This file is the one handoff for this repository. Every agent (Claude, Codex, Grok, ChatGPT or any other) reads it first and updates it before stopping: change the open-items table when you start or finish something, link longer notes from their row, and merge the handoff change into main at once rather than leaving it on a work branch. Private work has its own single handoff in Scrolls-private's docs/HANDOFF.md; nothing private goes here. Read AGENTS.md, the workflow and the current Git/PR state before continuing. Use plain sentences and preserve historical evidence bytes and hashes. Older dated sections are in HANDOFF-archive.md, newest first; when this file grows long, move its oldest dated sections there, unchanged (tests/test_repo.py fails past 40 KB).
| Item | State | Next step |
|---|---|---|
| Lasagna Fields data access and preview display (9 Nov) | PR353 released visible data links, selection-bound recipes and 68 stronger display previews. Follow-up fixes array Open actions that incorrectly navigated to raw Zarr roots: Open now targets array metadata; Copy URL and recipes keep the exact store addresses. Scientific arrays and preview assets are unchanged (page). | Verify metadata endpoints and actual Open/Copy controls, merge the scoped fix, then redeploy only the guide and verify live. |
| Ringleader ladders: second passing test, PHerc0172, made public (6 Nov) | Chase (6 Nov): skip the villa PR and publish the result here. The Ringleader page, novel finding 29 (README and FINDINGS.md), ringleader_ladders_0139_20261104 and ringleader_ladders_0172_20261106 in results.json, the release entry and a scores-only chart (chart_0172.py) add PHerc0172 z 9,680 to 9,904: fit alone 0.069, with ladders 0.288, plus-one control 0.037. No code change |
Merge once Chase says yes to the exact text |
| Token thrift, AGENTS.md rule 18 (6 Nov) | Done: rule 18, a shorter AGENTS.md, the Sonnet scout subagent; details in AI-AGENTS.md |
None |
| Umbilici 1.14.0: only each scan's best center line glows (4 Nov) | Done. On Chase's ask and his yes to the preview: the line kit centerline get and AutoPipeline take by default glows, the others are gray and labeled alternate or unused on the page, in the Data tables, downloads and API. Merged in PR #336; deployed as Space commit 56f7da24 from main at a204505; live recheck passed (page). |
None. If Chase wants Center Finder v2 to glow first, change rankOf in spaces/umbilici/app.js and its test. |
| Paris 4 Ink Map Space listed in the Connect checks (4 Nov) | Chase's yes (4 Nov): spaces/connect/spaces.json gains SharpMeow/paris4-ink-map (docker, page code in Scrolls-private), and spaces/connect/paris4-ink-map/ mirrors the live Space's connect.json at Space commit f521c528; the rendered connect.html, CONNECT.md and onescroll.html are byte-identical to the live ones. kit spaceconnect check: 15 Spaces, 0 problems. check --live: every link answers except a GET on /mcp (405, POST only, as on Vesuvius Radar); a POST tools/list there returns its seven tools. No Space was changed |
None after merge. Run kit spaceconnect scroll paris4-ink-map after any change to the Space |
| Findings 12 and 13 brought up to villa's state (4 Nov) | villa #2060 closed unmerged on 30 Oct (rendering moving to vc_reader); seed_anchor_xyz merged in villa #2063 on 30 Oct, closing #2059. README, FINDINGS.md, results.json, vc3dfix, false404-relay and MAP updated; the 27 Oct section moved to the archive |
Done |
| Novel findings 25 to 28 (4 Nov) | Merged in PR #331 with Chase's yes (4 Nov): README "Novel findings" and FINDINGS.md add sqrecipe's PHerc0332 PASS (25), villa's flow-gradient smoothing on PHerc0172 (PASS with a misplaced center line) and PHerc0332 (FAIL) (26) and villa's sheet tracer held-out bands and small-set fine-tuning (villa #2444, #2445) (28); numbers in results.json (spiral_smoothing_checks, villa_sheet_tracer_checks) and the claims ledger |
Done. Finding 27 was removed on 4 Nov at Chase's choice, until its review is done; its number stays unused |
afv-fibers: box .afv download fixed on Linux (4 Nov bug sweep) |
write_box_afv opened its output without URI support, so attaching each tile as file:...?mode=ro failed where Python's sqlite3 has URIs off (page). Merged in PR #328 as 11ea65e. Redeployed SharpMeow/afv-fibers from that commit under the standing yes: only fibers.py differed from main, uploaded as Space commit 02ff5c84 (rollback bbfa9088), RUNNING. Live recheck: a PHerc0139 box (z 10000, y 3200, x 3200, half 64) came back as an .afv with 144 fibers |
None |
| labelsnap: the shared band rule behind the finer-scan scores disclosed (4 Nov) | Chase approved the exact sentence on 4 Nov for public Scrolls, the Snapped Labels Space and the dataset card: "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)." Added beside each place that cites the 29 Oct seven-community-snaps numbers: README finding 22, FINDINGS 22, results.json (note), the labelsnap and snapped-labels release entries (through kit releases add; the summary limit rose from 1,500 to 1,700 characters so the dataset entry fits unchanged otherwise), the pipeline note, the labelsnap and snapped-labels pages, and the Space's README and page. No other wording or numbers changed; the charts are unchanged |
Merged in PR #329 as 9657b18. Redeployed the Snapped Labels Space from that commit as Space commit 38409bc3 (rollback d5212486), RUNNING; the dataset card SharpMeow/vesuvius-snapped-labels carries the sentence after the community-snaps paragraph and in a 4 Nov Changes entry, dataset commit 571bf1f0 (rollback 2a302471). Live recheck: the served page holds the sentence twice, the Space README and the card hold it, no long dashes. Nothing left |
| villaports: which villa #2440 removals break your scripts (4 Nov) | New kit command kit villaports (page), released as a community helper on Chase's OK (4 Nov): scan resolves every reference to villa's vesuvius and vc_reader packages in a folder against villa's tree at a commit (default main) and suggests a replacement from a map of 31 entries read from villa #2440 at baf5e605b; diff compares two villa commits. 31 tests, no network. Its run on this repository at villa main found 2 breaks: kit/spaceconnect.py line 148 (the Connect your tool snippet from vesuvius import Volume, gone in #2440) and tests/test_inklens.py line 268 (normalize_robust from vesuvius.image_proc) |
Port the two breaks: the Connect your tool snippet to vc_reader.reader.open_zarr, then rerender the Spaces' panels |
| Center Finder: trim line ends that leave the scroll; center lines v2.9 (4 Nov) | Done. scripts/umbilicus/lineends.py (check, trim; tests tests.test_umbilicus_lineends) cuts a line's end runs outside the scroll in the masked CT; Umbilici's build_lines.py loads the same test. Merged in PR #325. On Chase's yes (4 Nov) center lines v2.9 went up to SharpMeow/vesuvius-center-lines (dataset commit 705c2996, 17 files on top of v2.8 fdfb6103): trimmed copies of the 12 Center Finder v2 lines it cut (155 points, 167.3 mm of height), originals kept with superseded_by; 103 lines, 45 superseded; anonymous download matches SHA256SUMS. Release entries updated (page) |
None |
| Umbilici 1.13.0: line ends outside the scroll hidden (4 Nov) | Done. Merged in PR #323; deployed to the Space as commit 68de75ed from main at 48e9346 with data/lines.json built from center lines v2.9 and tools/lineends.py. Live recheck: RUNNING, served app.js and data/lines.json match the build, version 1.13.0, reads dataset revision 705c2996 (35 scans, 103 lines, 45 superseded), the 12 trimmed lines have no hidden end points, PHercMANB's polished line untested (corner check), no page errors (only 404s for empty CT chunks) (page) |
None |
Bucket Sentinel: kit sentinel, standing checks on the organizers' public uploads with an offset table (4 Nov) |
Merged in #322 (page, preregistration). Checks canvases against their tifxyz by CT phase correlation (affine or per-region model), pyramids, [1, 1, 1] scales (fallback from metadata.min.json), label grids and mesh frames; preflight before vc_render_tifxyz; kit/sentinel_table.py is the loader side. Planted-defect controls PASS on attempt 2 (attempt 1 failed on three code faults, fixed); frozen blind back-test PASS on the 14 ink/ folders the reserved-data rule leaves open (9 of 9 #2370 misaligned flagged, 0 of 4 aligned, 5 of 5 #2107 pyramids, MAN5 labels 59 px short); offsets in docs/sentinel/offsets.json (Model output). Stage option inputs bucket-sentinel (experimental); daily routine bucket-sentinel (scan only, and only where work/sentinel/exclude.json exists) |
Chase decides on the villa #2370 comment draft (contrib/villa-2370-offset-table.md); agents copy the private exclusion file to work/sentinel/exclude.json before scanning |
Exposure Atlas: kit expatlas, a leakage matrix from meshes (4 Nov, chosen by Chase) |
New kit command (page): ten released ink models' published training lists resolved to catalog segments (kit/data/expatlas-models.json), then, in each labeled segment's own scan, the share of its mesh within 60 um of the model's training meshes; builds on dlescanogithub's proposal in villa #2346. Instrument checks preregistered: version 1 FAILED its adjacent-winding control (21 of 30, prereg; multi-winding PHerc0814 meshes), version 2 PASSED all four (prereg). First matrix: 20 labeled segments by 10 models, records in data/expatlas/, exposure_atlas_checks in results.json; Model output. New stage exposure in the stage registry. 45 tests. An independent review (4 Nov) found open ends reading clear; fixed in 913e932 and the matrix rerun: 92 trained, 18 exposed, 3 trace, 6 selected, 32 clear, 49 unknown Villa comment draft (not posted): contrib/villa-2346-exposure-check-comment.md. Merged in #321 |
Chase: read the villa draft and post it or not. Then: decide whether held_out in results.json and the leaderboard should cite the matrix (d9v2 and Reader v2 on PHerc0009B ag487, d9v2 on PHerc0500P2 -1); add models as their lists appear |
| Item | State | Next step |
|---|---|---|
Training-run comparison for label suggestions: kit altrain and a fragment benchmark (3 Nov) |
New kit command (page): lays out a simulated labeling round with oracle labels and judges whether fine-tuning on kit labelsuggest's tiles beats random tiles, with a paired bootstrap and reverse-depth, shuffle and displaced-label controls. Benchmark on Frag2, Frag3 and Frag4 (docs/results.json, label_suggest_checks): fine-tuning ink_9um on the suggested tiles did not beat random tiles at any budget; at 5 % it was below random on two fragments; readout VOID by its controls (only Frag4 counted). |
None for this tool; a villa follow-up on PR #2381 is not proposed on this evidence |
| Prize snapshot of 3 Nov: First Letters at 8 more scrolls, $400,000 | New snapshot kit/data/prizes-2026-11-03.json (the 11 Oct file is kept) from the prizes page body read 3 Nov and villa #2431: 20 First Letters volumes, PHerc0306B awarded (announced 2 Nov). README table, prizes.md and the tool pages follow. The homepage card and the $2,140,000 pool figure still show the old $500,000; the Progress Prize line still names 31 Oct, so kit.prizewatch reports that one deadline until the page names November's |
When the page names the November Progress Prize deadline, add a new dated snapshot |
| labelsnap against scroll-label-snap: real drift chart (3 Nov) | Chart of the preregistered real drift test on 43 fresh public PHerc0139 sites (Model output): within 1 voxel of the two-scan recto face, labelsnap 0.798, A.Q. Snyder's scroll-label-snap 0.179 to 0.434 over four settings, published labels 0.173, with the caveats on the labelsnap page. PNGs in docs/assets/labelsnap/ (light and dark, sidecar labelsnap-realdrift.json), script scripts/experiments/lsnap-realdrift/chart.py, numbers in label_snap_checks of docs/results.json. Merged with this PR after CI and the leak check. 4 Nov: a caveat that labelsnap and the reference can both land on the inner neighbor's face where sheets touch, with the check that keeps the result, added to the page and to limits in results.json |
None; the private record stays in Scrolls-private |
| PHerc0139 snap development sites, public copy (3 Nov) | Branch claude/pherc0139-snap-sites: data/pherc0139-snap-sites/, 108 spent PHerc0139 development sites (972 vertices, windings w023 to w029, w031 to w034, w036 to w038) with two-scan recto face offsets and the placements of kit labelsnap and A.Q. Snyder's scroll-label-snap, CC BY-NC 4.0, the same three files given to A.Q. Snyder. Hit rates on 601 covered vertices (Model output, development only): published 0.298, labelsnap 0.854, scroll-label-snap 0.431 to 0.674. Linked from the docs index and the labelsnap page; label_snap_checks entry in results.json The leak check flagged 7 data cells (each counted in the CSV and the JSON) whose values coincide with unrelated numbers in private notes: site 38 k 4 face_2399_offset_vx, site 48 k 8 labelsnap_x, site 51 k 6 lsnap0e_x, site 61 k 6 lsnap0e_x, site 85 k 4 face_roi2_offset_vx, site 99 k 5 lsnap0_x, site 117 k 7 m7_x; Chase confirmed them as false alarms on 3 Nov, and the data stays byte-identical |
Done 3 Nov: Chase gave the file to A.Q. Snyder first, then this PR was opened and merged, as he decided |
| Item | State | Next step |
|---|---|---|
| Handoff headroom (31 Oct, workflow audit) | The 23 and 24 Oct sections moved unchanged to HANDOFF-archive.md; this file went from 39.9 to about 30 KB of its 40 KB limit. Possibly open items from them are listed under "Still binding" | Move the oldest dated section to the archive whenever this file passes about 35 KB |
| Ringleader, sqrecipe and offsetdiag released (31 Oct) | Code, tests, tool pages and results for Ringleader (two frozen PASSes, README finding 24), sqrecipe (PHerc0332 PASS) and offsetdiag (development only), with Chase's yes on 31 Oct; the Scrolls kit grouped entry limit is now 96 tools. Merged with this PR after CI, the leak check and an independent review | None. Ringleader v2 and any later Ringleader tests stay out of this repository until they are released the same way |
windlabel-box: --split-rule none and per-run gm_x (31 Oct) |
The split rule keeps a none option (params digest 6e3aff82495f): every chain step stays whole and flags.json marks the kept steps the v2 cut would remove as windlabel-check-split-step places to check, so the cut can be judged per scroll before switching it on (page); every run's export.json and --json answer also reports gm_x (n, median, p90, share above the v2 cut of 1.0). v1 c6d2eecbebc2 and v2 cd273aed7673 stay byte for byte; the same change ships with villa PR 2367, from jo-tud's cut replay on PHerc Paris 4. PR #311, 6 tests |
Done: merged as d965d41 with villa's head at fe2a324 and the jo-tud reply posted; take jo-tud's Paris 4 numbers and code if they arrive |
| Item | State | Next step |
|---|---|---|
| Lasagna Fields: twelve more Lasagnetta datasets and a decluttered main page (30 Oct) | Twelve public Lasagnetta 0.1.1 datasets (Chase's yes on the exact cards, 30 Oct): nine PHerc0139 volumes, PHerc0172 20241024131839, PHerc1667 and PHercMANBp at 1.129 um; each card states its release self-check. Mirrored into the public bucket SharpMeow/lasagna-fields. PR #300 lists them and redesigns Every field (Chase's yes, 30 Oct): four sections, folded field lists, a third of the height, and the off-screen Preview box fixed (page) | Done 30 Oct: #300 merged (d8ed50e), Space redeployed with the three changed files, each checked against the live copy first (9cbb7ec6). Logged-out recheck: files byte-identical to main, all twelve listed, all 22 Lasagnetta manifests load, no page errors or sideways scroll at 1400 and 390 px |
| AutoPipeline 0.1.11: self-intersections and flattening on VC3D builds after villa #2358 (30 Oct) | Released with Chase's yes (30 Oct): branch submission/progress-2026-10-fix11 at 22cecae, 184 files, after an agent leak and quality review; tag autopipeline-v0.1.11. The qa check runs villa's vc_tifxyz_selfcross or the same census kernel compiled from villa's source at build time (scrolls-kit selfcross build); flattening runs vc_flatten or flatboi. This pull request updates the README entry and the sentence on current builds in pipeline-october-2026.md and tools/autopipeline.md. |
GitHub release not marked Latest, from Chase's Mac; then check the tag, the release and the README entry live. docs/autopipeline/README.md still gives the 7154682e1 workaround: change it only with Chase's yes on the text. |
Label suggestions from reader disagreement: kit labelsuggest map (30 Oct) |
New kit command (page): ranks a segment's tiles by how much two or more readers' ink maps disagree and writes the top tiles as a mask to label first, with matched random and agree tiles and a pixel-shuffle control; numpy only, 13 tests. Two preregistered tests on the organizers' own label rounds (label_suggest_checks in results.json): lift 3.82 on PHerc1667 and 1.84 on PHerc0841 ag896 (bar 1.5); not shown to beat simple picks. Line labelsuggest in the Scrolls kit release entry |
None; a timed test with people (does labeling at suggested tiles save time) is untested |
Box copy of a zarr volume: kit zarrbox (30 Oct) |
New kit command (page), asked by Chase on 30 Oct: plan a box of a zarr or OME-Zarr volume (level-0 voxels, any level), copy the metadata and only the box's chunk files byte for byte with a sha256 per chunk in zarrbox.json, verify the copy against it and, with --source, against the bucket. Reuses kit tifxyz chunks' reader. Checked against villa's vesuvius 0.2.4 first: its extract_sparse_ome_zarr.py copies chunks near tifxyz surfaces, re-encoded, with no box mode and no checksums, so this is not a duplicate. Live check: a 256-voxel PHerc0139 box read equal to the bucket under zarr-python 3.4.2; the chunk past the box read as fill (control). 10 tests. Not in the README's kit tools list: the Scrolls kit entry is at its 64-tool cap. |
Merge the PR once checks pass; add a kit tools line when the list has room |
| Item | State | Next step |
|---|---|---|
| Parting result page for the October Progress Prize entry (29 Oct) | Done 30 Oct, merged to main with Chase's yes on the text: log of the preregistered contact test on PHercParis1Fr34 (PASS, gain +0.056 over base dinovol, bound +0.037; earlier PHerc1667 fragment FAIL at +0.045 and later PHercParis2Fr143 FAIL at +0.010 disclosed; fused contacts not solved), finding 23 in the README and FINDINGS, results.json parting_checks, glossary entry Parting. 30 Oct, by Chase's decision, the model file of every test is public as SharpMeow/parting-v1 (CC BY-NC 4.0, sha256 checked on download), and the page, scripts README, log and release entry say so. A villa draft PR of Parting waits on Chase's yes on its text |
|
| Swarm bursts in the agent notes (29 Oct) | Agents may run many subagents at once, each on a different route, to solve problems; preregistration, held-out rules and rule 8 still bind. Added to AGENTS.md standing decisions, AI-AGENTS.md and DECISIONS.md | Done |
| Windings made public: seamswitch, the frozen test chart, finding 21 (29 Oct) | New kit entry point python -m kit.seamswitch (page): seam-aware winding switches of a spiral fit on sheet grids, numpy only, with a demo and 19 synthetic tests; a line seamswitch in the Scrolls kit release entry. windlabel-box gets a chart of its six frozen tests and the 29 Oct comparison with E1 and winding-phase (docs/media/windlabel-box-heldout.svg, by scripts/media/windings_chart.py), three new windlabel_box_checks entries in results.json, README finding 21 and its FINDINGS text, all from numbers already on its page. Stage registry: new stages windings and spiralcheck with ours and the community's options. |
After merge: none. A public-data run of seamswitch on a real fit (a fit of a non-eligible scroll, scored with a baseline arm) would give it a first real number. |
Label snap tools on main: kit labelsnap, kit finesnap, kit snap (29 Oct) |
Pull request from claude/project-thread-czt16x: the code that made the Snapped Labels dataset, ported from the AutoPipeline 0.1.10 package with tests (kit/labelsnap.py, kit/finesnap.py, kit/snap.py, kit/snaptruth.py, kit/snapresolve.py and their support modules), pages labelsnap, finesnap and snap, label_snap_checks in results.json, finding 22, two charts (docs/media/label-snap-*.svg) and a synthetic GIF (docs/assets/labelsnap/), and two new tool lines in the Scrolls kit release entry. |
Approved for release on 29 Oct; open as a pull request, to be merged. Module pages for villafmt, vc3d-prep, vc3d-region, vc3d-orient and vc3d-landmarks are pending (pending.json). |
| Vesuvius Radar: a public radar for new scroll work (29 Oct) | New Space SharpMeow/vesuvius-radar (free CPU Docker Space, Chase's yes of 29 Oct): reads arXiv, GitHub (with villa PRs, issues and releases), Hugging Face, PyPI, Zenodo, Crossref and the Challenge blog logged out on a schedule, keeps 30 days, scores by keywords and by meaning (bge-small on CPU), with a 1 minute to 30 day lookback, reader views, a briefing, JSON, CSV, RSS and an MCP server. Source spaces/vesuvius-radar/, page tools/vesuvius-radar.md, tests tests/test_vesuvius_radar.py, README entry vesuvius-radar. First collection about 4 minutes, all 7 sources ok. Cleanup of 29 Oct (stacked PR on the branch claude/project-thread-3s1gtd): the list opens on Newest, 10 per page with a per-page picker, filters in four dropdowns with removable pills, the builders view is called Contributors, the people chart ranks by distinct pieces of work, two new hero stats, and three easter eggs. A blind rating of all 588 items found the old final score agrees with the ratings at Spearman 0.70; a learned head on the same embeddings (head.json, refit with train_head.py) now sets the final score: 0.86 in nested cross-validation, high band 61 items at precision 0.60 instead of 124 at 0.39. The ratings file is not in this repository. 30 Oct: version 1.1.1 live (PRs #285 and #287, redeployed 02:51 UTC): villa issue and pull request cards name who made the latest comment, review or change and what it said, with the opener as "opened by" (first live check: 32 of 53 villa items in a day show their update, villa #193 credits the newer commenter), the two card buttons open separate panels, and a cut-off GitHub download is retried instead of failing the whole source. 30 Oct evening: version 1.1.2 (PR #303): the minimum score no longer snaps back to 20 (Hugging Face's ?__theme= parameter had reset the saved filters), ChaseHendrick, SharpMeow and Sharpie count as one person (people in radar.json), the meaning switch reads "Match ideas, not words", and Refresh has a halo. 1.1.2 redeployed 23:02 UTC. Version 1.2.0 (PR #305) adds news searches (Google News, Bing News, Hacker News, scored like any listing, one story kept once) and GitHub org:ScrollPrize; 1.2.0 redeployed 23:16 UTC. Version 1.2.1 (PR #306) fixes review findings on the news: Hacker News discussions are a link (cards read "NaN comments"), the learned score gives news no blog weight, the team view drops press stories, the best copy of a repeated story is kept, and non-Latin titles keep their own ids. Checked and left for later: the open-data catalog, DataCite, ash2txt community uploads and Bluesky (each needs its own collector and a source color). |
Redeploy from merged main after changes (standing yes); watch GitHub's logged-out rate limit on the Sources tab, since a GitHub check now makes 3 to 9 more calls |
| Item | State | Next step |
|---|---|---|
| Coupled pipeline page, October entry (30 Oct) | pipeline-october-2026.md, PR #262: four steps (center line, trace with the contact barrier, label moved onto the papyrus, ink); windlabel-box, labelsuggest and Parting beside it; fibers (Weft) and joins (kollesis) listed under What did not work, their synthetic media removed; contact-barrier page says repair is opt in | Merge only after Chase approves the exact page |
Mac scripts on villa main (28 Oct) |
villa PRs #1812 and #1865 merged into villa main on 2026-10-08, so scripts/mac-verify.sh and scripts/mac-w045.sh no longer fetch pull/1865/head: both check out villa main at a pinned commit, VILLA_REV (default 860f52cbf, the merge of #1865; VILLA_REV=origin/main for the newest). villa main picks MPS by itself and has no device flag, so mac-verify's CPU reference now runs villa's entry point with MPS hidden from get_accelerator() (functools.lru_cache(lambda: False), since torch._dynamo reads __wrapped__) and CUDA_VISIBLE_DEVICES empty; every run's log must name its device (Using CPU device or Using MPS device), and EXPECT_GPU takes only mps or cpu. Six new tests in tests/test_mac_scripts.py; pages mac-verify, mac-w045 and mac.md updated. Off-Mac SMOKE=1 EXPECT_GPU=cpu run of mac-w045 in a Linux container on the pin: exit 0, the same six AUCs as the 7 Oct smoke run (it caught that a bare lambda breaks torch._dynamo). PR #270, branch claude/mac-verify-villa-main |
Review and merge; then a Mac run of bash scripts/mac-verify.sh rechecks CPU against MPS on the pinned commit |
| Ink Lens redeployed without the PHerc0139 w035 sample (28 Oct) | Redeployed SharpMeow/ink-lens from merged main 2e3c10c (PR #268) under the standing yes of 18 Oct for approved Spaces (rule 8): Space commit b71193a2 (rollback 978066cc), same ZeroGPU hardware. Only README.md, app.py and samples.json changed; every other Space file was already byte-identical to main, and the three new files are byte-identical too. Live recheck 28 Oct: the Space is RUNNING at b71193a2; /list_samples and every Sample menu list only 0841-w00, 0841-ag896 and 0841-ag405, and the page config names no PHerc0139 segment; each of the three ran on /lens_sample (256 px, one window); the 640 px 0841-w00 calibration gave gate ink, AUC 0.9147 as stored against 0.5954 reversed (Model output, development data), the same as the 24 Oct record |
Done. Nothing waits |
| Unroll 1.7.2: phone title fix (28 Oct) | Done, with Chase's yes of 28 Oct. On phones, with More controls open, the Read box above the panel covered the title. The panel's height now leaves room for that box. Pushed to the Space as commit 13a13964 (rollback: 757768f6); the Space is RUNNING and its five changed files match byte for byte. Playwright on the live Space: at 375 and 414 px the Read box now sits below the title and credit; desktop (1280 by 800) and phone landscape layouts measured the same as before. Tool page and release entry at 1.7.2 |
None |
| Snapped Labels dataset: snap loop (28 Oct) | Done, with Chase's yes of 28 Oct. snap-pherc0139-w033-z7600.gif (Model output, 720 px, 0.62 MB) and its credit sidecar are on the dataset (commit 9924782c; rollback 27c3525e), with a Model output caption on the card and SHA256SUMS updated; read back live, sha256 match. Made by scripts/media/snap_media.py real on PHerc0139 w033 at z 7600 |
None |
| Contact barrier page: held-out switch result on PHerc0139 (28 Oct) | New tool page tools/contact-barrier.md for the contact barrier, the sheet-switch cut released inside AutoPipeline 0.1.9 (2144cdf): synthetic demo stills, slow GIF and MP4 (scripts/media/barrier_demo.py, run on the released code), the held-out chart (scripts/media/barrier_chart.py) and contact_barrier_checks in results.json. Model output: on one PHerc0139 sector, two repeats, frozen rule, AHEAD of villa's vc_grow_seg_from_seed; reported beside it: a sector where villa's tracer was ahead with the older repair, the parent rule's earlier reads, and the PHerc0500P2 second-scroll test VOID at its data gate. Branch claude/switch-barrier-entry. |
Merged as #260 and checked logged out; independent check 28 Oct: every number matches the frozen readout, media rebuild byte-identical, and the scorer's half-winding check added to the page. If a second-scroll test passes under a frozen rule, add it to contact_barrier_checks and the page. The barrier's own unit tests come with the next AutoPipeline package or a port of kit switchfix to main. |
| Snapped Labels: Voxel exact tab live, two-scan result (28 Oct) | First deploy: PR #255 (1dcab57), Space commit dd7161c (rollback 3ba468f), PHerc0139 Not shown (first frozen test). Second deploy (this row, PR #265): the frozen two-scan test of 28 Oct, exported by the private exporter from the collected verdicts. PHerc0139 Measured on the covered share only (Model output): coverage 238 of 370 drawn code 0 sites (lower bound 60.0%, bar 40%), on covered sites 226 of 238 within 1 voxel (lower bound 92.0%, bar 90%), every control held; against the 1.129 µm scan alone Not shown (289 of 370, lower bound 74.3%); the first test kept on the card as earlier and spent. kit vxsummary gains coverage, the covered basis, single_truth, scope, second_truth and earlier; the hero card now says the 91% counts scorable vertices. Merged as 9aa8578; redeployed to SharpMeow/snapped-labels as Space commit 538ab067 (rollback: dd7161cf), RUNNING. Rechecked live in headless Chromium at 1360 px (light) and 390 px (dark): the tab shows, the served vx/summary.json equals the exporter output byte for byte, no horizontal scroll, no page errors. An independent review found one page issue (counts from a ?vx=URL file could carry markup), fixed before merge (page) |
When a newer frozen verdict lands, the private exporter rebuilds vx/summary.json; merge, then redeploy. Snapped Labels goes into no Progress Prize form until Chase says this PASS lifts his hold (asked, not answered) |
These come from sections now in HANDOFF-archive.md and still apply:
- README "Novel finding 1" keeps its w045 correction until v8in's PHerc0841 crop scores replace it (the owner's choice; "Correction 8 October 2026").
- Generic tools and public benchmarks stay public even when an artifact's filename includes
candidate("Evidence and next public work"). - Moved 4 November 2026 with the 27 October section, possibly still open: record generic public-data Jobs with
kit run hf-rowsafter each batch; redeploy Snapped Labels from mergedmainwhenspaces/snapped-labels/changes; rerunkit spaceconnect scrollafter any Space change. - Moved 4 November 2026 with the 26 October section, possibly still open: Fiber Checker's own tests for its untested hints and a VC3D overlay check of
fiber.tif; a faster Lasagna Fields model only after a held-out test against villa's. - Moved 31 October 2026 with the 23 and 24 October sections, possibly still open: Chase's own October form for research integrity (#167); whether Umbilici's lag and black flashes are gone on Chase's Mac; VC3D's relative-windings loader fix riding the combined VC3D PR for villa.
- When you change code, also run the separate numerical experiment regressions (TESTING.md), and never modify a script while it runs ("Environment").
Dated sections from 12 October 2026 and earlier, the 7 to 9 October corrections and the October review notes are in HANDOFF-archive.md, newest first. Search it; do not read it whole.