.gitignore

# ---- PHI: NEVER commit any of these ----
# Raw images / pixel data
*.dcm
*.dicom
*.tar
*.png
*.jpg
*.jpeg
*.tif
*.tiff
*.bmp
*.gif
*.webp
*.npy
*.npz

# Ground truth + annotation exports (contain PHI token text)
ground_truth/gt.csv
ground_truth/*.csv
# Pre-validation candidates and interrupted atomic writes. `build_gt.write_gt()` writes the
# FULL gt.csv table to `ground_truth/gt.csv.part` before validating it, and 10d's record
# writer writes `<image_id>.json.<pid>.<tid>.part` — both hold PHI, and NEITHER is matched by
# the `ground_truth/*.csv` or `ground_truth/review/` rules (`.part` is the final extension,
# and a killed process leaves the file behind). Unsuffixed `*.part` so any future atomic
# writer is covered the day it is written, not the day someone notices.
*.part
# Phase 10b sidecars, spelled out even though ground_truth/*.csv already covers them at
# their default paths: the back-map is the image_id -> SOPInstanceUID/series_uid link and
# the input list holds series UIDs + local paths. Both are PHI-adjacent wherever they live.
**/render_backmap*.csv
**/render_inputs*.csv
**/raw_response*
label-studio*/
label_studio*/
# Phase 10c / D-10c.8: Tesseract seed output. `ground_truth/seed/<image_id>.json` holds the
# read-back token text for each image — that IS the PHI, same class as gt.csv. Scoped to the
# `seed/` directory ONLY, never `ground_truth/*.json`, because the sibling summary must stay
# readable. NOTE (see the warning further down): git has no inline comments — every comment
# here is on its own line.
ground_truth/seed/
# Per-stratum seed dirs from diagnostic runs (e.g. ground_truth/seed_ct_scout/) hold the same
# read-back token text and are the same class of PHI. The trailing slash matches DIRECTORIES
# only, so the PHI-free ground_truth/seed_summary.json file is not caught by this rule.
ground_truth/seed_*/
# `seed_summary.json` is PHI-free (per-image counts only) so Arnav can still open it, but it
# is regenerable output and a per-image count table is not something to publish — ignore it
# as a build artifact, not as PHI (Phase 10c / D-10c.8).
ground_truth/seed_summary.json
# Phase 10d: ground_truth/review/<image_id>.json holds the human-confirmed token text —
# that IS the PHI, same class as gt.csv. The existing `ground_truth/*.csv` rule does NOT
# cover .json, so this rule is required, not redundant. Directory-scoped (trailing slash)
# so the PHI-free ground_truth/seed_summary.json sibling stays readable. review_* covers a
# future self-agreement round (ground_truth/review_r2/).
ground_truth/review/
ground_truth/review_*/

# Phase 10c / D-10c.2b: the tessdata_best language model Arnav downloads locally and passes
# via --tessdata-dir. Not PHI — a ~15MB upstream binary blob that does not belong in git.
# Its identity is already pinned the right way: the seeder records `traineddata_sha256` in
# every seed's provenance, so the model is committed by hash, not by content.
tessdata_best/
*.traineddata

# Bulk data dirs / renders / caches
data/
renders/
crops/
cache/
results/
# scripts/tag_census.py staging dir for downloaded series tars (PHI). NOTE: git has no
# inline comments — a trailing `# ...` becomes part of the pattern and the rule silently
# matches nothing, so every comment here must be on its own line.
census_tars/
# scripts/download_series_tars.py destination for the Step 3 series tars (PHI). Covered by
# *.tar already; the directory rule is here so an unexpected non-.tar file landing in it
# (a partial download, a gcloud temp file) cannot be committed either.
series_tars/

# Synthetic test fixtures: ignore generated images, keep the regenerator (D-0.3)
tests/fixtures/

# Logs (may capture filenames, tags, tokens)
*.log

# Local env / build artifacts
.venv/
__pycache__/
*.pyc
*.egg-info/
.pytest_cache/
.ruff_cache/

# Commit the *hash* of gt.csv, never its contents.
!ground_truth/gt.csv.sha256

# text_presence_*.csv = image_id + has_text boolean, PHI-free by construction. Negated so the
# recall denominator lives in git (CLAUDE.md §8: a stratum with no text-bearing images must be
# visible, or its recall reads a meaningless 100%).
!ground_truth/text_presence_*.csv

# NOTE: commit the *hash* of gt.csv, never its contents.
# manifest.csv contains quasi-identifiers (dates, pseudonyms) — keep it out of
# git unless your team has explicitly classified it as shareable.
manifest.csv

# source_paths.txt = 1000 gs:// paths to specific PHI-bucket series (linkable
# identifiers, Limited-Data-Set grade). Same class as manifest.csv — never commit.
source_paths.txt

# gt_sample_*.csv = a stratified sub-manifest (scripts/build_gt_sample.py). Same
# columns as manifest.csv, so same class — never commit. Commit the printed
# sample_sha256 in the plan/run metadata instead, which is what pins the set.
gt_sample_*.csv
# gt_pilot_*.csv = the same sub-manifest shape under a different name (the 10c/10d
# pilot draw). It carries series_uid, study_date and institution_name, so it is the
# same Limited-Data-Set class as manifest.csv — the gt_sample_*.csv rule above does
# NOT match this filename, which left gt_pilot_v1.csv committable until 2026-08-09.
gt_pilot_*.csv

# seed_summary*.json = PHI-free per-image count tables, ignored for the same reason
# as ground_truth/seed_summary.json above: regenerable build output, and a per-image
# count table is not something to publish. The exact-name rule above missed the
# per-stratum variants (seed_summary_v2.json, _pilot, _ct_axial, _ct_scout).
ground_truth/seed_summary*.json

# ---- NOT FOR THE PUBLIC REPO (gradienthealth/ocr-benchmarking) ----
# Personal working docs: internal planning, decision log, and learning journal.
plan.md
next-steps.md
insights.md
diff-update.md

# Presentation drafts and scratch tables. slides.md and promptslides.txt contain
# GRDN-prefixed ID-shaped strings and date-like strings that have not been cleared
# as non-PHI; census_out/tag_census.md is a DICOM tag census over the real study
# set. None of these belong in a public repo.
slides.md
promptslides.txt
table.csv
census_out/

# Internal detail Gradient wouldn't want public: these name the live PHI GCS
# bucket, the vendor/BAA legal matrix, the enforcement-layer inventory, and
# internal ops paths.
CLAUDE.md
transfer.sh
build_manifest.py
scripts/

# Git worktrees created by EnterWorktree (local scratch checkouts, never committed)
.claude/worktrees/
# Manually-created worktrees for parallel phase sessions. Nested INSIDE the project dir on
# purpose: the managed PHI guard and every managed deny glob are absolute-path scoped to
# /home/user/intern_ocr_transfer (phi_guard.py PROJECT_DIR), so a sibling ../repo-<phase>
# worktree would run with the root-owned rails silently disabled. Nesting keeps them armed.
.worktrees/