccby-step026000.npz",
     "sha256": "b14988852fc721cce71e80a66db0a8ea99a1de914ef2b724ecc6e711baa9ed8c",
     "step": 26000,
     "params": 1483241,
     "licence": "CC BY 4.0",
     "chosen_on": "lowest character error of 15 checkpoints on the first 600 Hell-Char validation lines (0.416 at step 26000)"
    },
    "letters-v2-real": {
     "file": "letters-v2-real-step030000.npz",
     "sha256": "44946dadf0195ddb565f097acec0dd1d777af9b6c66d83995c1f4497aa587477",
     "step": 30000,
     "params": 1483241,
     "licence": "CC BY-NC 4.0",
     "chosen_on": "the PHerc. 172 dev half"
    },
    "weights_in_git": false
   },
   "settings": {
    "pherc1667_references": "rebuilt on 14 October 2026 because the 9 October transcriptions were lost: three blind machine transcriptions of each of the same 91 label groups, only the 84 unanimous groups scored, 117 letters; letters-v2-real's PHerc. 1667 numbers here therefore differ from its first published card. Machine transcriptions of the drawings, not a papyrologist's reading",
    "pherc1667_reads": "scripts/letters/eval_p1667.py read with --source label, map, map-wide (letter height given) and map-region --letter-px auto",
    "pherc172": "Bodillium/Herculaneum-Scroll-Labels at commit 2266661, scripts/letters/eval_p172.py --columns auto, test half (with and without the letter language model) and dev half",
    "synthetic_pages": "scripts/letters/eval_pages.py defaults (24 pages, seed 11, columns off)",
    "hellchar": "2,000 validation lines of gen_dataset.py seed 9002 from a Hell-Char-only bank (Hell-Char at aeef846, 21 held-out papyri), scored in the trainer's order; letters-v2-ccby's error on the first 600 reproduces its training record's 0.416",
    "paired_comparison": "letters-v2-ccby minus letters-v2-real, computed before rounding; paired bootstrap, 20,000 resamples, 95% interval, over groups (PHerc. 1667), lines (PHerc. 172) or text lines (Hell-Char)",
    "verdict_rule": "interval wholly below zero: ccby errs less; wholly above zero: ccby errs more; otherwise not resolved",
    "controls_rule": "a read's number counts only when its error against references permuted across groups or lines is above its real error and its confident letters kept are above three times the larger null total; no read failed. On the PHerc. 172 and page reads the permuted-reference check was added after the test plan was fixed",
    "multiplicity": "the six scroll comparisons fixed in advance (four PHerc. 1667 rows, two PHerc. 172 test-half rows) carry no correction for multiplicity; the Hell-Char and dev-half comparisons were not fixed in advance"
   },
   "rows": [
    {
     "row": 1,
     "test": "PHerc. 1667 official ink labels (drawings), line and letter height given",
     "metric": "character error, ink only (plain)",
     "status": "scored readout",
     "units": "84 groups",
     "letters_v2_ccby": 0.043,
     "letters_v2_real": 0.017,
     "errors_ccby": 5,
     "errors_real": 2,
     "reference_letters": 117,
     "difference_ccby_minus_real": 0.026,
     "ci95": [
      -0.018,
      0.079
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.137,
       "letters_kept": 112,
       "null_letters_kept": [
        19,
        8
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.128,
       "letters_kept": 115,
       "null_letters_kept": [
        20,
        4
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 2,
     "test": "PHerc. 1667 2.4 um ink map windows, the same groups, line and letter height given",
     "metric": "free-ends error, ink only",
     "status": "scored readout",
     "units": "84 groups",
     "letters_v2_ccby": 0.462,
     "letters_v2_real": 0.376,
     "errors_ccby": 54,
     "errors_real": 44,
     "reference_letters": 117,
     "difference_ccby_minus_real": 0.085,
     "ci95": [
      0.017,
      0.155
     ],
     "verdict": "ccby errs more",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 0.872,
       "letters_kept": 86,
       "null_letters_kept": [
        17,
        6
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 0.889,
       "letters_kept": 103,
       "null_letters_kept": [
        12,
        7
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 3,
     "test": "PHerc. 1667 ink maps, windows about three lines tall, the reader finds the line (letter height given)",
     "metric": "free-ends error, ink only",
     "status": "scored readout",
     "units": "84 groups",
     "letters_v2_ccby": 0.479,
     "letters_v2_real": 0.41,
     "errors_ccby": 56,
     "errors_real": 48,
     "reference_letters": 117,
     "difference_ccby_minus_real": 0.068,
     "ci95": [
      0.0,
      0.142
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 0.906,
       "letters_kept": 248,
       "null_letters_kept": [
        31,
        25
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 0.889,
       "letters_kept": 265,
       "null_letters_kept": [
        27,
        15
       ],
       "result": "PASS"
      }
     },
     "note": "On the decision boundary: the interval's lower end is exactly 0.000, so not wholly above zero; not resolved"
    },
    {
     "row": 4,
     "test": "PHerc. 1667 ink maps, each segment's whole labelled region read once, nothing given (--letter-px auto)",
     "metric": "free-ends error, ink only",
     "status": "scored readout",
     "units": "84 groups",
     "letters_v2_ccby": 0.308,
     "letters_v2_real": 0.342,
     "errors_ccby": 36,
     "errors_real": 40,
     "reference_letters": 117,
     "difference_ccby_minus_real": -0.034,
     "ci95": [
      -0.118,
      0.048
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 0.812,
       "letters_kept": 551,
       "null_letters_kept": [
        124,
        51
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 0.838,
       "letters_kept": 557,
       "null_letters_kept": [
        107,
        36
       ],
       "result": "PASS"
      }
     },
     "note": "Passes the letters-kept control on totals; two of letters-v2-ccby's six segments are under three times their larger null on their own"
    },
    {
     "row": 5,
     "test": "PHerc. 172 hand-drawn labels, test half, columns found (--columns auto)",
     "metric": "character error over matched lines, ink only",
     "status": "scored readout, diagnostic: the column rule was set after this half was seen",
     "units": "22 lines",
     "letters_v2_ccby": 0.541,
     "letters_v2_real": 0.426,
     "reference_letters": 122,
     "difference_ccby_minus_real": 0.115,
     "ci95": [
      -0.023,
      0.265
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.139,
       "letters_kept": 118,
       "null_letters_kept": [
        18,
        13
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.045,
       "letters_kept": 120,
       "null_letters_kept": [
        23,
        5
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 6,
     "test": "PHerc. 172 hand-drawn labels, test half, columns found, with the letter language model",
     "metric": "character error over matched lines, with language model",
     "status": "scored readout, diagnostic: the column rule was set after this half was seen",
     "units": "22 lines",
     "letters_v2_ccby": 0.5,
     "letters_v2_real": 0.451,
     "reference_letters": 122,
     "difference_ccby_minus_real": 0.049,
     "ci95": [
      -0.048,
      0.155
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.104,
       "letters_kept": 118,
       "null_letters_kept": [
        18,
        13
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.064,
       "letters_kept": 120,
       "null_letters_kept": [
        23,
        5
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 7,
     "test": "Whole synthetic pages (eval_pages.py defaults: 24 pages, seed 11, columns off)",
     "metric": "character error over matched lines: ink-map look, drawn-label look, pooled",
     "status": "side by side; no paired tool exists for pages",
     "units": "24 text pages, 106 lines; 6 blank pages",
     "letters_v2_ccby": {
      "ink_map_look": 0.098,
      "drawn_label_look": 0.012,
      "pooled": 0.071
     },
     "letters_v2_real": {
      "ink_map_look": 0.087,
      "drawn_label_look": 0.012,
      "pooled": 0.064
     },
     "verdict": "no paired comparison",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 0.983,
       "letters_kept": 1213,
       "null_letters_kept": [
        142,
        62
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 0.978,
       "letters_kept": 1222,
       "null_letters_kept": [
        117,
        84
       ],
       "result": "PASS"
      }
     },
     "blank_pages": {
      "pages": 6,
      "letters_v2_ccby_letters_kept": 1,
      "letters_v2_real_letters_kept": 39,
      "note": "letters-v2-real's 39 all fell on one texture page, where its own scrambled-copy control marked them as noise"
     },
     "note": "Fonts of the container that ran the benchmark, the same pages for both readers"
    },
    {
     "row": 8,
     "test": "Hell-Char validation lines (seed 9002, 21 held-out papyri), lines 601 to 2,000, used to choose neither checkpoint",
     "metric": "character error",
     "status": "supplementary comparison, not fixed in advance",
     "units": "1288 text lines",
     "letters_v2_ccby": 0.404,
     "letters_v2_real": 0.43,
     "errors_ccby": 5435,
     "errors_real": 5785,
     "reference_letters": 13452,
     "difference_ccby_minus_real": -0.026,
     "ci95": [
      -0.034,
      -0.018
     ],
     "verdict": "ccby errs less",
     "controls": "no letters-kept control for shard lines; the error against shuffled references, checked after the plan was fixed, is above the real error for both readers"
    },
    {
     "row": 9,
     "test": "Hell-Char validation lines, the first 600, which letters-v2-ccby was chosen on",
     "metric": "character error",
     "status": "supplementary comparison, biased toward letters-v2-ccby",
     "units": "549 text lines",
     "letters_v2_ccby": 0.416,
     "letters_v2_real": 0.444,
     "errors_ccby": 2408,
     "errors_real": 2570,
     "reference_letters": 5783,
     "difference_ccby_minus_real": -0.028,
     "ci95": [
      -0.039,
      -0.017
     ],
     "verdict": "ccby errs less",
     "controls": "no letters-kept control for shard lines; the error against shuffled references, checked after the plan was fixed, is above the real error for both readers"
    },
    {
     "row": 10,
     "test": "PHerc. 1667 2.4 um ink map windows, line and letter height given (as row 2)",
     "metric": "plain character error, ink only",
     "status": "secondary metric",
     "units": "84 groups",
     "letters_v2_ccby": 0.786,
     "letters_v2_real": 0.709,
     "errors_ccby": 92,
     "errors_real": 83,
     "reference_letters": 117,
     "difference_ccby_minus_real": 0.077,
     "ci95": [
      -0.026,
      0.178
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.359,
       "letters_kept": 86,
       "null_letters_kept": [
        17,
        6
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.359,
       "letters_kept": 103,
       "null_letters_kept": [
        12,
        7
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 11,
     "test": "PHerc. 1667 ink maps, windows about three lines tall (as row 3)",
     "metric": "plain character error, ink only",
     "status": "secondary metric",
     "units": "84 groups",
     "letters_v2_ccby": 0.761,
     "letters_v2_real": 0.761,
     "errors_ccby": 89,
     "errors_real": 89,
     "reference_letters": 117,
     "difference_ccby_minus_real": 0.0,
     "ci95": [
      -0.105,
      0.102
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.359,
       "letters_kept": 248,
       "null_letters_kept": [
        31,
        25
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.359,
       "letters_kept": 265,
       "null_letters_kept": [
        27,
        15
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 12,
     "test": "PHerc. 1667 ink maps, whole labelled regions (as row 4)",
     "metric": "plain character error, ink only",
     "status": "secondary metric",
     "units": "84 groups",
     "letters_v2_ccby": 1.385,
     "letters_v2_real": 1.308,
     "errors_ccby": 162,
     "errors_real": 153,
     "reference_letters": 117,
     "difference_ccby_minus_real": 0.077,
     "ci95": [
      -0.085,
      0.237
     ],
     "verdict": "not resolved",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.983,
       "letters_kept": 551,
       "null_letters_kept": [
        124,
        51
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.906,
       "letters_kept": 557,
       "null_letters_kept": [
        107,
        36
       ],
       "result": "PASS"
      }
     }
    },
    {
     "row": 13,
     "test": "PHerc. 172 hand-drawn labels, dev half, columns found",
     "metric": "character error over matched lines, ink only",
     "status": "supplementary comparison, biased toward letters-v2-real, which was chosen on this half",
     "units": "25 lines",
     "letters_v2_ccby": 0.58,
     "letters_v2_real": 0.471,
     "reference_letters": {
      "ccby": 169,
      "real": 170
     },
     "difference_ccby_minus_real": 0.109,
     "ci95": [
      0.006,
      0.204
     ],
     "verdict": "ccby errs more",
     "controls": {
      "letters_v2_ccby": {
       "error_against_permuted_references": 1.184,
       "letters_kept": 160,
       "null_letters_kept": [
        23,
        21
       ],
       "result": "PASS"
      },
      "letters_v2_real": {
       "error_against_permuted_references": 1.171,
       "letters_kept": 174,
       "null_letters_kept": [
        24,
        16
       ],
       "result": "PASS"
      }
     },
     "note": "One small segment is within a factor of three of its nulls for both readers; the half passes on totals"
    }
   ],
   "verdict": "Model output: on scroll data letters-v2-ccby has not been shown to read better than letters-v2-real. It errs more on one of the four PHerc. 1667 scored rows (map windows with the line given, +0.085, 95% interval +0.017 to +0.155) and on the PHerc. 172 dev half that chose letters-v2-real; the other scroll comparisons are not resolved. It errs less on Hell-Char validation lines from held-out papyri, whose dataset gave it all of its real letters (that may favour it there; untested). It is the default for its licence, not because it reads better (Interpretation)",
   "reproduce": "docs/letter-reader.md, Reproduce the benchmarks (letters-v2-ccby against letters-v2-real): each read with --weights letters-v2-ccby and --weights letters-v2-real; eval_p1667.py compare REAL.json CCBY.json --n 20000 --level group --metric plain or free_ends; compare_p172.py REAL.json CCBY.json --n 20000 --level line --key ink_only or with_lm"
  },
  {
   "date": "2026-10-14",
   "source": "docs/letter-reader.md (Check the install: the offline demo, First reads, Letter size matters most, Big maps, The status, Where the letters are, The control map)",
   "claim": "Software checks, and Model output on synthetic images and on a published scroll's public map (PHerc. 1667 w013); not a reading, no hidden text",
   "check": "the guide's quoted outputs rerun with letters-v2-ccby when it became the default weights (kit letterdemo, scripts/letters/demo_p1667.py and the CLI default name it); letters-v2-real's earlier outputs stay in the entries above",
   "model": "letters-v2-ccby step 26000 (SHA-256 b1498885...), read from a local copy of the release file through KIT_LETTER_WEIGHTS; Python 3.13, numpy 2.5.3, 2 threads",
   "demo": {
    "text_page": {
     "status": "strong",
     "letters_kept": 96,
     "letters_drawn": 96,
     "null_letters_kept": [
      9,
      0,
      11,
      2
     ],
     "ratio": 8.73,
     "character_error": 0.0,
     "lines_found": 4
    },
    "noise_image": {
     "status": "none",
     "letters_kept": 0,
     "null_letters_kept": [
      8,
      0,
      6,
      1
     ]
    },
    "verdict": "DEMO PASS, exit code 0",
    "ci_step": "scrolls-letterread on the demo page PNG at --letter-px 32: 96 letters kept, strong"
   },
   "demo_other_seeds": {
    "seeds": 12,
    "seed_rule": "page seed SEED + 100k and noise seed NOISE_SEED + 100k, k = 1 to 12",
    "passed": 12,
    "character_error": 0.0,
    "text_letters_kept": [
     94,
     96
    ],
    "text_ratio_range": [
     6.71,
     9.6
    ],
    "noise_letters_kept": [
     0,
     1
    ],
    "noise_status": "none"
   },
   "w013_map_averaged_16_by_16": {
    "letter_px": 86,
    "lines": 13,
    "letters_kept": 120,
    "null_letters_kept": [
     31,
     20,
     33,
     47
    ],
    "status": "weak",
    "ratio": 2.55,
    "chance_per_line": 2.52,
    "kept_on_lines_0_to_9": 108,
    "orientation_letters_kept": {
     "as_is": 120,
     "mirrored": 96,
     "rot180": 96,
     "rot90": 29,
     "rot270": 27
    },
    "control_map_same_map": {
     "letters_kept": 120,
     "status": "none"
    },
    "control_map_tile_shuffled_copy": {
     "copy": "letterread.null_images(img, 86, n=1, seed=100)",
     "letters_kept": 42,
     "ratio": 2.86,
     "status": "weak",
     "letters_v2_real_same_copy": {
      "letters_kept": 37,
      "ratio": 3.3,
      "status": "weak"
     }
    }
   },
   "w013_demo_jpeg": {
    "read": "scripts/letters/demo_p1667.py: the bucket's ds8 JPEG averaged 2 by 2, --letter-px 86",
    "letters_kept": 121,
    "null_letters_kept": [
     30,
     20,
     34,
     47
    ],
    "status": "weak",
    "ratio": 2.57,
    "same_jpeg_at_letter_px_172": {
     "letters_kept": 121,
     "null_letters_kept": [
      30,
      20,
      34,
      47
     ],
     "status": "weak"
    }
   },
   "w013_whole_2_4um_map": {
    "letter_px_1376": {
     "block": 43,
     "letters_kept": 128,
     "null_letters_kept": [
      27,
      24,
      35,
      45
     ],
     "status": "weak",
     "peak_rss_mb": 248
    },
    "letter_mm_3_3_um_per_px_2_399": {
     "block": 42,
     "letters_kept": 120,
     "null_letters_kept": [
      30,
      25,
      41,
      35
     ],
     "status": "weak",
     "peak_rss_mb": 243
    },
    "note": "with zarr installed; wall times not isolated (another read ran beside them)"
   },
   "w013_regression_letters_v2_real": "letters, boxes, confidences and control counts with letters-v2-real identical to the read before this change (122 kept against 43, 33, 27, 28)",
   "reproduce": "KIT_LETTER_WEIGHTS=letters-v2-ccby-step026000.npz python -m kit letterdemo; python -m kit letterread w013-b16.tif --weights letters-v2-ccby --letter-px 86 [--control w013-b16.tif]; python scripts/letters/demo_p1667.py --segment w013"
  },
  {
   "date": "2026-10-24",
   "source": "docs/prereg/2026-10-24-letter-reader-icdar-h2h.md (Outcome)",
   "prereg": "docs/prereg/2026-10-24-letter-reader-icdar-h2h.md",
   "evidence": "docs/evidence/2026-10-24-letter-icdar-h2h/score.json",
   "evidence_sha256": "713ee5e4f55cd61696dbc9ed917f55ec519f99178532ee846493ea9e1e53c3f1",
   "fill_commit": "e39253010c6f48e243954ce7fc6dc1fe3ed29a30 (owner's workbench; see the note at the top of the public preregistration)",
   "claim": "Model output: Greek letters named on public papyrus photographs (ICDAR 2023 test answers, Zenodo 13825619), expert boxes given. Not a reading of any scroll.",
   "check": "preregistered head-to-head; 8,479 boxes, 516 lines, 28 texts; controls C1 chance, C2 tile-shuffled decoys, C6 prior alone, C7 language model alone; bootstrap over texts, 20,000 resamples",
   "letters": {
    "O5": 0.2907,
    "R1": 0.3399,
    "O2": 0.4485
   },
   "lines_cer": {
    "L5": 0.2407,
    "L-R1": 0.3202,
    "L-R1lm": 0.2521,
    "L-O2free": 0.5466,
    "C7": 0.8689
   },
   "comparisons": {
    "H1": {
     "d": -0.0492,
     "ci95": [
      -0.0612,
      -0.035
     ],
     "verdict": "ahead"
    },
    "H2": {
     "d": -0.0795,
     "ci95": [
      -0.096,
      -0.0611
     ],
     "verdict": "ahead"
    },
    "H2b": {
     "d": -0.0114,
     "ci95": [
      -0.0246,
      0.0033
     ],
     "verdict": "level"
    }
   },
   "controls_pass": true,
   "cost_usd": 3.9,
   "weights": "docs/tools/letterclass.md (GitHub release letterclass-v1, scripts/letters/letterclass-weights.json)"
  }
 ],
 "orientation_checks": [
  {
   "date": "2026-10-13",
   "source": "docs/logs/2026-10-14-upright-before-ink.md",
   "claim": "Model output on development data (PHerc0841 w00 labels); not a reading, not independent confirmation",
   "reader": "ink_9um seeds 42 and 43",
   "bootstrap": "paired block bootstrap, 2000 draws of 107 px blocks, 95 % intervals",
   "ablation_published_w00": {
    "reader": "ink_9um seed 42",
    "what": "published w00 surface volume turned in plane and turned back, cost in AUC against the unturned map",
    "turn_30_degrees": {
     "cost": 0.155,
     "ci95": [
      0.077,
      0.228
     ]
    },
    "turn_90_degrees": {
     "cost": 0.027,
     "ci95": [
      -0.02,
      0.078
     ]
    },
    "resampling_only": {
     "cost": 0.009,
     "ci95": [
      -0.004,
      0.023
     ]
    }
   },
   "own_trace": {
    "what": "one AutoPipeline trace of PHerc0841 from a seed near w00, w00 labels carried within 4 voxels and 30 degrees",
    "labelled_pixels": 352860,
    "angle_degrees": 49.275,
    "angle_consistency": 0.983,
    "published_map_auc_seed42": 0.798,
    "native_auc_seed42": 0.618,
    "upright_auc_seed42": 0.874,
    "upright_minus_native": {
     "seed42": {
      "diff": 0.256,
      "ci95": [
       0.148,
       0.368
      ]
     },
     "seed43": {
      "diff": 0.181,
      "ci95": [
       0.084,
       0.289
      ]
     }
    },
    "upright_minus_sham_45": {
     "seed42": {
      "diff": 0.213,
      "ci95": [
       0.078,
       0.36
      ]
     },
     "seed43": {
      "diff": 0.208,
      "ci95": [
       0.131,
       0.298
      ]
     }
    },
    "aa_published_turned_own_angle": {
     "seed42": {
      "diff": -0.018,
      "ci95": [
       -0.066,
       0.032
      ]
     },
     "seed43": {
      "diff": 0.0,
      "ci95": [
       -0.051,
       0.049
      ]
     }
    },
    "upright_minus_published_seed42": {
     "diff": -0.076,
     "ci95": [
      -0.164,
      0.015
     ]
    },
    "rule": "preregistered 2026-10-13: seed 42 upright beats native, recovers at least half the gap to the published map, beats the 45 degree sham, and seed 43 upright beats native",
    "verdict": "GO"
   },
   "held_out": "not run (labelled PHerc1447 windings, once); PHerc0841 ag405 labels not opened"
  },
  {
   "date": "2026-10-15",
   "source": "AutoPipeline fix4 release README, limits section (https://github.com/ChaseHendrick/Scrolls/blob/submission/progress-2026-10-fix4/submissions/progress-2026-10-fix4/autopipeline/README.md); docs/logs/2026-10-14-upright-before-ink.md",
   "claim": "Model output on development data (PHerc0841 w00 labels); a replication with a second reader, not a reading",
   "reader": "Reader v2",
   "what": "the same AutoPipeline trace and the same labelled pixels as the ink_9um test above, Reader v2 in place of ink_9um",
   "native_auc": 0.851,
   "upright_auc": 0.911,
   "sham_45_auc": 0.872,
   "upright_minus_sham_45": {
    "diff": 0.039,
    "ci95": [
     -0.005,
     0.091
    ]
   },
   "reversed_control": {
    "native": 0.476,
    "upright": 0.709
   },
   "forward_minus_reversed": {
    "native": 0.375,
    "upright": 0.202
   },
   "result": "upright minus sham is not clearly above zero for Reader v2, so ink.orient stays opt in; the turn also raised the depth-reversed control"
  }
 ],
 "umbilicus_checks": [
  {
   "date": "2026-10-13",
   "source": "scripts/umbilicus/README.md",
   "evidence": "docs/evidence/2026-10-14-umbilicus/",
   "claim": "Model output on the 13 published centre lines, each scroll held out of its own models' training; not an untouched test set, and the reference is one person's line per scroll",
   "what": "in-plane distance between the smoothed centre line and the published line at every slice (one every 64 level-0 voxels) inside the published line's height range, pooled over 13 scrolls",
   "references": "the three open-data bucket umbilicus files (PHerc0125, PHerc0211, PHerc0826) and Alexey Drobkov's ten lines, herculaneum-umbilici at de7642ac9b38266e95192489e22fb13b7e2d13b6",
   "folds": [
    [
     "PHerc0125",
     "PHerc0268",
     "PHerc1203"
    ],
    [
     "PHerc0211",
     "PHerc0800",
     "PHerc0191",
     "PHerc1545"
    ],
    [
     "PHerc0826",
     "PHerc1218",
     "PHerc0257"
    ],
    [
     "PHerc1447",
     "PHerc0358",
     "PHerc0813"
    ]
   ],
   "models": "first stage: U-Net, 3.3 million parameters, three slices in a 256 px frame, two seeds per fold averaged; zoom stage: the same U-Net on a 176 px level-3 window round the first answer, one per fold, two rounds; held-out answers computed on NVIDIA GPUs under bfloat16 autocast",
   "weights": {
    "release": "https://github.com/ChaseHendrick/Scrolls/releases/tag/umbilicus-v1",
    "files": 14,
    "sha256": "scripts/umbilicus/README.md and docs/evidence/2026-10-14-umbilicus/heldout-endrule.json",
    "in_git": false
   },
   "slices": 3414,
   "pooled": {
    "centroid_control": {
     "median_mm": 3.98,
     "p90_mm": 9.97,
     "within_1mm_pct": 7,
     "within_2mm_pct": 25
    },
    "first_stage": {
     "median_mm": 1.2,
     "p90_mm": 4.0,
     "within_1mm_pct": 43,
     "within_2mm_pct": 72
    },
    "both_stages": {
     "median_mm": 0.73,
     "p90_mm": 3.37,
     "within_1mm_pct": 60,
     "within_2mm_pct": 80
    }
   },
   "per_scroll_median_mm": {
    "PHerc0125": {
     "centroid_control": 2.36,
     "first_stage": 1.37,
     "both_stages": 0.46
    },
    "PHerc0211": {
     "centroid_control": 2.84,
     "first_stage": 0.99,
     "both_stages": 0.86
    },
    "PHerc0826": {
     "centroid_control": 3.25,
     "first_stage": 1.5,
     "both_stages": 1.29
    },
    "PHerc0191": {
     "centroid_control": 6.45,
     "first_stage": 2.04,
     "both_stages": 1.35
    },
    "PHerc0257": {
     "centroid_control": 3.41,
     "first_stage": 1.24,
     "both_stages": 0.65
    },
    "PHerc0268": {
     "centroid_control": 14.66,
     "first_stage": 4.04,
     "both_stages": 3.48
    },
    "PHerc0358": {
     "centroid_control": 5.79,
     "first_stage": 0.82,
     "both_stages": 0.25
    },
    "PHerc0800": {
     "centroid_control": 6.93,
     "first_stage": 1.56,
     "both_stages": 1.43
    },
    "PHerc0813": {
     "centroid_control": 7.0,
     "first_stage": 1.31,
     "both_stages": 0.82
    },
    "PHerc1203": {
     "centroid_control": 4.64,
     "first_stage": 0.51,
     "both_stages": 0.39
    },
    "PHerc1218": {
     "centroid_control": 1.8,
     "first_stage": 0.69,
     "both_stages": 0.37
    },
    "PHerc1447": {
     "centroid_control": 2.74,
     "first_stage": 1.07,
     "both_stages": 0.45
    },
    "PHerc1545": {
     "centroid_control": 3.03,
     "first_stage": 1.38,
     "both_stages": 1.34
    }
   },
   "median_of_scroll_medians_mm": {
    "first_stage": 1.31,
    "both_stages": 0.82
   },
   "rule": "written 13 October 2026 before each run: the first stage works if its pooled held-out median is at most 2.0 mm and it beats the centroid on at least 10 of 13 scrolls; the zoom stage is adopted if its pooled median is at least 0.1 mm lower than the first stage's and it is not worse on at least 9 of 13 scrolls",
   "verdict": "both passed: the first stage beats the centroid on 13 of 13 scrolls, and adding the zoom stage is not worse on 13 of 13",
   "end_trims": {
    "date": "2026-10-14",
    "calibration": "docs/evidence/2026-10-14-umbilicus/calibrate.json: the end-point rule fixed on the published centres of the 13 published scrolls; radius 4 mm, the smallest of 2, 3, 4 and 5 mm at which at most 0.5 % of middle slices fall under the bar (5.7, 1.2, 0.18 and 0.11 %); bar three quarters of 32 directions and run 2 mm set by meaning",
    "held_out": "docs/evidence/2026-10-14-umbilicus/heldout-endrule.json: the same held-out answers scored over all slices, over the slices the mount and cap trim keeps (keep_band) and over the slices both trims keep (end_rule, what predict.py writes), with endcheck.py score reading the bucket (mount and cap trim from level-5 planes, a stand-in for the level-3 area)",
    "check": "the all-slices scores equal the run records: 0.730 mm (both stages) and 1.197 mm (first stage) over 3,414 slices",
    "both_stages": {
     "all": {
      "slices": 3414,
      "median_mm": 0.73,
      "p90_mm": 3.371,
      "within_1mm_pct": 60.3,
      "within_2mm_pct": 79.6
     },
     "keep_band": {
      "slices": 3373,
      "median_mm": 0.717,
      "p90_mm": 3.206,
      "within_1mm_pct": 61.0,
      "within_2mm_pct": 80.4
     },
     "end_rule": {
      "slices": 3243,
      "median_mm": 0.686,
      "p90_mm": 2.97,
      "within_1mm_pct": 62.4,
      "within_2mm_pct": 81.8
     }
    },
    "first_stage": {
     "all": {
      "slices": 3414,
      "median_mm": 1.197,
      "p90_mm": 3.997,
      "within_1mm_pct": 43.1,
      "within_2mm_pct": 71.9
     },
     "keep_band": {
      "slices": 3373,
      "median_mm": 1.177,
      "p90_mm": 3.877,
      "within_1mm_pct": 43.7,
      "within_2mm_pct": 72.7
     },
     "end_rule": {
      "slices": 3242,
      "median_mm": 1.146,
      "p90_mm": 3.638,
      "within_1mm_pct": 44.5,
      "within_2mm_pct": 73.8
     }
    },
    "dropped_slices": "the end-point rule dropped scored slices on 10 scrolls (130 slices for both stages, 131 for the first stage; none on PHerc0125, 0211 and 0826); on 9 of the 10, in both arms, their median error was above the scroll's median over all its scored slices; the exception is PHerc0191",
    "primary": "the all-slices score stays the main score, because there the finder does not choose which slices count"
   },
   "limits": "ten of the 13 lines are one annotator's; the settings were chosen while looking at some of the same scrolls; PHerc0268 and PHerc0800 have more than one winding centre at some heights; no head-to-head with other finders on the same slices"
  },
  {
   "date": "2026-10-23",
   "source": "docs/tools/axispolish.md",
   "claim": "Measurement against villa's model-predicted lasagna sheets: the coverage-polished Center Finder line against villa's vc_gen_umbilicus on the ten scans of the 22 October polish test, 38 heights, rule frozen before any field was predicted, read once",
   "verdict": "LEVEL",
   "pooled_median_gain_wraps": -0.022,
   "interval_95": [
    -0.237,
    0.256
   ],
   "scans_ahead": "4 of 10 (6 needed)",
   "reported": {
    "unpolished_vs_vc_gen_umbilicus_wraps": 0.051
   },
   "controls": {
    "scan_center_wraps": -4.74,
    "ring_3mm_loss_wraps": 2.85
   },
   "limits": "coverage of model-predicted sheets, not the true axis"
  },
  {
   "date": "2026-10-28",
   "source": "docs/tools/umbilicus.md",
   "claim": "Model output and Measurement: Center Finder v2 (the anchored coverage polish behind the dataset's finder-v2 lines, before their later smoothing) against villa's vc_gen_umbilicus in two tests, each frozen before its heights were read and read once",
   "coverage": {
    "scans": "PHerc0846A and PHercMAN5, 2.4 um",
    "heights": 96,
    "pooled_median_gain_wraps": 0.352,
    "interval_95": [
     0.274,
     0.476
    ],
    "per_scan_wraps": {
     "PHerc0846A": 0.259,
     "PHercMAN5": 0.476
    },
    "verdict": "BETTER",
    "controls": "scan center, 3 mm ring round villa's point and v2 moved 3 mm all below villa"
   },
   "person_line": {
    "reference": "Erwin Nieuwlaar's PHerc0343 line, github.com/Nieuwlaar/pherc343-first-letters e8370f1 (MIT)",
    "heights": 48,
    "median_mm": {
     "v2": 1.83,
     "vc_gen_umbilicus": 1.86,
     "released_finder_line": 1.39
    },
    "difference_interval_95_mm": [
     -0.56,
     0.76
    ],
    "v2_nearer_blocks": "1 of 11",
    "verdict": "LEVEL"
   },
   "disclosed": "a first coverage test the same morning was VOID (its scorer read two 2.4 um scans' grids at the wrong scale)",
   "limits": "coverage reads villa's own predicted sheets; one person's line on one scroll; the released v2 lines were smoothed afterward in place of the tests' smoothing"
  }
 ],
 "strokenet_checks": [
  {
   "date": "2026-10-14",
   "source": "docs/strokenet.md",
   "evidence": "docs/evidence/2026-10-14-strokenet/",
   "claim": "Model output on six held-out PHerc. Paris 4 segments (one scroll, one ink model, rule-made labels), read once under docs/prereg/2026-10-14-strokenet.md",
   "run": "20261014T141348Z",
   "verdict": "StrokeNet misses its bar on T6_dev, FPR_blank; simple scores meet 9 of 9 counted bar items; bar met by 0 of 3 seeds; adds over simple scores on T4 of the primary measures (T2, T4, T8); not interpretable as text structure: T2, T2_20, T3, T3_raw, T7, T7_raw",
   "released": {
    "T1": 0.9899,
    "T1_AP": 0.9518,
    "T2": 0.9999,
    "T3": 1.0,
    "T4": 0.9936,
    "T4_paired": 0.9741,
    "T6": 0.6393,
    "T7": 0.9666,
    "T8": 0.6433,
    "ECE": 0.0422,
    "FPR_blank": 0.0213,
    "FPR_dense": 0.8842
   },
   "adds_over_simple_scores": {
    "T4": {
     "reference": "lines",
     "difference": 0.1412,
     "segments_positive": "6 of 6",
     "sign_flip_p": 0.015625
    }
   },
   "controls_test": {
    "permuted_T1": 0.6647,
    "artifact_T2": 0.9987,
    "artifact_T4": 0.5589
   }
  },
  {
   "date": "2026-10-14",
   "source": "docs/strokenet.md",
   "evidence": "docs/evidence/2026-10-14-strokenet/gridpilot-result.json",
   "claim": "Model output: one 9.2 mm region of PHerc. Paris 4 w028-037 re-rendered with the organisers' checkpoint (scrollprize/ink_canonical_2um r152_3ddec_v2_l5_epoch13) and villa's recipe at a4979b99, under docs/prereg/2026-10-14-gridpilot.md",
   "run": "20261014T161031Z",
   "verdict": "stride 128: tiling effect RMS 25.759 (noise 0.000, r with published 1.000); stride 64: reduced (RMS 12.307, 0.48 of stride 128), r with stride 128 0.983; stride 32: reduced (RMS 6.315, 0.25 of stride 128), r with stride 128 0.980",
   "reproduction_r": 0.999999,
   "tiling_effect_rms_of_255": {
    "128": 25.759,
    "64": 12.307,
    "32": 6.315
   },
   "tiles_per_s_b300": 107.0
  },
  {
   "date": "2026-10-22",
   "source": "docs/strokenet.md#the-two-density-score",
   "evidence": "Not published here: the run's report; the numbers below are copied from it",
   "claim": "Model output on 11 public PHerc0500P2 segments with ink_9um maps, against text and blank labels made by a fixed rule from the team's 2.215 um new_canon_autoresearch_recipe ink map of each segment (model output, not human labels), read once under a preregistration frozen before any test map existed",
   "score": "two-density: s = z(logit d48) + 0.25 z(logit d128), z within each segment over every scored 48 px label; kit/strokenet.py two_density_analyze",
   "run": "not published",
   "claim_wording": "ahead of ink density inside segments on PHerc0500P2 with ink_9um",
   "verdict": "ahead (seed 42, primary): lower 95 % bound of the mean within-segment AUC difference above 0, both controls passed",
   "segments": 11,
   "labels": {
    "text": 62220,
    "blank": 6468
   },
   "reader_exposure": "ink_9um's label dataset lists no PHerc0500P2 segment (S7, docs/logs/2026-10-16-reader-exposure.md)",
   "seed42_primary": {
    "mean_within_segment_d": 0.04,
    "d_ci95": [
     0.028,
     0.05
    ],
    "within_segment_auc_s": 0.701,
    "within_segment_auc_density": 0.662,
    "segments_s_above_density": "11 of 11",
    "per_segment_gain": {
     "min": 0.01,
     "median": 0.029,
     "max": 0.104
    },
    "controls": {
     "reverse_depth_within_auc_s": {
      "value": 0.574,
      "rule": "at least 0.10 below 0.701",
      "pass": true
     },
     "shifted_labels_median_of_20_within_auc_s": {
      "value": 0.507,
      "rule": "at least 0.10 below 0.701",
      "pass": true
     }
    }
   },
   "seed43_descriptive": {
    "mean_within_segment_d": 0.048,
    "segments_s_above_density": "11 of 11",
    "controls": {
     "reverse_depth": {
      "pass": false
     }
    }
   },
   "limits": [
    "ink_9um reads this scroll weakly: density's mean within-segment AUC 0.66, below 0.5 on one segment; a gain over a weak baseline says nothing about strong readers",
    "one scroll, one reader for the verdict, a single read",
    "three test pairs share papyrus, so the interval is likely too narrow",
    "the 128 px window also reads the map's footprint; untested",
    "the weights were set on maps of other segments"
   ]
  }
 ],
 "stackcheck_checks": [
  {
   "date": "2026-10-14",
   "source": "docs/logs/2026-10-14-pherc0139-segment-defects.md",
   "tool": "kit/stackcheck.py (python -m kit stackcheck)",
   "claim": "Measured from the published PHerc0139 segments and umbilicus (geometry only; no CT, model or ink map); the note's CT look at 12 places is Interpretation and is not in these numbers",
   "frame": "full-resolution voxels of volume 20260102150214 (2.399 um); angles about the umbilicus 20260102150214-umbilicus-20260826112529.json at the same z; distances in um; areas in cm2",
   "runs": {
    "A": {
     "segments": "w033 to w037, 2.399 um copies",
     "settings": "--stride 2 --upsample 1 --plane-z 8600 28800 200 --coincide-um 18.72",
     "seconds": 49.8,
     "flagged": {
      "order": [
       "w034/w035"
      ],
      "order_vertices": [
       "w033",
       "w034",
       "w035",
       "w036",
       "w037"
      ],
      "coincident_pairs": [
       "w034/w035"
      ],
      "missing_sectors": [
       "w035"
      ],
      "missing_runs": [
       "w034/w037"
      ],
      "skipped_windings": [
       "w033",
       "w034",
       "w035",
       "w036",
       "w037"
      ],
      "turn_overlaps": [
       "w033",
       "w034",
       "w036",
       "w037"
      ],
      "stack_ends": [
       "w035",
       "w036"
      ]
     },
     "not_checked": []
    },
    "B": {
     "segments": "w023 to w040, 9.362 um copies mapped with catalog:PHerc0139:20250728140407:20260102150214",
     "settings": "--plane-z 8600 28800 200 --coincide-um 18.72",
     "seconds": 144.8,
     "flagged": {
      "order": [
       "w031/w032",
       "w034/w035"
      ],
      "order_vertices": [
       "w033",
       "w034",
       "w035",
       "w036",
       "w037"
      ],
      "coincident_pairs": [
       "w031/w032",
       "w034/w035"
      ],
      "missing_sectors": [
       "w035"
      ],
      "missing_runs": [
       "w034/w037",
       "w034/w038"
      ],
      "skipped_windings": [
       "w028",
       "w031",
       "w032",
       "w033",
       "w034",
       "w035",
       "w036",
       "w037",
       "w038"
      ],
      "turn_overlaps": [
       "w023",
       "w024",
       "w025",
       "w026",
       "w027",
       "w028",
       "w029",
       "w030",
       "w031",
       "w032",
       "w033",
       "w034",
       "w037",
       "w038",
       "w039",
       "w040"
      ],
      "stack_ends": [
       "w035",
       "w036",
       "w037",
       "w038",
       "w039"
      ]
     },
     "not_checked": []
    },
    "B_defaults": {
     "segments": "as run B",
     "settings": "the defaults: 50 planes over the stack's height, --coincide-um 20",
     "seconds": 161.6,
     "flagged": {
      "order": [
       "w031/w032",
       "w034/w035"
      ],
      "order_vertices": [
       "w034",
       "w035",
       "w036",
       "w037"
      ],
      "coincident_pairs": [
       "w031/w032",
       "w032/w033",
       "w034/w035"
      ],
      "missing_sectors": [
       "w035"
      ],
      "missing_runs": [
       "w034/w037",
       "w034/w038",
       "w034/w039",
       "w034/w040"
      ],
      "skipped_windings": [
       "w028",
       "w031",
       "w032",
       "w033",
       "w034",
       "w035",
       "w036",
       "w037",
       "w038"
      ],
      "turn_overlaps": [
       "w023",
       "w024",
       "w025",
       "w026",
       "w027",
       "w028",
       "w029",
       "w030",
       "w031",
       "w032",
       "w033",
       "w034",
       "w037",
       "w038",
       "w039",
       "w040"
      ],
      "stack_ends": [
       "w035",
       "w036",
       "w037",
       "w038",
       "w039"
      ]
     },
     "not_checked": []
    }
   },
   "A": {
    "w035_turn_deg_per_line_q10_50_90": [
     190.5,
     258.3,
     292.5
    ],
    "gap_extent": {
     "z": [
      9000.0,
      27800.0
     ],
     "planes": 95,
     "start_deg_q10_50_90": [
      327.4,
      342.0,
      5.0
     ],
     "end_deg_q10_50_90": [
      71.0,
      80.0,
      84.6
     ],
     "width_deg_min_median_max": [
      33.0,
      98.0,
      120.0
     ]
    },
    "w034_to_w036_um": {
     "gap": {
      "rays": 7427,
      "median_um": 16.5,
      "q25_75_um": [
       -48.8,
       107.4
      ],
      "inside_share": 0.42
     },
     "present": {
      "rays": 24349,
      "median_um": 242.1,
      "q25_75_um": [
       157.4,
       365.4
      ],
      "inside_share": 0.012
     }
    },
    "w033_to_w037_median_um": {
     "gap": 251.5,
     "present": 561.2
    },
    "gap_by_height_w034_to_w036_median_um": [
     {
      "z": [
       8600.0,
       10600.0
      ],
      "median_um": -46.5
     },
     {
      "z": [
       10800.0,
       12800.0
      ],
      "median_um": -7.0
     },
     {
      "z": [
       13000.0,
       14800.0
      ],
      "median_um": 27.0
     },
     {
      "z": [
       15000.0,
       16800.0
      ],
      "median_um": 65.0
     },
     {
      "z": [
       17000.0,
       18800.0
      ],
      "median_um": 110.4
     },
     {
      "z": [
       19000.0,
       20800.0
      ],
      "median_um": 80.2
     },
     {
      "z": [
       21000.0,
       22800.0
      ],
      "median_um": -20.8
     },
     {
      "z": [
       23000.0,
       24800.0
      ],
      "median_um": -96.4
     },
     {
      "z": [
       25000.0,
       26800.0
      ],
      "median_um": -57.7
     },
     {
      "z": [
       27000.0,
       28800.0
      ],
      "median_um": 174.0
     }
    ],
    "gap_vertices_cm2": {
     "w034": 9.105,
     "w036": 7.459
    },
    "w034_w035_rays": {
     "rays": 24489,
     "inverted_share": 0.1023,
     "inverted_beyond_coincide_share": 0.0675,
     "inverted_depth_um_q25_50_75": [
      13.2,
      32.4,
      66.8
     ],
     "within_coincide_share": 0.092
    },
    "missing_run_w034_w037": {
     "rays": 716,
     "share": 0.0203,
     "w034_to_w037_median_um": 293.5,
     "all_present_median_um": 402.4
    },
    "w034_skip_to_w036_beyond_coincide_cm2": 1.288,
    "pair_spacing_against_nearby_pairs_min_max": [
     0.732,
     1.016
    ]
   },
   "B": {
    "w035_turn_deg_per_line_q10_50_90": [
     200.3,
     257.1,
     291.4
    ],
    "gap_extent": {
     "z": [
      9000.0,
      27800.0
     ],
     "planes": 95,
     "start_deg_q10_50_90": [
      327.0,
      341.0,
      5.0
     ],
     "end_deg_q10_50_90": [
      71.0,
      80.0,
      84.6
     ],
     "width_deg_min_median_max": [
      27.0,
      95.0,
      122.0
     ]
    },
    "w034_to_w036_um": {
     "gap": {
      "rays": 7644,
      "median_um": 18.2,
      "q25_75_um": [
       -47.5,
       111.1
      ],
      "inside_share": 0.415
     },
     "present": {
      "rays": 24256,
      "median_um": 241.7,
      "q25_75_um": [
       157.2,
       365.1
      ],
      "inside_share": 0.012
     }
    },
    "w033_to_w037_median_um": {
     "gap": 255.2,
     "present": 560.6
    },
    "gap_by_height_w034_to_w036_median_um": [
     {
      "z": [
       8600.0,
       10600.0
      ],
      "median_um": -45.5
     },
     {
      "z": [
       10800.0,
       12800.0
      ],
      "median_um": -6.2
     },
     {
      "z": [
       13000.0,
       14800.0
      ],
      "median_um": 27.2
     },
     {
      "z": [
       15000.0,
       16800.0
      ],
      "median_um": 68.5
     },
     {
      "z": [
       17000.0,
       18800.0
      ],
      "median_um": 110.7
     },
     {
      "z": [
       19000.0,
       20800.0
      ],
      "median_um": 82.3
     },
     {
      "z": [
       21000.0,
       22800.0
      ],
      "median_um": -15.9
     },
     {
      "z": [
       23000.0,
       24800.0
      ],
      "median_um": -96.5
     },
     {
      "z": [
       25000.0,
       26800.0
      ],
      "median_um": -48.3
     },
     {
      "z": [
       27000.0,
       28800.0
      ],
      "median_um": 164.0
     }
    ],
    "gap_vertices_cm2": {
     "w034": 9.031,
     "w036": 7.215
    },
    "w034_w035_rays": {
     "rays": 24361,
     "inverted_share": 0.1004,
     "inverted_beyond_coincide_share": 0.0667,
     "inverted_depth_um_q25_50_75": [
      13.1,
      31.4,
      66.1
     ],
     "within_coincide_share": 0.0913
    },
    "missing_run_w034_w037": {
     "rays": 725,
     "share": 0.0206,
     "w034_to_w037_median_um": 301.5,
     "all_present_median_um": 401.4
    },
    "w034_skip_to_w036_beyond_coincide_cm2": 1.277,
    "pair_spacing_against_nearby_pairs_min_max": [
     0.727,
     1.242
    ]
   },
   "control_pairs_four_windings_apart_B": {
    "note": "pairs four windings apart lying wholly inside or wholly outside w035; median spacing on the gap's rays against the rays at w035's heights where it is present",
    "pairs": [
     {
      "pair": [
       "w023",
       "w027"
      ],
      "gap_median_um": 858.2,
      "present_median_um": 861.0,
      "ratio": 0.997
     },
     {
      "pair": [
       "w024",
       "w028"
      ],
      "gap_median_um": 864.3,
      "present_median_um": 771.3,
      "ratio": 1.121
     },
     {
      "pair": [
       "w025",
       "w029"
      ],
      "gap_median_um": 749.8,
      "present_median_um": 699.4,
      "ratio": 1.072
     },
     {
      "pair": [
       "w026",
       "w030"
      ],
      "gap_median_um": 744.5,
      "present_median_um": 712.3,
      "ratio": 1.045
     },
     {
      "pair": [
       "w027",
       "w031"
      ],
      "gap_median_um": 734.1,
      "present_median_um": 715.9,
      "ratio": 1.025
     },
     {
      "pair": [
       "w028",
       "w032"
      ],
      "gap_median_um": 623.7,
      "present_median_um": 683.6,
      "ratio": 0.912
     },
     {
      "pair": [
       "w029",
       "w033"
      ],
      "gap_median_um": 636.9,
      "present_median_um": 642.6,
      "ratio": 0.991
     },
     {
      "pair": [
       "w030",
       "w034"
      ],
      "gap_median_um": 589.7,
      "present_median_um": 616.8,
      "ratio": 0.956
     },
     {
      "pair": [
       "w036",
       "w040"
      ],
      "gap_median_um": 740.9,
      "present_median_um": 726.6,
      "ratio": 1.02
     }
    ],
    "w033_w037_ratio": 0.455
   },
   "coincident_pairs_within_coincide_share_B": {
    "w023/w024": 0.0005,
    "w024/w025": 0.0002,
    "w025/w026": 0.0005,
    "w026/w027": 0.0006,
    "w027/w028": 0.0011,
    "w028/w029": 0.0016,
    "w029/w030": 0.0067,
    "w030/w031": 0.0021,
    "w031/w032": 0.0386,
    "w032/w033": 0.0083,
    "w033/w034": 0.0074,
    "w034/w035": 0.0913,
    "w035/w036": 0.0042,
    "w036/w037": 0.0076,
    "w037/w038": 0.0022,
    "w038/w039": 0.0004,
    "w039/w040": 0.0016
   },
   "turn_overlaps_B": {
    "note": "past_full_turn_deg_median is the median per grid line; within_coincide_share and distance_um_median are of the part of the doubly traced end that has the neighbour along its normal (matched_share)",
    "segments": [
     {
      "segment": "w023",
      "past_full_turn_deg_median": 28.2,
      "piece": "outer",
      "lies_on": "w024",
      "matched_share": 0.995,
      "within_coincide_share": 0.744,
      "distance_um_median": 8.3
     },
     {
      "segment": "w024",
      "past_full_turn_deg_median": 28.4,
      "piece": "outer",
      "lies_on": "w025",
      "matched_share": 0.993,
      "within_coincide_share": 0.533,
      "distance_um_median": 16.7
     },
     {
      "segment": "w025",
      "past_full_turn_deg_median": 36.1,
      "piece": "outer",
      "lies_on": "w026",
      "matched_share": 1.0,
      "within_coincide_share": 0.707,
      "distance_um_median": 8.7
     },
     {
      "segment": "w026",
      "past_full_turn_deg_median": 65.7,
      "piece": "outer",
      "lies_on": "w027",
      "matched_share": 0.993,
      "within_coincide_share": 0.733,
      "distance_um_median": 5.1
     },
     {
      "segment": "w027",
      "past_full_turn_deg_median": 47.9,
      "piece": "outer",
      "lies_on": "w028",
      "matched_share": 0.994,
      "within_coincide_share": 0.634,
      "distance_um_median": 8.5
     },
     {
      "segment": "w028",
      "past_full_turn_deg_median": 44.8,
      "piece": "outer",
      "lies_on": "w029",
      "matched_share": 0.997,
      "within_coincide_share": 0.751,
      "distance_um_median": 2.8
     },
     {
      "segment": "w029",
      "past_full_turn_deg_median": 41.0,
      "piece": "outer",
      "lies_on": "w030",
      "matched_share": 0.996,
      "within_coincide_share": 0.75,
      "distance_um_median": 3.4
     },
     {
      "segment": "w030",
      "past_full_turn_deg_median": 36.2,
      "piece": "outer",
      "lies_on": "w031",
      "matched_share": 0.998,
      "within_coincide_share": 0.69,
      "distance_um_median": 6.3
     },
     {
      "segment": "w031",
      "past_full_turn_deg_median": 33.8,
      "piece": "outer",
      "lies_on": "w032",
      "matched_share": 0.999,
      "within_coincide_share": 0.577,
      "distance_um_median": 13.3
     },
     {
      "segment": "w032",
      "past_full_turn_deg_median": 32.6,
      "piece": "outer",
      "lies_on": "w033",
      "matched_share": 1.0,
      "within_coincide_share": 0.438,
      "distance_um_median": 22.8
     },
     {
      "segment": "w033",
      "past_full_turn_deg_median": 30.0,
      "piece": "outer",
      "lies_on": "w034",
      "matched_share": 0.997,
      "within_coincide_share": 0.387,
      "distance_um_median": 32.4
     },
     {
      "segment": "w034",
      "past_full_turn_deg_median": 27.6,
      "piece": "outer",
      "lies_on": "w035",
      "matched_share": 0.232,
      "within_coincide_share": 0.191,
      "distance_um_median": 75.9
     }
    ]
   },
   "checks": "bucket runs of the note's two commands gave reports and masks identical to runs on local copies of the same files; tests/test_stackcheck.py plants each defect in a synthetic spiral",
   "limits": "one scroll; the umbilicus is a published line; the ray checks use rays that cross each segment of a pair once; neighbours along a normal are searched up to 600 um at about the same height and angle; the innermost and outermost segments get no missing-sector check"
  }
 ],
 "ink_lens_checks": [
  {
   "date": "2026-10-20",
   "source": "scripts/experiments/2026-10-20-ink-lens-gate/results.json",
   "prereg": "docs/prereg/2026-10-20-ink-lens-gate.md",
   "tool": "spaces/ink-lens (tools/check_gate.py)",
   "claim": "Model output: d9v2 with three depth windows on the three PHerc0841 640 px crops (held out from d9v2 and ink_9um, not from Hecate); development windows chosen with the labels, not a blind test",
   "settings": "d9v2_ft-012000.pth (sha256 50d2ad0e...a9f966), windows 0, 2, 4 of the middle 21 layers, villa 8aa507c12 flat inference, normalisation batched (engine.robust_torch), overlap 0.5, hann, CPU fp32; kit.auc inner 64",
   "rule": "ink-lens-gate-v1",
   "crops": {
    "0841-w00": {
     "status": "ink",
     "excess": 15.4345,
     "threshold": 0.527333,
     "ink_kept_share": 0.53927,
     "background_kept_share": 0.03108,
     "auc_forward_3win": 0.9147,
     "auc_reverse_3win": 0.5955,
     "swapped_status": "none",
     "swapped_excess": 0.0
    },
    "0841-ag896": {
     "status": "ink",
     "excess": 21.1552,
     "threshold": 0.490485,
     "ink_kept_share": 0.52252,
     "background_kept_share": 0.07064,
     "auc_forward_3win": 0.8299,
     "auc_reverse_3win": 0.5606,
     "swapped_status": "none",
     "swapped_excess": 0.0
    },
    "0841-ag405": {
     "status": "ink",
     "excess": 17.6333,
     "threshold": 0.552445,
     "ink_kept_share": 0.49663,
     "background_kept_share": 0.07073,
     "auc_forward_3win": 0.8392,
     "auc_reverse_3win": 0.6842,
     "swapped_status": "none",
     "swapped_excess": 0.0
    }
   },
   "no_papyrus": "w00 rows and columns 0 to 640: refused by the engine (all zero), so the gate itself was not exercised off the papyrus",
   "pass": {
    "labeled_crops": true,
    "swapped": true,
    "no_papyrus": true,
    "all": true
   }
  }
 ],
 "text_restoration_checks": [
  {
   "date": "2026-10-22",
   "source": "docs/prereg/2026-10-22-greek-restoration.md",
   "evidence": "docs/evidence/greek-restoration-20261022/score.json",
   "claim": "Model output: hosted text models and an order-8 Kneser-Ney letter model restoring hidden letters in the second half of Philodemus, De Musica (First1KGreek tlg1595.tlg221, Kemke), public edited text, under a rule frozen before any request; no scroll data",
   "claim_wording": "no hosted model beats the letter model; all three do clearly worse",
   "verdict": "FAIL",
   "rule": "beats KN: parsed share >= 0.90, gap letter accuracy >= KN + 0.05, one-sided letter-weighted bootstrap lower bound (20,000 resamples, seed 0, quantile 0.05/3) above 0; memorised: probe letter accuracy >= 0.50",
   "items": {
    "gaps": 190,
    "gaps_per_length": 38,
    "gap_lengths": [
     1,
     2,
     3,
     5,
     8
    ],
    "probes": 40,
    "window_symbols": 150,
    "seed": 20261022,
    "test_letters": 29095
   },
   "requests": 690,
   "replies": 690,
   "priced_usd": 0.066,
   "cap_usd": 1.2,
   "kn": {
    "model": "order-8 interpolated modified Kneser-Ney letter model, general (16 million symbols of Perseus and First1KGreek) mixed 0.728 and 0.272 with Volumina Rhetorica (tlg1595.tlg370); beam 256, 24 symbols of right context",
    "letter_acc": 0.331,
    "probe_acc": 0.0814,
    "acc_by_gap_length": {
     "1": 0.8947,
     "2": 0.6711,
     "3": 0.5,
     "5": 0.3,
     "8": 0.1316
    },
    "exact_by_gap_length": {
     "1": 0.8947,
     "2": 0.6053,
     "3": 0.3158,
     "5": 0.1053,
     "8": 0.0
    }
   },
   "hosted": [
    {
     "model": "deepseek-ai/DeepSeek-V3-0324:deepinfra",
     "letter_acc": 0.1814,
     "diff_from_kn": -0.1496,
     "diff_lower_one_sided": -0.2064,
     "parsed_share": 1.0,
     "probe_acc": 0.0845,
     "acc_by_gap_length": {
      "1": 0.4474,
      "2": 0.2895,
      "3": 0.2719,
      "5": 0.1368,
      "8": 0.1151
     },
     "exact_by_gap_length": {
      "1": 0.4474,
      "2": 0.2105,
      "3": 0.0789,
      "5": 0.0,
      "8": 0.0
     }
    },
    {
     "model": "moonshotai/Kimi-K2-Instruct-0905:novita",
     "letter_acc": 0.1814,
     "diff_from_kn": -0.1496,
     "diff_lower_one_sided": -0.2073,
     "parsed_share": 0.9217,
     "probe_acc": 0.0789,
     "acc_by_gap_length": {
      "1": 0.4474,
      "2": 0.3289,
      "3": 0.1667,
      "5": 0.1895,
      "8": 0.1118
     },
     "exact_by_gap_length": {
      "1": 0.3947,
      "2": 0.1579,
      "3": 0.0263,
      "5": 0.0,
      "8": 0.0
     }
    },
    {
     "model": "Qwen/Qwen3-235B-A22B-Instruct-2507:deepinfra",
     "letter_acc": 0.1177,
     "diff_from_kn": -0.2133,
     "diff_lower_one_sided": -0.2748,
     "parsed_share": 1.0,
     "probe_acc": 0.0775,
     "acc_by_gap_length": {
      "1": 0.3421,
      "2": 0.1184,
      "3": 0.1754,
      "5": 0.1211,
      "8": 0.0658
     },
     "exact_by_gap_length": {
      "1": 0.3421,
      "2": 0.0263,
      "3": 0.0263,
      "5": 0.0,
      "8": 0.0
     }
    }
   ],
   "settings": "Hugging Face router, temperature 0, at most 48 output tokens, provider pinned in the model name, one prompt per task",
   "pins": {
    "manifest.json": "ed544db8e3faaf0642217589d0dba44bac5466940145a5846747a82f6f401991",
    "key.json": "177c43b61f77734da5fe88784a75dc268f3404e43af8dbe6e64f97cb93a35405",
    "baseline.json": "99175529f2e39545c8e321aa3c3aa03904bfa9ed0cdaa0cf0ca3d8c5c63b7e0a",
    "corpus": "5c7c4684b7a4515ebd8840b963757b5a7680cfd2c986375a487b3d66bdb594bf"
   },
   "reproduced": "2026-10-22: score.json rescored byte-identical from the frozen replies; corpus and the three prep files rebuilt byte-identical from the pinned clones with kit/greekfill.py",
   "disclosure": "the test half was scored once before (bits per symbol, 16 Oct 2026) in an earlier unpublished letter-model evaluation whose frozen KN settings are reused; the hosted models may have seen De Musica in training (probe accuracy about 0.08 for every arm)",
   "limits": [
    "one text, one editor's normalisation",
    "edited text, not papyrus damage patterns, with word spaces that papyri lack",
    "three hosted models, one prompt each, one run",
    "letter accuracy does not reward near misses",
    "KN always returns exactly L letters and one of its components was trained on another Philodemus text"
   ]
  }
 ],
 "hopbench_checks": [
  {
   "date": "2026-10-24",
   "tool": "kit hopbench",
   "command": "python -m kit hopbench run --pred m7",
   "claim": "Model output: the one-sheet check (kit.sheetgap) reading an organizers' public surface prediction, on planted sheet hops at real contacts between PHerc0139 hand-traced windings w023 to w040 (scan 20250728140407, 9.362 um); geometry only, never a reading",
   "prediction": "PHerc0139/representations/predictions/surfaces/20250728140407-surface-20260413222639-surface-m7-L0-th0.2.zarr",
   "manifest_digest": "e1fc4cac80a05995b8cb55d14534fac3580cfaf7f88b580b023b33738a332645",
   "rule_digest": "4f5abd102a26",
   "settings": {
    "seed": 20261016,
    "clean_mm": 20.0,
    "windows_per_winding": 1,
    "plant_mm": 6.0,
    "sites_per_pair": 4,
    "contact_vox": 4.0,
    "copy_vox": 8.0,
    "far_vox": 6.0,
    "min_side_vox2": 1600.0,
    "thick_quantile": 0.95,
    "thick_cap": 10
   },
   "max_thick_vox": 4,
   "plants": {
    "planted": 136,
    "counted": 64,
    "band": 10,
    "ramp": 54,
    "gap_joined": 11,
    "gap_share": 0.1719,
    "gap_wilson95": [
     0.0988,
     0.2821
    ],
    "gap_joined_band": 2,
    "gap_joined_ramp": 9,
    "seated_joined": 26,
    "seated_share": 0.4062
   },
   "clean": {
    "gap": {
     "median_square_mm": 3.505,
     "max_square_mm": 5.3,
     "windows_ge_12.5mm": 0
    },
    "seated": {
     "median_square_mm": 5.015,
     "max_square_mm": 7.97,
     "windows_ge_12.5mm": 0
    },
    "truth": {
     "median_square_mm": 5.865,
     "max_square_mm": 13.83,
     "windows_ge_12.5mm": 1
    }
   },
   "clean_windows": 18,
   "missing_chunks": 0,
   "seconds": 2206,
   "machine": "4-core Intel Xeon 2.80 GHz cloud container, CPU only; the two runs partly overlapped",
   "source": "docs/tools/hopbench.md"
  },
  {
   "date": "2026-10-24",
   "tool": "kit hopbench",
   "command": "python -m kit hopbench run --pred recto090",
   "claim": "Model output: the one-sheet check (kit.sheetgap) reading an organizers' public surface prediction, on planted sheet hops at real contacts between PHerc0139 hand-traced windings w023 to w040 (scan 20250728140407, 9.362 um); geometry only, never a reading",
   "prediction": "PHerc0139/representations/predictions/surfaces/20250728140407-surface-20250701154204-surface-recto-090.zarr",
   "manifest_digest": "e1fc4cac80a05995b8cb55d14534fac3580cfaf7f88b580b023b33738a332645",
   "rule_digest": "93ff3b6dafda",
   "settings": {
    "seed": 20261016,
    "clean_mm": 20.0,
    "windows_per_winding": 1,
    "plant_mm": 6.0,
    "sites_per_pair": 4,
    "contact_vox": 4.0,
    "copy_vox": 8.0,
    "far_vox": 6.0,
    "min_side_vox2": 1600.0,
    "thick_quantile": 0.95,
    "thick_cap": 10
   },
   "max_thick_vox": 8,
   "plants": {
    "planted": 136,
    "counted": 64,
    "band": 10,
    "ramp": 54,
    "gap_joined": 13,
    "gap_share": 0.2031,
    "gap_wilson95": [
     0.1227,
     0.3171
    ],
    "gap_joined_band": 2,
    "gap_joined_ramp": 11,
    "seated_joined": 43,
    "seated_share": 0.6719
   },
   "clean": {
    "gap": {
     "median_square_mm": 3.905,
     "max_square_mm": 6.85,
     "windows_ge_12.5mm": 0
    },
    "seated": {
     "median_square_mm": 6.355,
     "max_square_mm": 11.13,
     "windows_ge_12.5mm": 0
    },
    "truth": {
     "median_square_mm": 5.865,
     "max_square_mm": 13.83,
     "windows_ge_12.5mm": 1
    }
   },
   "clean_windows": 18,
   "missing_chunks": 0,
   "seconds": 2389,
   "machine": "4-core Intel Xeon 2.80 GHz cloud container, CPU only; the two runs partly overlapped",
   "source": "docs/tools/hopbench.md"
  }
 ],
 "windlabel_box_checks": [
  {
   "date": "2026-10-24",
   "source": "docs/tools/windlabel-box.md",
   "evidence": "docs/evidence/windlabel-box-20261024/summary.json",
   "claim": "Model output: kit windlabel-box relative windings on 18 fresh PHerc0139 boxes of 512 voxels from the open-data m7 and lasagna grad_mag, scored against PHerc0139's published hand windings w023 to w040 under a readout fixed before any chain was grown",
   "claim_wording": "split chains 0.9758 step accuracy on 13,969 judged steps, above the 0.95 bar; one scroll, held out in position only",
   "verdict": "PASS",
   "rule": "fixed before any chain was grown: PASS when the split chains' pooled step accuracy is at least 0.95, its one-sided 95 % block bootstrap lower bound above 0.92, at least 70 % of the uncut chains' judged steps kept, and the accuracy above the random control's 97.5th percentile; VOID when the skip or shift control is above 0.5, or fewer than 2,000 steps or 24 blocks are judged; one run",
   "chain_rule": "v1 0e53e0ae5016",
   "split_rule": "c6d2eecbebc2",
   "pooled": {
    "v1": {
     "steps": 19613,
     "step_accuracy": 0.9223,
     "blocks": 40,
     "lower_bound_95": 0.908
    },
    "split": {
     "steps": 13969,
     "step_accuracy": 0.9758,
     "blocks": 40,
     "lower_bound_95": 0.9683,
     "kept_share": 0.7122,
     "skip_control": 0.0218,
     "shift_control": 0.0645
    },
    "random_control": {
     "draws": 1000,
     "seed": 0,
     "median": 0.9229,
     "p97_5": 0.9253
    },
    "bootstrap": {
     "blocks": "box x 30 degree sector of the chain's seed",
     "resamples": 10000,
     "seed": 0,
     "one_sided": 0.95
    },
    "void_reasons": [],
    "verdict": "PASS"
   },
   "dev_boxes": {
    "boxes": 6,
    "split_step_accuracy": 0.9415,
    "without_umbilicus": 0.8819,
    "note": "earlier development boxes, under the bar; the shortfall sits near z 7000 where predictions and hand windings disagree"
   }
  },
  {
   "date": "2026-10-24",
   "source": "docs/tools/windlabel-box.md",
   "evidence": "docs/evidence/windlabel-box-1667-20261024/summary.json",
   "claim": "Model output: kit windlabel-box relative windings, split rule v1, on 18 PHerc1667 boxes of 512 voxels (scan 20231117161658) from m7 run with the published model and the published lasagna grad_mag, scored against PHerc1667's published hand windings w028 to w041 under a readout fixed before any box had a prediction",
   "claim_wording": "split chains 0.9693 step accuracy on 4,362 judged steps but 68.3 % kept, under the 70 % floor",
   "verdict": "FAIL",
   "rule": "fixed before any box had a prediction: PASS when the split chains' pooled step accuracy is at least 0.95, its one-sided 95 % block bootstrap lower bound above 0.92, at least 70 % of the uncut chains' judged steps kept, and the accuracy above the random control's 97.5th percentile; VOID when the skip or shift control is above 0.5, or fewer than 2,000 steps or 24 blocks are judged; one run",
   "chain_rule": "v1 0e53e0ae5016",
   "split_rule": "v1 c6d2eecbebc2",
   "pooled": {
    "v1": {
     "steps": 6389,
     "step_accuracy": 0.956,
     "blocks": 28,
     "lower_bound_95": 0.9314
    },
    "split": {
     "steps": 4362,
     "step_accuracy": 0.9693,
     "blocks": 27,
     "lower_bound_95": 0.9508,
     "kept_share": 0.6827,
     "skip_control": 0.0435,
     "shift_control": 0.0416
    },
    "random_control": {
     "draws": 1000,
     "seed": 0,
     "median": 0.9537,
     "p97_5": 0.9564
    },
    "bootstrap": {
     "blocks": "box x 30 degree sector of the chain's seed",
     "resamples": 10000,
     "seed": 0,
     "one_sided": 0.95
    },
    "void_reasons": [],
    "verdict": "FAIL"
   }
  },
  {
   "date": "2026-10-24",
   "source": "docs/tools/windlabel-box.md",
   "evidence": "docs/evidence/windlabel-box-1667-20261024/summary.json",
   "claim": "Model output: kit windlabel-box relative windings, split rule v2, on 20 fresh PHerc1667 boxes of 512 voxels (scan 20231117161658) from m7 run with the published model and the published lasagna grad_mag, scored against PHerc1667's published hand windings w028 to w041 under a readout fixed before any box had a prediction",
   "claim_wording": "split chains 0.9786 step accuracy on 5,938 judged steps, lower bound 0.9671, 75.9 % kept; the 1.0 cut was chosen after the v1 result, on spent boxes",
   "verdict": "PASS",
   "rule": "fixed before any box had a prediction: PASS when the split chains' pooled step accuracy is at least 0.95, its one-sided 95 % block bootstrap lower bound above 0.92, at least 70 % of the uncut chains' judged steps kept, and the accuracy above the random control's 97.5th percentile; VOID when the skip or shift control is above 0.5, or fewer than 2,000 steps or 24 blocks are judged; one run",
   "chain_rule": "v1 0e53e0ae5016",
   "split_rule": "v2 cd273aed7673",
   "pooled": {
    "v1": {
     "steps": 7826,
     "step_accuracy": 0.9486,
     "blocks": 35,
     "lower_bound_95": 0.9288
    },
    "split": {
     "steps": 5938,
     "step_accuracy": 0.9786,
     "blocks": 35,
     "lower_bound_95": 0.9671,
     "kept_share": 0.7588,
     "skip_control": 0.0165,
     "shift_control": 0.0772
    },
    "random_control": {
     "draws": 1000,
     "seed": 0,
     "median": 0.9499,
     "p97_5": 0.9527
    },
    "bootstrap": {
     "blocks": "box x 30 degree sector of the chain's seed",
     "resamples": 10000,
     "seed": 0,
     "one_sided": 0.95
    },
    "void_reasons": [],
    "verdict": "PASS"
   }
  },
  {
   "date": "2026-10-29",
   "source": "docs/tools/windlabel-box.md",
   "claim": "Model output: kit windlabel-box relative windings, split rule v2, on 24 fresh PHerc1667 boxes (scan 20231117161658) from m7 and the published lasagna, scored against PHerc1667's published hand windings w028 to w037 under a readout fixed before any box had a prediction, with two community winding tools scored on the same boxes",
   "claim_wording": "split chains 0.9706 of 7,937 judged steps, lower bound 0.9575, 75.4 % kept; ahead of winding-phase at matched coverage (0.978 against 0.942, lead lower bound +0.029) and of E1 on the steps both judged (0.971 against 0.948, lead lower bound +0.017); held out by position, not by scroll",
   "verdict": "PASS",
   "rule": "fixed before any box had a prediction: PASS when the split chains' pooled step accuracy is at least 0.95, its one-sided 95 % block bootstrap lower bound above 0.92, at least 70 % of the uncut chains' judged steps kept, and the accuracy above the random control's 97.5th percentile; INCONCLUSIVE when the accuracy is at least 0.95 and the lower bound at most 0.92; FAIL otherwise; VOID when the skip or shift control is above 0.5; a community tool is comparable when it judged at least 200 items, shares at least 10 blocks with windlabel-box in its decisive comparison and its own control held (E1 also its label-free units check); one read of the truth",
   "chain_rule": "v1 0e53e0ae5016",
   "split_rule": "v2 cd273aed7673",
   "code": "public Scrolls a77ef4a",
   "baselines": {
    "E1": "pscamillo/constraint-gauge a72c423, winding count across each of the chains' own steps",
    "winding-phase": "millerandmuller/winding-phase 95583f3, certified hops"
   },
   "pooled": {
    "v1": {
     "steps": 10523,
     "step_accuracy": 0.9323,
     "blocks": 44,
     "lower_bound_95": 0.9036
    },
    "split": {
     "steps": 7937,
     "step_accuracy": 0.9706,
     "blocks": 43,
     "lower_bound_95": 0.9575,
     "kept_share": 0.7543,
     "skip_control": 0.0434,
     "shift_control": 0.0357
    },
    "random_control": {
     "draws": 1000,
     "seed": 0,
     "median": 0.9363,
     "p97_5": 0.939
    },
    "bootstrap": {
     "blocks": "box x 30 degree sector of the chain's seed",
     "resamples": 10000,
     "seed": 0,
     "one_sided": 0.95
    },
    "void_reasons": [],
    "verdict": "PASS"
   },
   "comparison": {
    "E1": {
     "step_accuracy": 0.914,
     "shuffled_truth": 0.226,
     "matched_steps": {
      "windlabel_box": 0.971,
      "E1": 0.948,
      "lead_lower_bound": 0.017
     }
    },
    "winding-phase": {
     "step_accuracy": 0.927,
     "repaired_control": 0.216,
     "matched_coverage": {
      "windlabel_box": 0.978,
      "winding-phase": 0.942,
      "lead_lower_bound": 0.029
     }
    }
   }
  },
  {
   "date": "2026-10-29",
   "source": "docs/tools/windlabel-box.md",
   "claim": "Model output: kit windlabel-box relative windings, split rule v2, on 27 boxes of held-out PHerc0172 (z 11,660 to 13,000 of the 7.91 um scan) with lasagna from villa's lasagna-preprocess predict3d, scored against the published hand windings under a readout fixed before any box had a prediction; second readout after a disclosed axis-check fix frozen before it",
   "claim_wording": "split chains 0.9515 of 16,599 judged steps, lower bound 0.9408, 77.7 % kept; ahead of winding-phase at matched coverage (0.954 against 0.906, lead lower bound +0.028); E1 not comparable (its units check failed on our input)",
   "verdict": "PASS",
   "rule": "fixed before any box had a prediction: PASS when the split chains' pooled step accuracy is at least 0.95, its one-sided 95 % block bootstrap lower bound above 0.92, at least 70 % of the uncut chains' judged steps kept, and the accuracy above the random control's 97.5th percentile; INCONCLUSIVE when the accuracy is at least 0.95 and the lower bound at most 0.92; FAIL otherwise; VOID when the skip or shift control is above 0.5; a community tool is comparable when it judged at least 200 items, shares at least 10 blocks with windlabel-box in its decisive comparison and its own control held (E1 also its label-free units check); one read of the truth",
   "chain_rule": "v1 0e53e0ae5016",
   "split_rule": "v2 cd273aed7673",
   "code": "public Scrolls a77ef4a",
   "baselines": {
    "E1": "pscamillo/constraint-gauge a72c423, winding count across each of the chains' own steps",
    "winding-phase": "millerandmuller/winding-phase 95583f3, certified hops"
   },
   "pooled": {
    "v1": {
     "steps": 21372,
     "step_accuracy": 0.9341,
     "blocks": 54,
     "lower_bound_95": 0.9212
    },
    "split": {
     "steps": 16599,
     "step_accuracy": 0.9515,
     "blocks": 54,
     "lower_bound_95": 0.9408,
     "kept_share": 0.7767,
     "skip_control": 0.1876,
     "shift_control": 0.134
    },
    "random_control": {
     "draws": 1000,
     "seed": 0,
     "median": 0.9346,
     "p97_5": 0.9363
    },
    "bootstrap": {
     "blocks": "box x 30 degree sector of the chain's seed",
     "resamples": 10000,
     "seed": 0,
     "one_sided": 0.95
    },
    "void_reasons": [],
    "verdict": "PASS"
   },
   "comparison": {
    "E1": {
     "comparable": false,
     "why": "label-free units check failed on the input we built (median step 0 where 1 is expected)"
    },
    "winding-phase": {
     "step_accuracy": 0.898,
     "items": 46076,
     "repaired_control": 0.183,
     "matched_coverage": {
      "windlabel_box": 0.954,
      "winding-phase": 0.906,
      "lead_lower_bound": 0.028
     }
    }
   },
   "disclosed": "a first readout read this scroll void on an axis check that used one winding over the whole band; the fix was frozen before a single second readout, which also reproduced the PHerc1667 result exactly"
  },
  {
   "date": "2026-10-29",
   "source": "docs/tools/windlabel-box.md",
   "claim": "Model output: kit windlabel-box relative windings, split rule v2, on 20 fresh boxes of a second held-out PHerc0172 band (z 10,710 to 11,510, never scored for windings), scored against the published hand windings under a readout fixed before any box had a prediction, with both community winding tools comparable",
   "claim_wording": "split chains 0.955 of 32,411 judged steps, lower bound 0.9471, 82.8 % kept; ahead of E1 on the steps both judged (0.955 against 0.944, lead lower bound +0.009) and of winding-phase at matched coverage (0.952 against 0.934, lead lower bound +0.005); both leads are small",
   "verdict": "PASS",
   "rule": "fixed before any box had a prediction: PASS when the split chains' pooled step accuracy is at least 0.95, its one-sided 95 % block bootstrap lower bound above 0.92, at least 70 % of the uncut chains' judged steps kept, and the accuracy above the random control's 97.5th percentile; INCONCLUSIVE when the accuracy is at least 0.95 and the lower bound at most 0.92; FAIL otherwise; VOID when the skip or shift control is above 0.5; a community tool is comparable when it judged at least 200 items, shares at least 10 blocks with windlabel-box in its decisive comparison and its own control held (E1 also its label-free units check); one read of the truth",
   "chain_rule": "v1 0e53e0ae5016",
   "split_rule": "v2 cd273aed7673",
   "code": "public Scrolls a77ef4a",
   "baselines": {
    "E1": "pscamillo/constraint-gauge a72c423, winding count across each of the chains' own steps",
    "winding-phase": "millerandmuller/winding-phase 95583f3, certified hops"
   },
   "pooled": {
    "v1": {
     "steps": 39155,
     "step_accuracy": 0.9455,
     "blocks": 50,
     "lower_bound_95": 0.9376
    },
    "split": {
     "steps": 32411,
     "step_accuracy": 0.955,
     "blocks": 50,
     "lower_bound_95": 0.9471,
     "kept_share": 0.8278,
     "skip_control": 0.0963,
     "shift_control": 0.0802
    },
    "random_control": {
     "draws": 1000,
     "seed": 0,
     "median": 0.9455,
     "p97_5": 0.9466
    },
    "bootstrap": {
     "blocks": "box x 30 degree sector of the chain's seed",
     "resamples": 10000,
     "seed": 0,
     "one_sided": 0.95
    },
    "void_reasons": [],
    "verdict": "PASS"
   },
   "comparison": {
    "E1": {
     "comparable": true,
     "matched_steps": {
      "windlabel_box": 0.955,
      "E1": 0.944,
      "lead_lower_bound": 0.009
     }
    },
    "winding-phase": {
     "step_accuracy": 0.932,
     "items": 75639,
     "matched_coverage": {
      "windlabel_box": 0.952,
      "winding-phase": 0.934,
      "lead_lower_bound": 0.005
     }
    }
   },
   "disclosed": "villa's lasagna-preprocess predict3d (db3dbea) failed twice the same way on one box's crop; that crop's upper end was moved out 4 voxels and nothing else changed"
  },
  {
   "date": "2026-11-04",
   "source": "docs/tools/windlabel-box.md",
   "claim": "Model output: villa's spiral fit (commit e2254ebe, 30,000 steps, seeds 1 to 3, the same settings for every arm) of PHerc0172 (scan 20241024131838, 7.91 um), fit window z 7,424 to 8,448, with and without kit windlabel-box relative windings (split rule none, 12 boxes in the window), scored once by the exact share of truth crossing pairs on the published hand windings in z 7,680 to 8,064, at least 150 voxels from every site an earlier test had scored; rule frozen before any fit",
   "claim_wording": "the windings lowered the exact share at every seed: 0.459, 0.470 and 0.478 without them, 0.410, 0.392 and 0.396 with them; gain -0.069 (95 % interval -0.091 to -0.046); FAIL. The windings beat a plus-one control by +0.076 (+0.043 to +0.109), so the fit reads them",
   "verdict": "FAIL",
   "rule": "PASS needs the lower end of the 95 % interval of the gain on all pairs above 0, the lower end on pairs one winding apart at least -0.01, and the plus-one control beaten",
   "pairs": 16970,
   "clusters": 303,
   "E": {
    "without_1": 0.459,
    "without_2": 0.47,
    "without_3": 0.478,
    "with_1": 0.41,
    "with_2": 0.392,
    "with_3": 0.396
   },
   "gain": {
    "all_pairs": [
     -0.069,
     -0.091,
     -0.046
    ],
    "one_winding_apart": [
     -0.06,
     -0.077,
     -0.042
    ]
   },
   "windings_minus_plus_one_control": [
    0.076,
    0.043,
    0.109
   ],
   "limits": "one band of one scroll, three seeds, the split cut off (split rule none, not v2), a lasagna field from a run of villa's lasagna-preprocess predict3d made for the test (PHerc0172 has none published)",
   "umbilicus": "both PHerc0172 step accuracy tests (29 Oct) used the umbilicus-lines-v1 PHerc0172 line, off center between about z 9,950 and 12,200 (docs/evidence/2026-10-21-umbilicus-lines/README.md); their axis checks passed"
  }
 ],
 "contact_barrier_checks": [
  {
   "date": "2026-10-22",
   "tool": "kit switchfix barrier (AutoPipeline qa.repair method barrier)",
   "source": "docs/tools/contact-barrier.md",
   "claim": "Model output: AutoPipeline's trace with the contact barrier against villa's vc_grow_seg_from_seed on PHerc0139 (scan 20250728140407, 9.362 um), seeded on published hand-traced report windings that AutoPipeline's trace settings and the barrier were never developed or tested on (the barrier's parent rule, kit.sheetgap 24da24815958, read scores of report windings in another band, z 4352 to 4864, and the records cannot rule out that some of these were among them, so they are held out in this band but not completely unseen), one sector z 5248 to 5760, y 2688 to 3712, x 4352 to 5376, two repeats, rule frozen before any sector, seed, trace or score existed; geometry only, never a reading",
   "verdict": "AHEAD",
   "rule": "gated correct area C (windings passing both gates with at least 0.10 cm2 covered) higher than villa's by more than M = 0.10 cm2 plus both arms' change between repeats, in both repeats, at precision no worse than villa's less 0.02",
   "windings": [
    "w049",
    "w051",
    "w055",
    "w057",
    "w058"
   ],
   "barrier_digest": "3165f76adf4c1a964acfe67f6d0df892b199e15d4ab7dfa21e2c6aa566ac657c",
   "villa": "e0bbb8b",
   "D": {
    "rep1": 0.724,
    "rep2": 0.804
   },
   "M": 0.18,
   "barrier": {
    "rep1": {
     "passing": 2,
     "C": 0.724,
     "covered_cm2": 1.737,
     "wrong_cm2": 0.061,
     "P": 0.9661
    },
    "rep2": {
     "passing": 2,
     "C": 0.804,
     "covered_cm2": 1.593,
     "wrong_cm2": 0.063,
     "P": 0.962
    }
   },
   "villa_arm": {
    "rep1": {
     "passing": 0,
     "C": 0,
     "covered_cm2": 1.789,
     "wrong_cm2": 3.983,
     "P": 0.3099
    },
    "rep2": {
     "passing": 0,
     "C": 0,
     "covered_cm2": 1.677,
     "wrong_cm2": 14.382,
     "P": 0.1044
    }
   },
   "uncut_beside": {
    "rep1": {
     "passing": 0,
     "covered_cm2": 1.818,
     "wrong_cm2": 12.328,
     "P": 0.1285
    },
    "rep2": {
     "passing": 0,
     "covered_cm2": 1.956,
     "wrong_cm2": 12.6,
     "P": 0.1344
    }
   },
   "per_winding": "docs/media/contact-barrier-a15-data.json",
   "units": "covered_cm2, C and P: higher is better; wrong_cm2 (area on other windings): lower is better",
   "controls": "sector controls passed for all five windings (round trip 0.991 to 0.998; a copy of each winding moved half a winding, 9.5 voxels, either way kept 0, and a 1 voxel move kept 0.999 or more, so the scorer could fail); AutoPipeline surfaces held the seed anchor and villa's did not; two repeats; a barrier refusal would count against AutoPipeline (none happened)",
   "limits": "one sector of one scroll, five windings, two repeats, one community tracer in AutoPipeline's harness; three of five windings still failed the switch gate with the barrier"
  },
  {
   "date": "2026-10-22",
   "tool": "AutoPipeline triage repair (before the barrier)",
   "source": "docs/tools/contact-barrier.md",
   "claim": "Model output: on another PHerc0139 sector (z 4352 to 4864, tune windings w032, w037, w038, w040), AutoPipeline with its earlier triage repair against the same villa tracer",
   "verdict": "BEHIND",
   "C": {
    "autopipeline": {
     "rep1": 1.718,
     "rep2": 1.707
    },
    "villa": {
     "rep1": 1.846,
     "rep2": 1.845
    }
   },
   "margin": 0.112,
   "note": "both arms at precision 0.999 to 1.0; a different repair and development windings, reported so the barrier's result is not read as a general win"
  },
  {
   "date": "2026-10-28",
   "tool": "kit switchfix barrier (AutoPipeline qa.repair method barrier)",
   "source": "docs/tools/contact-barrier.md",
   "claim": "Second-scroll test on PHerc0500P2 (13 catalog wrap meshes, automatic surface meshes and not hand traces, on the 9.362 um scan with its published m7 prediction), preregistered before any trace",
   "verdict": "VOID",
   "why": "gate G1: four neighboring wrap pairs (w003-w004, w009-w010, w010-w011, w012-w013) were 2 voxels apart on 0.95, 0.957, 0.72 and 0.854 of samples against 0.98, leaving 4 seeded wraps of the 6 the rule needs; no trace, control or readout ran; the VOID is about the reference meshes, not any tracer; the region is spent for this test"
  }
 ],
 "label_snap_checks": [
  {
   "date": "2026-10-25",
   "tool": "kit labelsnap snap (face snap)",
   "source": "docs/tools/labelsnap.md",
   "claim": "Model output: published PHerc0139 surface label vertices snapped onto the recto face of the 9.362 um label scan, scored against the recto face of the independent 1.129 um scan registered per site, on fresh sites, rule fixed before the run; geometry only, never a reading",
   "sites_scored": 59,
   "vertices": 444,
   "endpoint": "mean gain toward the finer scan's recto face along the vertex normal, label-scan voxels, per-site means, site bootstrap 95 % interval",
   "gain_merged_band_rule": [
    1.71,
    1.38,
    2.07
   ],
   "gain_default": [
    2.37,
    2.01,
    2.76
   ],
   "controls": "the sham (the snap's move reversed) moved labels away; a random sign gave no gain",
   "disclosure": "11 of the 59 sites lie on three segments the Snapped Labels release later held back for evaluation; the test ran before that hold and its per-site rows were not kept"
  },
  {
   "date": "2026-10-27",
   "tool": "kit labelsnap snap (face snap)",
   "source": "docs/tools/labelsnap.md",
   "claim": "Model output: published PHerc0814 surface label vertices (12 automatic segments on the 9.362 um scan) snapped onto the recto face, scored against the recto face of the 1.129 um scan registered per site, on 60 fresh sites, rule fixed before the run; geometry only, never a reading",
   "sites_drawn": 60,
   "sites_scored": 42,
   "vertices": 310,
   "within_1_voxel": {
    "published": 0.242,
    "labelsnap": 0.71
   },
   "within_2_voxels": {
    "labelsnap": 0.82
   },
   "median_distance_voxels": {
    "published": 2.32,
    "labelsnap": 0.45
   },
   "gain": [
    1.78,
    1.4,
    2.21
   ],
   "controls": {
    "sham": -2.38,
    "random_sign": -0.58,
    "positive_control_2399um_face": 2.4
   },
   "merged_band_rule_cost": [
    0.52,
    0.32,
    0.75
   ],
   "note": "18 sites fell below the registration floor and were not scored; the merged-band rule's cost made the default drop it"
  },
  {
   "date": "2026-10-27",
   "tool": "kit finesnap snap (bands finder)",
   "source": "docs/tools/finesnap.md",
   "claim": "Model output: PHerc0139 surface label vertices snapped through a per-site registration onto the recto face of the 2.399 um scan, scored against the recto face of the 1.129 um scan registered per site, on 60 fresh sites, rule fixed before the run; geometry only, never a reading",
   "sites_scored": 60,
   "vertices": 476,
   "within_1_voxel": {
    "published": 0.237,
    "labelsnap": 0.794,
    "finesnap": 0.912,
    "positive_control_2399um_face": 0.903
   },
   "within_2_voxels": {
    "finesnap": 0.93
   },
   "median_distance_voxels": {
    "published": 2.37,
    "labelsnap": 0.37,
    "finesnap": 0.24
   },
   "gain_finesnap": [
    2.85,
    2.43,
    3.28
   ],
   "finesnap_minus_labelsnap": [
    0.65,
    0.41,
    0.93
   ],
   "controls": {
    "sham": -3.12
   },
   "limits": "the finder shares the truth's face definition and the truth is chained through the 2.4 um frame; misses concentrate where the 1.129 um band is split (33 % against 6 %); 11 of the 60 sites lie on three segments the Snapped Labels release later held back for evaluation, the test ran before that hold and its per-site rows were not kept"
  },
  {
   "date": "2026-10-29",
   "tool": "kit labelsnap snap against seven community snaps",
   "source": "docs/tools/snapped-labels.md",
   "claim": "Model output: on fresh PHerc0814 sites (1.129 um truth) and PHerc0841 sites (2.403 um truth), mean distance along the published normal to the finer scan's recto face, per site, then pooled; rule fixed before the run; geometry only, never a reading",
   "sites_scored": {
    "PHerc0814": 26,
    "PHerc0841": 45,
    "pooled": 71
   },
   "labelsnap_minus_published": [
    2.153,
    1.821,
    2.507
   ],
   "labelsnap_minus_vc_objrefine_tuned": [
    1.892,
    1.48,
    2.349
   ],
   "within_1_voxel": {
    "PHerc0814": {
     "published": 0.19,
     "labelsnap": 0.55,
     "finesnap": 0.85,
     "best_community": 0.22
    },
    "PHerc0841": {
     "published": 0.19,
     "labelsnap": 0.66,
     "best_community": 0.21
    }
   },
   "finesnap_minus_labelsnap_PHerc0814": [
    0.93,
    0.53,
    1.32
   ],
   "note": "Scored against a finer scan's face found with labelsnap's own band rule; with a face rule sharing neither, the lead held at about half the size (+1.04, +1.23 voxels).",
   "compared": "vc_objrefine (villa 382e48d35ae8) as shipped and with thresholds tuned on PHerc0139, recto-offset r2 (commit 16bc31f) with and without its gate, surface_placement_audit (villa PR #2147), RectoRefine's midline, a surface prediction snap",
   "limits": "two scrolls and these sites only; PHerc0814 is held out by site, not by scroll; the planted toward-verso control was weak; midline and nearest-peak snaps aim at the sheet center by design; on PHerc0841 the truth scan is the one the release's finesnap read, so the released PHerc0841 meshes were not scored; one-voxel exactness is not claimed"
  },
  {
   "date": "2026-11-03",
   "tool": "kit labelsnap snap against scroll-label-snap on real drift",
   "source": "scripts/experiments/lsnap-realdrift/chart.py",
   "claim": "Model output: a preregistered real drift test, scored once (Hugging Face Job 6aea627fc78863f9c6320c59). On 43 fresh PHerc0139 sites in one band (z 6194 to 7890), per-vertex distance along the published normal, in 9.362 um label voxels, to the recto face where the 2.403 um region scan and the 2.399 um scan agree within 1 voxel; 95 % cluster bootstrap intervals over 150-voxel cells; both snaps read only the label scan; geometry only, never a reading",
   "compared": "A.Q. Snyder's scroll-label-snap, https://github.com/aqsnyder/scroll-label-snap (MIT), pinned at 347ea5b4f3f7e346197034b7df599a4709aa0bbe, untuned, run on 13 x 13 vertex crops",
   "sites_drawn": 43,
   "sites_covered": 28,
   "vertices": 387,
   "vertices_scored": 173,
   "within_1_voxel": {
    "published": [
     0.173,
     0.105,
     0.257
    ],
    "labelsnap": [
     0.798,
     0.732,
     0.859
    ],
    "m7_snap": [
     0.214,
     0.132,
     0.317
    ],
    "scroll_label_snap_defaults": [
     0.179,
     0.093,
     0.257
    ],
    "scroll_label_snap_target_offset_0": [
     0.353,
     0.252,
     0.463
    ],
    "scroll_label_snap_target_offset_0_template_weight_0": [
     0.434,
     0.31,
     0.56
    ],
    "scroll_label_snap_authors_settings": [
     0.393,
     0.268,
     0.518
    ]
   },
   "median_signed_distance_voxels": {
    "published": [
     3.21,
     2.05,
     4.78
    ],
    "labelsnap": [
     0.02,
     -0.12,
     0.08
    ],
    "m7_snap": [
     1.89,
     0.88,
     3.19
    ],
    "scroll_label_snap_defaults": [
     2.69,
     2.16,
     3.16
    ],
    "scroll_label_snap_target_offset_0": [
     1.34,
     1.01,
     2.59
    ],
    "scroll_label_snap_target_offset_0_template_weight_0": [
     0.91,
     0.35,
     2.32
    ],
    "scroll_label_snap_authors_settings": [
     1.19,
     0.93,
     2.21
    ]
   },
   "interval_format": "[value, low, high]; signed distance positive means inside the sheet",
   "readout": "labelsnap ahead of all four scroll-label-snap settings (frozen rule, Bonferroni intervals over the four)",
   "controls": "the label scan's own face 0.954 within 1 voxel, 0.069 and 0.104 with the reference moved 1.5 voxels either way; the two reference scans agree within 1 voxel on 0.861 of 201 vertices where both name a face",
   "labelsnap_abstain_share": 0.191,
   "self_consistency_planted_shift": {
    "about": "108 earlier development sites (972 vertices): share returned within 1 voxel of the tool's own unshifted result after a planted 3 voxel shift inward (+3) or outward (-3); not accuracy; author's settings not run",
    "labelsnap": {
     "+3": 0.694,
     "-3": 0.563
    },
    "m7_snap": {
     "+3": 0.872,
     "-3": 0.867
    },
    "scroll_label_snap_defaults": {
     "+3": 0.079,
     "-3": 0.061
    },
    "scroll_label_snap_target_offset_0": {
     "+3": 0.202,
     "-3": 0.303
    },
    "scroll_label_snap_target_offset_0_template_weight_0": {
     "+3": 0.493,
     "-3": 0.717
    }
   },
   "limits": "one scroll and one band, 128 of the 173 scored vertices on w038; the two reference scans are both about 2.4 um at a similar energy and may share a face bias; scroll-label-snap is built for whole segments and ran untuned on small crops, and its default keeps the mesh's own offset by design; labelsnap's defaults were set on other PHerc0139 sites; a within-1 hit shows agreement with two other scans, not that the placement is on the papyrus; labelsnap and the reference both take the recto face of the nearest bright band, so where two sheets touch they can land on the inner neighbor's face together (check of 4 November 2026: another published segment crosses within 3.5 voxels inward of the face, or inside the band, at 12 of the 173 vertices; leaving out those 12 and the 19 vertices whose band is at least 12 voxels thick in both reference scans, 29 in all, keeps labelsnap ahead of all four settings at 0.826, and scoring those 29 against the outer sheet's face, 4.4 voxels in from the band's far side, keeps it ahead at 0.717)"
  },
  {
   "date": "2026-11-03",
   "tool": "kit labelsnap against scroll-label-snap (A.Q. Snyder) on spent PHerc0139 development sites",
   "source": "docs/data/pherc0139-snap-sites/README.md",
   "claim": "Model output, development only (spent sites, not a held-out test): share of covered vertices where a placement lies within 1 voxel of the recto face that the 1.129 um scan (20260413113053) and the ROI2 scan (20260319133554) agree on, positions in volume 20250728140407 level 0; geometry only, never a reading",
   "sites": 108,
   "windings": "w023 to w029, w031 to w034, w036 to w038",
   "covered_vertices": 601,
   "hit_rate": {
    "published": 0.298,
    "m7_snap": 0.301,
    "scroll_label_snap_defaults": 0.431,
    "scroll_label_snap_offset0": 0.579,
    "scroll_label_snap_offset0_template_weight0": 0.674,
    "labelsnap": 0.854
   },
   "compared": "scroll-label-snap (https://github.com/aqsnyder/scroll-label-snap) at commit 347ea5b4, unmodified, on 13 x 13 vertex crops",
   "data": "docs/data/pherc0139-snap-sites/pherc0139_twoscan_vertices.csv (sha256 14fb0a7c5a0bd3623bad9c80ab87b8b2d78e4e6a656f73940813cf92beb336a7), CC BY-NC 4.0",
   "limits": "one scroll, spent development sites, small crops; scroll-label-snap is built for whole meshes and was untuned here; labelsnap's face finder resembles the truth's face definition, so it may have a home advantage; not a general ranking"
  },
  {
   "date": "2026-10-30",
   "tool": "kit snap face per vertex on a +-16 voxel window with a +-12 fallback and no contrast floor (FACE_SNAP_WINDOW16; labelsnap --half 16 --fallback-half 12 --min-contrast 0) against community snaps on the PHercMANBp fragment",
   "source": "https://github.com/ScrollPrize/villa/pull/2368 (update of 30 Oct 2026); docs/tools/labelsnap.md",
   "claim": "Model output, a comparison with every rule fixed before the read and scored once: share of vertices within 1 voxel of the recto face in a second registered scan, an abstention counted as a miss; geometry only, never a reading; a comparison, not proof of 1 voxel accuracy",
   "scans": "label 20251216152116 (2.399 um, read at 9.596 um); truth 20260427100434 (1.129 um, read at 9.03 um), linked by its published transform.json and registered per site",
   "sites_drawn": 300,
   "sites_with_face": 220,
   "within_1_voxel": {
    "window16": 0.927,
    "rectorefine_recto_edge": 0.691,
    "published": 0.245,
    "recto_offset_r2_ungated": 0.232,
    "recto_offset_r2_min_snr_2": 0.236,
    "vc_objrefine_shipped": 0.059,
    "vc_objrefine_full_resolution": 0.027,
    "vc_objrefine_tuned": 0.205
   },
   "window16_hits": 204,
   "lead_over_best_lower_bound": 0.17,
   "lead_bound_note": "one-sided 97.5 %, clustered by 150 voxel cell; recto-offset failed its own push-back check and was not counted",
   "compared": "RectoRefine recto edge (ScrollFiesta bce8a65), recto-offset r2 (16bc31f), vc_objrefine (villa 382e48d), grow3d snap, RectoRefine midline, surface_placement_audit (villa PR #2147)",
   "limits": "one fragment, one truth scan at about the label voxel size; the truth face is found with the same half-level face rule, which favors a method built on it; scored by the test harness on its own profiles, equal to the snap's per-vertex offsets on synthetic profiles; run without the default contrast floor"
  },
  {
   "date": "2026-10-29",
   "tool": "the same +-16 window with a +-12 fallback against community snaps on a band of PHerc0332",
   "source": "https://github.com/ScrollPrize/villa/pull/2368 (update of 30 Oct 2026); docs/tools/labelsnap.md",
   "claim": "Model output, a test with every rule fixed before the read and scored once; the comparison held, the 0.90 exactness bar was not reached; geometry only, never a reading",
   "truth": "two 3.240 um scans of PHerc0332, which agree with each other within 1 voxel at 0.869",
   "sites_with_face": 326,
   "within_1_voxel": {
    "window16": 0.782,
    "rectorefine_recto_edge": 0.604
   },
   "lead_over_best_lower_bound": 0.12,
   "exactness": {
    "answered_within_1_voxel": 0.859,
    "upper_bound": 0.896,
    "bar": 0.9
   },
   "limits": "one band of one scroll; the exactness bar sat above what the two truth scans reach together"
  }
 ],
 "parting_checks": [
  {
   "date": "2026-10-29",
   "tool": "Parting (contact classifier; code released under Apache-2.0, model file SharpMeow/parting-v1 on Hugging Face)",
   "source": "docs/logs/2026-10-29-parting-fragment-contact.md",
   "claim": "Model output: Parting, a contact classifier on base dinovol (scrollprize/dinovol_ps8_9.6um_101626) token, CT and normal-profile features trained on PHerc0139 and PHerc0172 hand-traced windings, against base dinovol's raw token cosine on the same sites, telling a same-layer point from a next-layer point at touching sites (both gaps 3 to 8 voxels) of PHercParis1Fr34 (Frag3), a fragment of PHerc. Paris 1 that neither model saw; truth the team's six layer segments of October 2024; standardized 54 keV 3.24 um volume box-averaged to 7.91 um; rule frozen before any of the test's 3D CT was read, read once",
   "claim_wording": "touching AUC 0.8952 against 0.8392, gain +0.0560 with 2.5 % bound +0.0368 on 150 sites over 5 layer pairs; one pass in three fragments, after a FAIL on PHerc1667Cr1Fr3 (+0.045) and before a FAIL on PHercParis2Fr143 (+0.0103)",
   "verdict": "PASS",
   "rule": "PASS when the touching AUC is at least 0.75, the gain over base dinovol at least 0.05 and the 2.5 % bound of the gain (20,000 resamples of layer pairs, seed 0) above 0; VOID when fewer than 100 touching sites or 5 pairs, base dinovol under 0.85 on apart sites, or the mean AUC of 1,000 shuffled-label permutations outside 0.5 plus or minus 3 standard errors",
   "touching": {
    "sites": 150,
    "pairs": 5,
    "model": 0.8952,
    "model_bound": 0.8452,
    "dino_base": 0.8392,
    "delta": 0.056,
    "delta_bound": 0.0368
   },
   "controls": {
    "permutation_mean": 0.50039,
    "permutation_half_width": 0.00311,
    "single_permutation": 0.5413,
    "pipeline_apart_dino_base": 0.9438,
    "apart_sites": 75
   },
   "per_pair_delta": {
    "L0-L1": 0.0256,
    "L1-L2": 0.0711,
    "L2-L3": 0.0256,
    "L3-L4": 0.0533,
    "L4-L5": 0.0233
   },
   "without_L4-L5_delta": 0.057,
   "reported_not_judged": {
    "fused": {
     "sites": 38,
     "model": 0.633,
     "dino_base": 0.6108
    },
    "touching_gap_le4": {
     "sites": 3,
     "model": 0.3333,
     "dino_base": 0.3333
    }
   },
   "compute": {
    "flavor": "rtx-pro-6000",
    "seconds": 425,
    "usd": 0.37
   },
   "earlier_attempt": {
    "object": "PHerc1667Cr1Fr3 (Frag5)",
    "date": "2026-10-29",
    "sites": 203,
    "pairs": 7,
    "model": 0.96,
    "dino_base": 0.915,
    "delta": 0.045,
    "delta_bound": 0.02,
    "verdict": "FAIL"
   },
   "later_attempt": {
    "object": "PHercParis2Fr143 (Frag2)",
    "date": "2026-10-29",
    "sites": 270,
    "pairs": 9,
    "apart_sites": 135,
    "model": 0.9754,
    "model_bound": 0.9615,
    "dino_base": 0.9651,
    "delta": 0.0103,
    "delta_bound": -0.0005,
    "permutation_mean": 0.50057,
    "pipeline_apart_dino_base": 0.9988,
    "per_pair_delta_range": [
     -0.029,
     0.03
    ],
    "usd": 0.64,
    "verdict": "FAIL",
    "note": "Interpretation: base dinovol near its ceiling, so the largest possible gain was +0.035, under the +0.05 bar"
   }
  }
 ],
 "label_suggest_checks": [
  {
   "date": "2026-10-29",
   "tool": "kit labelsuggest map, --measure vote (the default)",
   "source": "docs/tools/labelsuggest.md",
   "claim": "Measurement: the 20 tiles the vote measure ranked first held more of the committee's errors than random tiles, on three PHerc1667 segments; scored against the organizers' own ink label rounds in the public Scroll Prize ink dataset (the ink folder of scrollprize/datasets on Hugging Face, listed in docs/external.md): version 1 labels and supervision (7 March 2026), version 2 (24 March 2026), and the organizers' model predictions published in the same segment folders and dated between the two rounds (the committee of readers). An error is a tile pixel where the committee's call (mean probability at least 0.5) differs from the version 2 label, inside the area that version 2 newly supervised",
   "claim_wording": "On PHerc1667 the 20 tiles the reader-disagreement pick ranked first held 3.82 times the average tile's error share (geometric mean over three segments, bar 1.5, p 0.0001 in each); a label-free ink-edge pick did as well (4.17), so the tool is not shown to beat simple picks",
   "verdict": "PASS",
   "rule": "fixed and frozen before any test segment was read: PASS when the vote measure's geometric-mean top-20 lift over usable segments is at least 1.5 and at least three quarters of the usable segments (3 of 3 here) have lift above 1 with p below 0.05 against 10,000 random 20-tile draws; VOID when fewer than 3 segments are usable or the pixel-shuffle control (mean lift 0.8 to 1.25) or the planted control (at least 15 of 20 found) fails; a segment is usable only when the committee's AUC drops by at least 0.10 with the labels moved 32 px; one run",
   "segments": {
    "usable": 3,
    "lost_to_a_runner_fault": 1,
    "readers_per_segment": 7,
    "tiles": [
     1183,
     354,
     431
    ]
   },
   "lift_top20": {
    "vote": 3.82,
    "margin": 4.08,
    "edge": 4.17,
    "ink_density": 1.26
   },
   "controls": {
    "pixel_shuffle_geomean": 0.935,
    "pixel_shuffle_band": [
     0.8,
     1.25
    ],
    "planted_found_min": 17,
    "planted_bar": 15,
    "committee_auc_in_place": [
     0.922,
     0.943,
     0.978
    ],
    "committee_auc_labels_moved_32px": [
     0.753,
     0.743,
     0.767
    ]
   },
   "attempt": "attempt 1 on a first scroll",
   "limits": "one scroll; the edge and margin picks, both label-free, lifted at least as much, so the PASS is not a win over simple picks; errors are model calls against later labels, not label errors; whether labeling at these tiles saves a person's time is not tested; the organizers may have looked at these predictions while labeling version 2, and the latest predictions may have trained on interim version 2 labels (both untested); the test runner is not in this repository"
  },
  {
   "date": "2026-10-29",
   "tool": "kit labelsuggest map, --measure edge",
   "source": "docs/tools/labelsuggest.md",
   "claim": "Measurement: on a second scroll, PHerc0841 segment ag896 (auto_grown_20260220144552896), the 20 tiles a label-free ink-edge pick ranked first held more of the committee's errors than random tiles; scored against the organizers' own ink label rounds in the public Scroll Prize ink dataset (the ink folder of scrollprize/datasets on Hugging Face, listed in docs/external.md): version 1 labels and supervision (7 March 2026), version 2 (24 March 2026), and the organizers' model predictions published in the same segment folders and dated between the two rounds (the committee of readers). An error is a tile pixel where the committee's call (mean probability at least 0.5) differs from the version 2 label, inside the area that version 2 newly supervised",
   "claim_wording": "On a second scroll (PHerc0841 ag896, one segment, three readers, 108 tiles), the 20 tiles a label-free ink-edge pick suggested held 1.84 times the average share of the committee's disagreements with the organizers' later labels (bar 1.5, p 0.0005 against random 20-tile picks). This was attempt 2 at a second-scroll test, and the edge pick was chosen on spent development data after a misplacement control fixed in advance failed and was made reported only once its numbers were seen. The pick is no better than margin or vote by any test we ran",
   "verdict": "PASS",
   "rule": "fixed and frozen before any PHerc0841 file was read: PASS when the edge pick's top-20 lift is at least the bar (1.5 for 60 to 239 eligible tiles, 2.0 from 240) with p below 0.05 against 10,000 random 20-tile draws; VOID when the segment is unusable or a control fails (planted at least 15 of 20, broken readout at most 5 of 20, random-position mean lift 0.8 to 1.25); one run",
   "segments": {
    "usable": 1,
    "readers": 3,
    "tiles": 108,
    "average_error_share": 0.147
   },
   "lift_top20": {
    "edge": 1.84,
    "margin": 1.96,
    "vote": 1.87,
    "ink_density": 1.14
   },
   "p_random": 0.0005,
   "controls": {
    "planted_found": 20,
    "planted_bar": 15,
    "broken_readout_found": 0,
    "broken_bar": 5,
    "random_position_mean_lift": 1.0014,
    "committee_auc_in_place": 0.934,
    "committee_auc_labels_moved_32px": 0.685
   },
   "attempt": "attempt 2 at a second-scroll test (the first was never frozen, so no data was spent on it)",
   "limits": "one segment of one scroll, three readers, one of them a prediction whose file name says reverse; the edge pick was chosen on PHerc1667 data already used by the first test, after a disclosed amendment; margin and vote lifted as much or more, and no comparison between picks was a preregistered claim; errors are model calls against later labels, not label errors; whether labeling at these tiles saves a person's time is not tested; the organizers may have looked at these predictions while labeling version 2, and the latest predictions may have trained on interim version 2 labels (both untested); the test runner is not in this repository"
  },
  {
   "date": "2026-11-03",
   "tool": "kit altrain (training-run comparison) with kit labelsuggest's edge measure as SUGGEST",
   "source": "docs/tools/altrain.md",
   "claim": "Measurement on model output: a simulated labeling round with oracle labels on three Paris fragments (Frag2 PHercParis2Fr143, Frag3 PHercParis1Fr34, Frag4 PHercParis1Fr39: the public 54 keV 3.24 um exposed-surface volumes of dl.ash2txt.org/fragments mean-pooled 3 x 3 x 3 to 9.72 um, 21 layers, with their infrared-photograph ink labels pooled the same way). Per fragment a fixed label-free layout split the labeled area into 256 px test blocks and a pool of 64 px tiles one tile away from them. Each arm revealed the labels of k pool tiles: SUGGEST (top k by labelsuggest's edge measure on scrollprize/ink_9um seeds 42 and 43, revision 7109667e), UNCERTAIN (top k by entropy of seed 42), RANDOM (5 draws) and SHUF (top k of shuffled SUGGEST scores, a control). ink_9um seed 42, which trained on no fragment, was fine-tuned 300 steps on each arm's tiles at 5, 10 and 20 % of the pool, 3 seeds, and mapped the test blocks forward and with reversed depth. Rule, controls and bars were preregistered and frozen before any fine-tune, after an independent review; each fragment was read once.",
   "claim_wording": "Not shown: fine-tuning on the tiles the edge measure picked did not beat random tiles at any budget on any fragment (no lower bound above 0), and at 5 % it was below random on Frag3 and Frag4 (about -0.10 AUC). The fragment readout is VOID by its rule: only Frag4 counted; Frag2 and Frag3 failed the displaced-label control (the base reader's map does not track their labels by the 0.05 margin), and some arms on Frag2 and Frag3 also crossed the 0.55 reverse-depth bar",
   "verdict": "VOID",
   "counted": [
    "frag4"
   ],
   "passing": [],
   "rule": "PASS needs two counted segments each with SUGGEST minus RANDOM's paired hierarchical bootstrap lower bound (2000 draws resampling test blocks and runs, 2.5th percentile) above 0, and its forward gain larger than its reverse-depth gain, at two or more of the three budgets; a segment is VOID when any arm's mean reverse-depth AUC exceeds 0.55, shuffled suggestions differ from RANDOM (99 % interval pooled over budgets), the base reader's test AUC beats the labels-moved-32-px AUC by less than 0.05, or the pool or test is too small",
   "segments": {
    "frag2": {
     "state": "VOID",
     "why_void": [
      "labels do not track the CT: base AUC 0.644531 in place against 0.598413 moved 32 px",
      "budget 0.05: RANDOM mean reverse-depth AUC 0.561457 > 0.55",
      "budget 0.05: SHUF mean reverse-depth AUC 0.554669 > 0.55",
      "budget 0.10: SUGGEST mean reverse-depth AUC 0.575576 > 0.55",
      "budget 0.10: RANDOM mean reverse-depth AUC 0.562512 > 0.55",
      "budget 0.10: SHUF mean reverse-depth AUC 0.561964 > 0.55",
      "budget 0.20: SUGGEST mean reverse-depth AUC 0.572629 > 0.55",
      "budget 0.20: UNCERTAIN mean reverse-depth AUC 0.572831 > 0.55",
      "budget 0.20: RANDOM mean reverse-depth AUC 0.567018 > 0.55",
      "budget 0.20: SHUF mean reverse-depth AUC 0.564143 > 0.55"
     ],
     "base_test_auc": 0.6445,
     "base_test_auc_labels_moved_32px": 0.5984,
     "pool_tiles": 912,
     "test_blocks": 57,
     "budgets": {
      "0.05": {
       "mean_auc": {
        "SUGGEST": 0.713062,
        "UNCERTAIN": 0.709675,
        "RANDOM": 0.718903,
        "SHUF": 0.719426
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.538023,
        "UNCERTAIN": 0.526999,
        "RANDOM": 0.561457,
        "SHUF": 0.554669
       },
       "suggest_minus_random": {
        "est": -0.005841,
        "lower": -0.031624,
        "upper": 0.019753
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.017593,
        "lower": -0.012866
       },
       "uncertain_minus_random": {
        "est": -0.009228,
        "lower": -0.034837,
        "upper": 0.017208
       },
       "shuffled_minus_random": {
        "est": 0.000522,
        "lower": -0.020258,
        "upper": 0.02274
       },
       "suggest_revealed_ink_over_random": 1.6
      },
      "0.10": {
       "mean_auc": {
        "SUGGEST": 0.752589,
        "UNCERTAIN": 0.749727,
        "RANDOM": 0.735805,
        "SHUF": 0.748152
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.575576,
        "UNCERTAIN": 0.549666,
        "RANDOM": 0.562512,
        "SHUF": 0.561964
       },
       "suggest_minus_random": {
        "est": 0.016784,
        "lower": -0.006201,
        "upper": 0.040847
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.003721,
        "lower": -0.026498
       },
       "uncertain_minus_random": {
        "est": 0.013922,
        "lower": -0.010232,
        "upper": 0.039265
       },
       "shuffled_minus_random": {
        "est": 0.012347,
        "lower": -0.010515,
        "upper": 0.037486
       },
       "suggest_revealed_ink_over_random": 1.59
      },
      "0.20": {
       "mean_auc": {
        "SUGGEST": 0.769624,
        "UNCERTAIN": 0.768109,
        "RANDOM": 0.760662,
        "SHUF": 0.762181
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.572629,
        "UNCERTAIN": 0.572831,
        "RANDOM": 0.567018,
        "SHUF": 0.564143
       },
       "suggest_minus_random": {
        "est": 0.008962,
        "lower": -0.006585,
        "upper": 0.025196
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.003351,
        "lower": -0.021457
       },
       "uncertain_minus_random": {
        "est": 0.007447,
        "lower": -0.012224,
        "upper": 0.026509
       },
       "shuffled_minus_random": {
        "est": 0.00152,
        "lower": -0.014823,
        "upper": 0.018946
       },
       "suggest_revealed_ink_over_random": 1.47
      }
     },
     "shuffle_control": {
      "est": 0.004796,
      "lo": -0.010185,
      "hi": 0.022511,
      "holds": true
     }
    },
    "frag3": {
     "state": "VOID",
     "why_void": [
      "labels do not track the CT: base AUC 0.567359 in place against 0.568890 moved 32 px",
      "budget 0.10: RANDOM mean reverse-depth AUC 0.550020 > 0.55",
      "budget 0.20: RANDOM mean reverse-depth AUC 0.551229 > 0.55",
      "budget 0.20: SHUF mean reverse-depth AUC 0.561037 > 0.55"
     ],
     "base_test_auc": 0.5674,
     "base_test_auc_labels_moved_32px": 0.5689,
     "pool_tiles": 270,
     "test_blocks": 13,
     "budgets": {
      "0.05": {
       "mean_auc": {
        "SUGGEST": 0.557527,
        "UNCERTAIN": 0.534149,
        "RANDOM": 0.663957,
        "SHUF": 0.637735
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.49202,
        "UNCERTAIN": 0.515826,
        "RANDOM": 0.540043,
        "SHUF": 0.524565
       },
       "suggest_minus_random": {
        "est": -0.10643,
        "lower": -0.164077,
        "upper": -0.054272
       },
       "suggest_gain_forward_minus_reverse": {
        "est": -0.058408,
        "lower": -0.106297
       },
       "uncertain_minus_random": {
        "est": -0.129808,
        "lower": -0.206895,
        "upper": -0.066683
       },
       "shuffled_minus_random": {
        "est": -0.026222,
        "lower": -0.075023,
        "upper": 0.023561
       },
       "suggest_revealed_ink_over_random": 0.49
      },
      "0.10": {
       "mean_auc": {
        "SUGGEST": 0.658821,
        "UNCERTAIN": 0.615172,
        "RANDOM": 0.691748,
        "SHUF": 0.682795
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.513725,
        "UNCERTAIN": 0.546873,
        "RANDOM": 0.55002,
        "SHUF": 0.546678
       },
       "suggest_minus_random": {
        "est": -0.032927,
        "lower": -0.094024,
        "upper": 0.017189
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.003369,
        "lower": -0.060417
       },
       "uncertain_minus_random": {
        "est": -0.076576,
        "lower": -0.132745,
        "upper": -0.03289
       },
       "shuffled_minus_random": {
        "est": -0.008953,
        "lower": -0.049636,
        "upper": 0.02702
       },
       "suggest_revealed_ink_over_random": 0.89
      },
      "0.20": {
       "mean_auc": {
        "SUGGEST": 0.689616,
        "UNCERTAIN": 0.639359,
        "RANDOM": 0.710025,
        "SHUF": 0.696217
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.52867,
        "UNCERTAIN": 0.547591,
        "RANDOM": 0.551229,
        "SHUF": 0.561037
       },
       "suggest_minus_random": {
        "est": -0.020408,
        "lower": -0.063909,
        "upper": 0.016703
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.00215,
        "lower": -0.048348
       },
       "uncertain_minus_random": {
        "est": -0.070665,
        "lower": -0.103645,
        "upper": -0.029062
       },
       "shuffled_minus_random": {
        "est": -0.013807,
        "lower": -0.050526,
        "upper": 0.015431
       },
       "suggest_revealed_ink_over_random": 1.12
      }
     },
     "shuffle_control": {
      "est": -0.016328,
      "lo": -0.046052,
      "hi": 0.009285,
      "holds": true
     }
    },
    "frag4": {
     "state": "COUNTED",
     "why_void": [],
     "base_test_auc": 0.6419,
     "base_test_auc_labels_moved_32px": 0.5822,
     "pool_tiles": 485,
     "test_blocks": 24,
     "budgets": {
      "0.05": {
       "mean_auc": {
        "SUGGEST": 0.574422,
        "UNCERTAIN": 0.671241,
        "RANDOM": 0.675365,
        "SHUF": 0.668893
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.48714,
        "UNCERTAIN": 0.509413,
        "RANDOM": 0.512247,
        "SHUF": 0.523867
       },
       "suggest_minus_random": {
        "est": -0.100943,
        "lower": -0.16311,
        "upper": -0.044128
       },
       "suggest_gain_forward_minus_reverse": {
        "est": -0.075835,
        "lower": -0.171948
       },
       "uncertain_minus_random": {
        "est": -0.004124,
        "lower": -0.040483,
        "upper": 0.029335
       },
       "shuffled_minus_random": {
        "est": -0.006472,
        "lower": -0.053808,
        "upper": 0.038615
       },
       "suggest_revealed_ink_over_random": 1.56
      },
      "0.10": {
       "mean_auc": {
        "SUGGEST": 0.69113,
        "UNCERTAIN": 0.730165,
        "RANDOM": 0.706613,
        "SHUF": 0.688187
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.509547,
        "UNCERTAIN": 0.516982,
        "RANDOM": 0.536664,
        "SHUF": 0.49883
       },
       "suggest_minus_random": {
        "est": -0.015483,
        "lower": -0.055753,
        "upper": 0.022213
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.011635,
        "lower": -0.05581
       },
       "uncertain_minus_random": {
        "est": 0.023551,
        "lower": -0.009012,
        "upper": 0.056703
       },
       "shuffled_minus_random": {
        "est": -0.018426,
        "lower": -0.061165,
        "upper": 0.02381
       },
       "suggest_revealed_ink_over_random": 1.52
      },
      "0.20": {
       "mean_auc": {
        "SUGGEST": 0.749102,
        "UNCERTAIN": 0.742769,
        "RANDOM": 0.733703,
        "SHUF": 0.743848
       },
       "mean_reverse_depth_auc": {
        "SUGGEST": 0.529017,
        "UNCERTAIN": 0.522257,
        "RANDOM": 0.522538,
        "SHUF": 0.524368
       },
       "suggest_minus_random": {
        "est": 0.0154,
        "lower": -0.007651,
        "upper": 0.037919
       },
       "suggest_gain_forward_minus_reverse": {
        "est": 0.00892,
        "lower": -0.024877
       },
       "uncertain_minus_random": {
        "est": 0.009067,
        "lower": -0.017937,
        "upper": 0.036857
       },
       "shuffled_minus_random": {
        "est": 0.010146,
        "lower": -0.022654,
        "upper": 0.040614
       },
       "suggest_revealed_ink_over_random": 1.36
      }
     },
     "shuffle_control": {
      "est": -0.004918,
      "lo": -0.031151,
      "hi": 0.021427,
      "holds": true
     }
    }
   },
   "constants_sha256": "4fe0bf2d9f45318d814cffa27f21d25834f1b482e31f91e3b39f23e12004e00b",
   "scope": "One base reader, one short recipe, fragments only (exposed surfaces, not buried ink). The private preregistration also had a scroll segment, which stays private. A suggested tile is a place to label, never a reading; the labels were already drawn, so this says nothing of a person's labeling time.",
   "why": "A villa maintainer asked for a training run comparison before taking the active-learning tool upstream (villa PR #2381).",
   "constants_sha256_note": "This is the public module's constants hash. It differs from the hash of the frozen rule the run used only in the segment list; every other constant is the same.",
   "rounding_note": "suggest_revealed_ink_over_random is rounded to 2 decimals; the base AUC fields are at 4 decimals; the arm AUCs and differences are as the readout printed them (6 decimals).",
   "base_reader_training_data": "scrollprize/ink_9um: its label dataset README, in the ink_9um folder of scrollprize/datasets on Hugging Face (its URL is in docs/external.md; file date 14 Aug 2026), lists 24 aligned segments of PHerc. 0139, Scroll 1667, PHerc. Paris 4 and PHerc. 0814 and 5 native PHerc. 0139 segments, and no fragment; model card https://huggingface.co/scrollprize/ink_9um"
  }
 ],
 "sqrecipe_checks": [
  {
   "key": "sqrecipe_test_0332_20261030",
   "date": "2026-10-30",
   "tool": "scripts/sqrecipe/sqrecipe.py test (kit.sqrecipe, kit.sqfid)",
   "source": "docs/tools/sqrecipe.md",
   "label": "Measurement of the CT (frozen test, PHerc0332: DLS 7.910 um 53 keV 20231117143551 native against DLS 3.240 um 53 keV 20231027191953 level 1 resampled onto its grid, both scored against ESRF 2.399 um 78 keV 20251211183505 level 1)",
   "question": "In fresh 128-voxel regions of PHerc0332's 7.910 um grid, does the DLS 3.240 um 53 keV acquisition, resampled onto that grid, keep more sheet fidelity against an independent ESRF 2.399 um scan than the native DLS 7.910 um 53 keV acquisition? Fidelity is band-passed normalized cross-correlation at the best integer shift (sqfid).",
   "rule": "fixed before the read, read once: PASS when share(d > 0) and median d, d = F(B) - F(N), have 95 % cluster-bootstrap lower bounds above 0.5 and 0, the processing control holds (share(F(B) > F(N')) lower bound above 0.5) and the degradation control holds (median(F(Bdeg) - F(N)) at most half of median d); VOID under 60 regions or 10 clusters, or when the outcome control or border check of N or B, or the sign-flip control, fails",
   "verdict": "PASS",
   "regions": 120,
   "primary_share": 0.925,
   "primary_share_ci": [
    0.872,
    0.968
   ],
   "primary_median_d": 0.042,
   "primary_median_ci": [
    0.035,
    0.055
   ],
   "processing_share": 0.942,
   "signflip": "5 of 200",
   "outcome_controls": "held",
   "limits": "one acquisition pair on one scroll; it measures CT against CT, not ink and not readability; PHerc0500P2's two scans are a near-identical recipe pair (a null pair), so there is no clean second scroll yet",
   "attempts": "second freeze; the first never read any region (its smoke check refused it); the outcome rule (shift search plus or minus 10 voxels, sigmas 0.7 and 2.0, reference at level 1) was amended after development looks on spent regions, before this freeze"
  }
 ],
 "offsetdiag_checks": [
  {
   "key": "offsetdiag_dev_20261030",
   "date": "2026-10-30",
   "tool": "scripts/offsetdiag (ink offset diagnostic; code released under Apache-2.0)",
   "source": "docs/logs/2026-10-30-ink-offset.md",
   "label": "Model output",
   "status": "development only, no frozen test",
   "claim": "Model output: Reader V3 (domenicor046/reader-v3) at layers 6 to 23, hann, overlap 0.5, both directions, on PHerc0841 w00 and ag896 (9.366 um scan 20250821151531) inside the development scoring boxes of earlier ink rounds, with the surface moved along the published stack's normal (minus is toward layer 0); AUC inside the supervision mask with the reverse-depth control, as stored and texture registrations, arm minus our unshifted render R0 with a paired block bootstrap",
   "gate": "before any shift, Reader V3 on the published boxes reproduced its earlier scores exactly",
   "auc_loss_vs_R0": {
    "minus_4": {
     "w00": 0.07,
     "ag896": 0.04
    },
    "plus_4": {
     "w00": 0.02,
     "ag896": 0.02,
     "note": "not significant"
    },
    "minus_8": {
     "w00": 0.25,
     "ag896": 0.21
    },
    "plus_8": {
     "w00": 0.16,
     "ag896": 0.12
    }
   },
   "toward_layer_0_cost_ratio": [
    1.5,
    4.1
   ],
   "snap_to_lasagna_sheet_center": {
    "auc_loss": [
     0.03,
     0.08
    ],
    "beat_same_size_controls": true,
    "controls": "ANTI (same moves, other way) and RAND (field turned 180 degrees)"
   },
   "second_smaller_model": "a smaller ink model of ours (not released) lost less per shift; Reader V3 stayed above it at every shift, smallest gap 0.03",
   "renderer_effect": {
    "w00": 0.007,
    "ag896": -0.009,
    "note": "our render at offset 0 against the published surface volume, Reader V3"
   },
   "limits": "development only; two segments of one scroll; one main model; both segments already used in earlier development; not a test, no level claimed"
  }
 ],
 "ringleader_checks": [
  {
   "key": "ringleader_test2_20261030",
   "about": "Ringleader v1 (slicelab.PARAMS with V1, slab3d.ZPARAMS) 3D winding labels on m7 sheet pixels of held-out PHerc0139 z 3,240 to 3,400 (frame 20250728140407, 41 slices every 4th, in two stacks), scored by on-winding E0 against 14 published hand windings, beside villa's spiral fit at its default settings (villa commit e2254ebe, seeds 1 to 3) read one label per Ringleader piece; rule and bars frozen before the read; read once",
   "label": "Model output",
   "verdict": "PASS",
   "rule": "PASS when Ringleader's pooled E0 is at least 0.85, Ringleader minus the villa fit's E0 is at least 0.40 at every seed, and Ringleader's planted, shift and jitter controls behave on every slice",
   "slices": 41,
   "e0_ringleader": 0.902,
   "bar_e0": 0.85,
   "villa_fit_e0_range": [
    0.149,
    0.179
   ],
   "margins": [
    0.72,
    0.75
   ],
   "bar_margin": 0.4,
   "controls": "planted, shift and jitter controls failed on all 41 slices, as they should",
   "references": "14 published PHerc0139 hand windings",
   "independent_rerun": "MATCH (a separate rerun of the frozen readout gave the same result)",
   "metric": "on-winding E0: share of matched m7 sheet pixels whose label minus true turn equals the slice's most common offset, pooled weighted by matched labeled pixels",
   "limits": "E0 counts m7 sheet pixels only; truth is hand windings on one scroll; it labels windings and does not trace a surface. On PHerc0332, m7 may have trained on that scroll's older scans or meshes, and earlier tests of ours had read the same band, so it is not fully held out there; PHerc0139 is held out by band only, since the settings and bars were set on development stacks of the same scroll. villa's spiral fit is not built to label every sheet of a slice and drifts by several turns, so the margin is over its default fit read this way, not over a tuned winding labeler.",
   "source": "docs/tools/ringleader.md#checks-and-controls",
   "date": "2026-10-30"
  },
  {
   "key": "ringleader_0332_l4_20261030",
   "about": "Ringleader v1 3D winding labels (as in ringleader_test2_20261030) on m7 sheet pixels of PHerc0332 scan 20251211183505 z 3,760 to 4,340 (146 slices every 4th, in seven stacks), scored by on-winding E0 against organizer mesh m0828 (about 3 turns) placed by the catalog transform alone, beside villa's spiral fit at its default settings (villa commit e2254ebe, seeds 1 to 3) read per piece, a constant labeling and two negative labelings made from Ringleader's; read once",
   "label": "Model output",
   "verdict": "PASS",
   "rule": "PASS when Ringleader's pooled E0 is at least the larger of 0.50 and the constant labeling plus 0.15, Ringleader beats each villa seed and the constant labeling by at least 0.15, its controls pass and both negative labelings fail",
   "slices": 146,
   "stacks": 7,
   "e0_ringleader": 0.659,
   "bar_e0": 0.595,
   "villa_fit_e0_range": [
    0.349,
    0.398
   ],
   "villa_margins": [
    0.26,
    0.31
   ],
   "constant_e0": 0.445,
   "constant_margin": 0.21,
   "negatives": {
    "permutation": {
     "e0": 0.302,
     "verdict": "FAIL"
    },
    "sector_swap": {
     "e0": 0.5,
     "verdict": "FAIL"
    }
   },
   "villa_below_constant": "every villa seed scored below the constant labeling, so the margin over the constant labeling carries the claim",
   "independent_rerun": "MATCH (a separate rerun of the frozen readout gave the same result)",
   "later_tries": "Two later tries on other scrolls read VOID and would have missed their bars (on PHerc0172, 0.458 against 0.50), so it is claimed on two scrolls, not more",
   "limits": "E0 counts m7 sheet pixels only; truth is one organizer mesh of about 3 turns; it labels windings and does not trace a surface. On PHerc0332, m7 may have trained on that scroll's older scans or meshes, and earlier tests of ours had read the same band, so it is not fully held out there; PHerc0139 is held out by band only, since the settings and bars were set on development stacks of the same scroll. villa's spiral fit is not built to label every sheet of a slice and drifts by several turns, so the margin is over its default fit read this way, not over a tuned winding labeler.",
   "source": "docs/tools/ringleader.md#checks-and-controls",
   "date": "2026-10-30",
   "slices_scored": 142
  },
  {
   "key": "ringleader_ladders_0139_20261104",
   "about": "Ringleader v1 sure pieces on 10 evidence slices of PHerc0139 (z 1,920 to 1,956) given to villa's spiral fit (commit e2254ebe, 30,000 steps, fit window z 1,024 to 2,048 of its L2 grid) as relative windings by ladders.py's rule; arms A (fit alone), E (round 1 ladders), G (grown ladders), Q (G's plus-one control), seeds 1 to 3; 86 slices of fresh z 1,960 to 2,300 scored by on-winding E0 against 11 published hand windings; read once",
   "label": "Model output",
   "verdict": "PASS",
   "rule": "PASS when G minus A is at least +0.10 pooled and above 0 at every seed, and G minus Q is at least 0.05 at every seed",
   "slices": 86,
   "seeds": [
    1,
    2,
    3
   ],
   "e0": {
    "A": [
     0.219,
     0.23,
     0.232
    ],
    "E": [
     0.36,
     0.38,
     0.38
    ],
    "G": [
     0.415,
     0.414,
     0.397
    ],
    "Q": [
     0.103,
     0.105,
     0.112
    ],
    "ringleader": 0.79
   },
   "e0_mean": {
    "A": 0.227,
    "E": 0.373,
    "G": 0.409,
    "Q": 0.107
   },
   "g_minus_a_pooled": 0.182,
   "g_minus_a": [
    0.196,
    0.184,
    0.165
   ],
   "g_minus_q_range": [
    0.285,
    0.312
   ],
   "limits": "One band of one scroll; a later frozen test on another scroll did not clear its +0.10 bar. The taught fit stays below Ringleader's own labels (0.41 against 0.79). The test harness and data staging are not released.",
   "source": "docs/tools/ringleader.md#checks-and-controls",
   "date": "2026-11-04"
  },
  {
   "key": "ringleader_ladders_0172_20261106",
   "about": "The ringleader_ladders_0139_20261104 method, constants and pass rule, unchanged, on PHerc0172 scan 20241024131838 (7.91 um): our run of the organizers' m7 model on the band, villa's lasagna model for the fit's field, a center line picked from three public candidates by a label-free rule; Ringleader v1 sure pieces on 10 evidence slices (z 9,640 to 9,676) as relative windings; villa's spiral fit (commit e2254ebe, 30,000 steps, fit window z 9,280 to 10,304), arms A, E, G and Q, seeds 1 to 3; 57 slices of z 9,680 to 9,904 (every 4th, three stacks of 19) scored by on-winding E0 against the organizers' hand windings w052 to w095 (44); read once",
   "label": "Model output",
   "verdict": "PASS",
   "rule": "PASS when G minus A is at least +0.10 pooled and above 0 at every seed, and G minus Q is at least 0.05 at every seed; void unless Ringleader itself leads mean A by at least 0.20 (added before this test's freeze, after the failed test on another scroll)",
   "slices": 57,
   "refs": 44,
   "seeds": [
    1,
    2,
    3
   ],
   "e0": {
    "A": [
     0.07,
     0.066,
     0.072
    ],
    "E": [
     0.268,
     0.267,
     0.27
    ],
    "G": [
     0.293,
     0.293,
     0.278
    ],
    "Q": [
     0.038,
     0.038,
     0.035
    ],
    "ringleader": 0.497
   },
   "e0_mean": {
    "A": 0.069,
    "E": 0.268,
    "G": 0.288,
    "Q": 0.037
   },
   "g_minus_a_pooled": 0.219,
   "g_minus_a": [
    0.224,
    0.226,
    0.206
   ],
   "g_minus_q": [
    0.256,
    0.255,
    0.243
   ],
   "teacher_lead": 0.428,
   "by_distance_mean": {
    "bins_slices": [
     [
      0,
      40
     ],
     [
      40,
      80
     ],
     [
      80,
      120
     ],
     [
      120,
      160
     ],
     [
      160,
      200
     ],
     [
      200,
      240
     ]
    ],
    "A": [
     0.061,
     0.063,
     0.064,
     0.08,
     0.079,
     0.077
    ],
    "G": [
     0.327,
     0.321,
     0.311,
     0.246,
     0.246,
     0.237
    ]
   },
   "attempts": "Eighth freeze of this test; the first seven stopped on technical faults before any score was read; read once",
   "limits": "A second pass on a band picked after a failed test on another scroll, with the void rule added after that failure (under it the failed test would read void), so not a planned replication. Ringleader v1 had run on other PHerc0172 bands, never on this band's windings. m7's training data is not published; whether villa's fit settings were tuned on PHerc0172 is unknown; arm A is villa's fit at the PHerc0139 test's settings, not its best configuration. Stack C and ten slices of stack B lose one sector to a box held out for another test. The taught fit stays below Ringleader's own labels (0.29 against 0.50). The harness and data staging are not released.",
   "source": "docs/tools/ringleader.md#checks-and-controls",
   "date": "2026-11-06"
  }
 ],
 "exposure_atlas_checks": [
  {
   "key": "expatlas_instrument_v1_attempt2_20261104",
   "tool": "kit/expatlas.py (python -m kit expatlas)",
   "catalog": "metadata.json Last-Modified 2 Nov 2026 11:32:29 GMT, ETag 274f64b9261982444dfd2881f48ae766",
   "about": "Planted-duplicate instrument checks of the Exposure Atlas, rule version 1 (docs/prereg/2026-11-04-expatlas-instrument.md, frozen at 0ec8999; attempt 1 killed for memory before scoring), run at 34c3683 in Hugging Face Job 6aeaa27a0e0338d1bb7f85ec on 30 public PHerc0814 and PHerc0009B meshes",
   "label": "Model output",
   "verdict": "FAIL",
   "rule": "PASS when (a) at least 29 of 30 planted copies are called exposed, (b) at least 29 of 30 copies moved a further 100 um along the normal are not, (c) all 30 cross-scroll pairs are not, and (d) both directions of the real positive pair are",
   "a_exposed": 30,
   "b_not_exposed": 21,
   "c_not_exposed": 30,
   "d_exposed": 2,
   "of": 30,
   "a_f_same_min": 0.804,
   "a_f_same_median": 0.927,
   "b_f_same_max": 0.25,
   "c_f_same_max": 0.009,
   "note": "all ten (b) failures (9 exposed, 1 trace) are large auto-grown PHerc0814 meshes on the 9.362 um scan; a post-hoc self-cover diagnostic (Interpretation) found 7 to 45 % of their vertices with another part of the same mesh 60 to 160 um away, so the 100 um copy lands on the mesh's own next winding; the matrix was not computed",
   "source": "docs/data/expatlas/plant-v1-attempt2-2026-11-04.json"
  },
  {
   "key": "expatlas_instrument_v2_20261104",
   "tool": "kit/expatlas.py (python -m kit expatlas)",
   "catalog": "metadata.json Last-Modified 2 Nov 2026 11:32:29 GMT, ETag 274f64b9261982444dfd2881f48ae766",
   "about": "Planted-duplicate instrument checks, rule version 2 (docs/prereg/2026-11-04-expatlas-instrument-v2.md, frozen at 7741f8b; control (b) scored only on vertices with no other part of their own mesh within 160 um; new seed, 28 of 30 meshes shared with version 1's draw), Hugging Face Job 6aeaa4410e0338d1bb7f866e",
   "label": "Model output",
   "verdict": "PASS",
   "rule": "as version 1, with (b) passing when at least 95 % of at least 25 scored meshes are not called exposed",
   "a_exposed": 30,
   "b_not_exposed": 30,
   "b_scored": 30,
   "c_not_exposed": 30,
   "d_exposed": 2,
   "of": 30,
   "a_f_same_min": 0.804,
   "a_f_same_median": 0.927,
   "b_f_same_max": 0.037,
   "b_trace": 9,
   "c_f_same_max": 0.01,
   "d_f_same": [
    0.0505,
    0.9012
   ],
   "note": "the nine (b) traces (0.005 to 0.037, free vertices only) are the multi-winding PHerc0814 meshes; on all their vertices the same copies give 0.039 to 0.259, and matrix cells are scored on all vertices; free vertices were 54 to 100 % of each mesh. Not independent of version 1: the free-vertex rule came from a diagnostic of version 1's failed meshes, and 28 of the 30 source meshes are version 1's. (d) passes at 0.0505 against the 0.05 bar in one direction, a near miss. (a) goes forward and back through the same catalog matrix, so it tests the instrument, not the transforms' accuracy",
   "source": "docs/data/expatlas/plant-v2-2026-11-04.json",
   "b_all_vertex_f_same_range_multiwinding": [
    0.039,
    0.259
   ]
  },
  {
   "key": "expatlas_matrix_20261104",
   "tool": "kit/expatlas.py (python -m kit expatlas)",
   "catalog": "metadata.json Last-Modified 2 Nov 2026 11:32:29 GMT, ETag 274f64b9261982444dfd2881f48ae766",
   "about": "Exposure Atlas matrix after the version 2 PASS: 20 labeled segments of PHerc0009B, PHerc0500P2, PHerc0814, PHerc0841, PHerc1667 and PHercParis4 against 10 released ink models (kit/data/expatlas-models.json), rerun locally at 913e932 after review fixes (first run Hugging Face Job 6aeaa4cfc78863f9c6322578 at 7741f8b: same shares; ink_canonical_2um's 8 PHercParis4 cells read clear there, unknown now); no labels or CT read",
   "label": "Model output (the calls); Sourced fact or Community report (the lists, per model)",
   "cells": {
    "trained": 92,
    "exposed": 18,
    "trace": 3,
    "selected": 6,
    "clear": 32,
    "unknown": 49
   },
   "exposed_beyond_names": {
    "d9v2 on PHerc0009B ag487 (its own test list)": 0.43,
    "d9v2 on PHerc0500P2 -1 (its own test list)": 0.263,
    "Reader v2 and Reader V3 on PHerc0009B ag487": 0.43,
    "Reader v2 and Reader V3 on PHerc0500P2 -1": 0.086,
    "v8in, v8in-1447, A-hn on PHercParis4 20230929220926": 0.146,
    "v8in, v8in-1447, A-hn on PHercParis4 20231007101619": 0.101,
    "v8in, v8in-1447, A-hn on PHerc1667 w023": 0.056,
    "v8in, v8in-1447, A-hn on PHerc0500P2 -1": 0.056
   },
   "evaluation_pairs": {
    "pairs": 92,
    "exposed": 18,
    "trace": 9,
    "examples": "PHerc0841 w00 and ag896 0.35; PHercParis4 pairs 0.08 to 0.25; PHerc1667 w028 and w029 0.20 (named as neighboring windings: a sheet switch or windings closer than 60 um, the geometry cannot tell)"
   },
   "source": "docs/data/expatlas/matrix-2026-11-04.json",
   "cross_scan_cells": 0
  }
 ],
 "bucket_sentinel_checks": [
  {
   "key": "sentinel_controls_20261104",
   "date": "2026-11-04",
   "label": "Model output",
   "verdict": "PASS",
   "about": "Bucket Sentinel (kit sentinel, commit 9fcc215) planted-defect controls on 20 public S3 segment surface volumes of PHerc0343P, PHerc0814 and PHerc0009B, each rendered from its own mesh (docs/sentinel/dev-pool.json); attempt 2, after attempt 1 failed on three implementation faults",
   "rule": "PASS when recall over must-flag plants is at least 0.95, the offset error is at most 2 px in at least 95 percent of caught shift and resample plants, and at most 5 percent of negatives and bases flag",
   "plants": 252,
   "recall": 0.9881,
   "shift_error_median_px": 0.341,
   "shift_within_2px": 1.0,
   "negatives_and_bases": 56,
   "false_flag_rate": 0.0,
   "bases_aligned": 18,
   "bases_not_measured": 2,
   "source": "docs/prereg/2026-11-04-bucket-sentinel.md#4-dev-runs-filled-at-the-freeze"
  },
  {
   "key": "sentinel_backtest_20261104",
   "date": "2026-11-04",
   "label": "Model output",
   "verdict": "PASS",
   "about": "Bucket Sentinel (commit a7ecd62) frozen blind back-test on the 14 Hugging Face ink/ folders of scrollprize/datasets that the reserved-data rule leaves open, against the defects villa #2370, #2107 and #2369 report (Community report); the reporters' tables were not read; run once",
   "rule": "B1 at least 8 of 9 misaligned canvases flagged; B2 none of 4 aligned; B3 at least 4 of 5 truncated pyramids and none of 9 others; B4 MAN5_outer_3 labels flagged 59 px short",
   "b1_flagged": "9/9",
   "b2_false_flags": "0/4",
   "b3_flagged": "5/5",
   "b3_false_flags": "0/9",
   "b4": "flagged, short by 59 and 59 px",
   "agreement": "PHerc0009B constant (-19.20, -19.29) px against about (-19.6, -19.6) in #2370; Paris 4 w01_r15 slopes 0.00427 and -0.00388 against 0.004263 and -0.003876",
   "offset_table": "docs/sentinel/offsets.json",
   "limits": "14 of the bucket's ink/ folders; the folders the reserved-data rule keeps closed were not opened; one bucket version",
   "source": "docs/prereg/2026-11-04-bucket-sentinel.md#outcome"
  }
 ],
 "spiral_smoothing_checks": [
  {
   "key": "spiral_smoothing_0172_l3_20261030",
   "date": "2026-10-30",
   "about": "villa's spiral fit (villa commit e2254ebe) with optimizer_flow_grad_smoothing on (sigma 32 voxels, no labels) against villa's defaults, plus a same-seed repeat of the defaults, seeds 1 to 3, on held-out PHerc0172 (volume 20241024131838, 7.91 um), fit window z 9,984 to 10,560, scored in the +x sector (y 2,112 to 4,416, x 5,952 to 7,488) against the organizers' hand windings w052 to w095: seam switches per 1000 sheet neighbor pairs (kit.seamswitch defaults) and exact share of truth pairs",
   "label": "Model output",
   "verdict": "PASS",
   "rule": "fixed before the read, read once: PASS when D = S(smoothing) - S(defaults) is at most -7.5 per 1000 at every seed and the exact-share guard G = mean(E(smoothing) - E(defaults)) is at least -0.02; VOID unless the same-seed repeat moves S by less than 7.5 at every seed",
   "switches_per_1000": {
    "defaults": [
     228.258,
     230.568,
     230.638
    ],
    "smoothing": [
     220.733,
     218.562,
     215.233
    ],
    "repeat": [
     227.979,
     232.327,
     236.271
    ]
   },
   "D_per_seed": [
    -7.525,
    -12.006,
    -15.406
   ],
   "bar_D": -7.5,
   "repeat_shift_per_seed": [
    -0.279,
    1.759,
    5.633
   ],
   "exact_share": {
    "defaults": [
     0.3217,
     0.3429,
     0.3576
    ],
    "smoothing": [
     0.3466,
     0.3484,
     0.35
    ]
   },
   "G": 0.0076,
   "G_ci95_cluster_bootstrap": [
    -0.02,
    0.0339
   ],
   "guard": -0.02,
   "truth_pairs": 3826,
   "neighbor_pairs_per_fit": 143259,
   "windings_scored": 32,
   "independent_rerun": "MATCH",
   "limits": "one scroll, one band, one sector, three seeds; seed 1 clears the bar by 0.025 per 1000 (about 3 switches of 143,259 neighbor pairs); the same-seed repeat moved up to 5.633, 75 % of the bar; the switch measure was chosen as primary after the exact-share primary failed on another scroll in development",
   "disclosure": "found after the read (30 Oct 2026, development): the center line this test used for its truth lines and both fits, the PHerc0172 line of release umbilicus-lines-v1, sits 3.9 to 12.6 mm off the hand windings' center in this band (two methods). Both arms shared it, so the comparison is fair and the PASS stands, but it shows the smoothing helping a straining fit with a misplaced center line (45 to 58 winding jumps per 1000 at the defaults), not a general switch cut (Interpretation)",
   "source": "docs/FINDINGS.md",
   "fit_window_z": [
    9984,
    10560
   ]
  },
  {
   "key": "spiral_smoothing_0332_l4_20261030",
   "date": "2026-10-30",
   "about": "the same comparison on a second scroll, PHerc0332 (scan 20251211183505, 2.399 um, fitted at level 2, 9.596 um), fit window level 2 z 3,200 to 3,712 on the organizers' open-data level 2 lasagna field and their hand center line, scored in level 2 x 678 to 3,254, y 1,127 to 3,075 against the organizers' meshes 20240711124827-20240618142020 and 20240828190516-20240716140050",
   "label": "Model output",
   "verdict": "FAIL (the repeat control held; the smoothing added switches at every seed; the guard held)",
   "rule": "the level 3 rule above, fixed before the read, read once",
   "switches_per_1000": {
    "defaults": [
     124.711,
     119.003,
     124.693
    ],
    "smoothing": [
     125.018,
     133.026,
     129.32
    ],
    "repeat": [
     122.944,
     117.028,
     131.574
    ]
   },
   "D_per_seed": [
    0.307,
    14.023,
    4.626
   ],
   "repeat_shift_per_seed": [
    -1.768,
    -1.975,
    6.881
   ],
   "G": -0.0013,
   "G_ci95_cluster_bootstrap": [
    -0.0134,
    0.0112
   ],
   "truth_pairs": 3676,
   "neighbor_pairs_per_fit": 110888,
   "independent_rerun": "MATCH",
   "limits": "one band on one scroll; truth from two organizer meshes; the center line checks out here and the default fits had about 1.5 winding jumps per 1000",
   "source": "docs/FINDINGS.md"
  }
 ],
 "villa_sheet_tracer_checks": [
  {
   "key": "villa_sheet_tracer_20261104",
   "date": "2026-11-04",
   "about": "villa's released sheet tracer scrollprize/sheet_tracer_7pt_gap60 (rev 443497f0, last.pt): its held-out bands and fine-tuning on a small patch set (villa main baf5e60; prepare_scroll.py read at 4eecd08)",
   "label": "Sourced fact (villa source and checkpoint config), reported upstream",
   "heldout_rule": "middle 20 % of each scroll's z extent of the segments listed in metadata.json on the day prepare_scroll.py ran, with a 256 voxel gap; the bands themselves are not published and villa #2440 deleted prepare_scroll.py",
   "heldout_fraction": 0.2,
   "gap_voxels": 256,
   "pherc1667_level2_band_from_todays_metadata": [
    3987,
    5790
   ],
   "small_set_repro": "one PHerc0841 w00 patch (z 13,440 to 14,336, 3,293 nodes) with the checkpoint's split_block 1024 and val_fraction 0.1 stops before the first step: RuntimeError: Could not draw 64 val episodes; with villa PR 2445 and train.val_episodes=0 it trained 3 steps on CPU and wrote last.pt",
   "issue": "https://github.com/ScrollPrize/villa/issues/2444",
   "pull_request": "https://github.com/ScrollPrize/villa/pull/2445",
   "source": "docs/FINDINGS.md",
   "patch_nodes": 3293,
   "patch_z": [
    13440,
    14336
   ]
  }
 ]
}