Archive recognizer teacher pilot and unapproved geometry/text review data

data/recognizer-data-v3-line-review-v1-20261005/README.md

@@ -0,0 +1,79 @@
+# Recognizer DATA ENGINE teacher pilot: result and crop review
+
+## Audited Remote Run
+
+The returned archive SHA256 is
+`e7f54254c33eeadaf6c873539d2657ca751909cb2f0cb66c4b815bf25ea27811`.
+All 22 members and checksum coverage passed verification against the frozen
+input ZIP, code commit `e24545b22b65d42937fd88714ff99c7b81e50f90`, model
+revisions, reference-free inputs and pinned import versions. Token termination
+and processor geometry were checked and proposals/report reproduced from raw
+predictions. The evidence reports Tesla T4, Python 3.13.15, Torch 2.11.0+cu130
+and Transformers 4.57.6. No inference was rerun locally.
+
+- Qwen: 64 returned, 63 EOS, one 512-token cap/repetition, no missing inputs.
+- TrOCR: 64 returned, 64 EOS, no caps or missing inputs.
+- Proposals: **62 disagreements, one abstention, one agreement**.
+- The agreement differs from the unreviewed source label: both models output
+  `Ołob` where the label contains `Oſob`. Agreement is not a ground-truth rule.
+- Raw emitted glyph counts: Qwen `ſ` 4, `á` 31, `ɇ` 0; TrOCR `ſ` 0, `á` 3,
+  `ɇ` 0. These are output counts, **not recall or accuracy**.
+- Zero clean training examples, zero gold labels, no SOTA claim. No CER/WER
+  quality ranking is computed against these unverified crop/label pairs.
+
+## Confirmed Crop/Label Problem
+
+Visual inspection of `NA1_FT__433925__r011__line000` shows only a word
+fragment in a 204x50 crop. Its unreviewed label is the entire line
+`Na nieſkończoną BOGU Chwałɇ.`. The frozen original 903x109 region shows
+that entire line. This is not solely a recognizer failure: the existing
+component grouping retained a fragment while line-count equality accepted
+the label. Count equality does not establish crop/reference alignment.
+
+This is an agent inspection, not a user review event or gold transcription.
+The original crop and label remain unchanged. At least this pair must be
+rejected or re-cropped and independently reviewed before training. The
+remaining 63 crops are **not** assumed geometrically correct.
+
+## Review Now, Do Not Rerun Colab
+
+Prepared 64 line crops with 35 checksum-verified original source-region JPEGs.
+The offline editor shows the line, full region, original weak label and both
+raw OCR outputs. It exports a separate `slayer-recognizer-line-review-v1`
+history with crop hash, region hash, text changes, reviewer and geometry choice.
+Existing page-review exports keep their previous schema and behavior.
+
+1. Inspect the line crop first. If it contains exactly one complete line,
+   choose **Cała jedna linia**. If it is a fragment, clips letters, combines
+   multiple lines or belongs to a different source line, choose **Odrzuć
+   wycinek** and record the reason. Do not complete cut text from context.
+2. For complete crops, edit the transcription to match the visible image,
+   retaining old spelling, punctuation, abbreviations and historical glyphs.
+   Model/source labels are suggestions, not authority. Choose **Zweryfikowano
+   wzrokowo** only after checking the text; ambiguous readings remain pending.
+3. Save each decision, export history and return the JSON. The first entry is
+   the confirmed fragment, followed by the agreement and capped output.
+   An image-only geometry choice is distinct from a text-verification choice.
+
+The editor blocks verified text with unreviewed geometry, and requires a reason
+for a rejected crop. Rejected geometry stays explicit in queue and history.
+Even verified text does not turn a rejected crop into a training example.
+This stage creates no training labels automatically. Later import must require
+both complete geometry and verified diplomatic text, preserve all originals,
+and maintain document/split and duplicate controls before student training.
+
+## Reproduce The Offline Review
+
+Use the frozen public review data bundle and GitHub code; the bundle contains
+only images and JSON/JSONL metadata, not HTML/JS/Python or model weights.
+After unpacking it, run the current repository command against its directory:
+
+```powershell
+python -m training.build_annotation_review --manifest input/manifest.jsonl --diagnostics diagnostics.json --contexts contexts.json --geometry-review --output review
+```
+
+142 focused CPU tests passed after the auditor/editor changes. Isolated
+Playwright/Edge checks at 1366x900 and 390x844 verified crop/context rendering,
+mandatory geometry and rejection reason, export schema, reload persistence,
+legacy mode and no console errors or mobile horizontal overflow. QA-generated
+events are test-only and are excluded from research evidence and training data.

data/recognizer-data-v3-line-review-v1-20261005/checksums.json

@@ -0,0 +1,6 @@
+{
+  "recognizer-data-v3-line-review-data.zip": "2c514d52bf1254aa6b7e3318e6aa2f6c15b24fa4f229a16865ceaff7c61cdc1b",
+  "review-report.json": "5ad210025d179d4239e063e19e04917f6f48534578625b499b199f45181b1ab0",
+  "README.md": "e223bcc78af9ce0c367bd52b990bc618e5d3e3d6bc733b487ae419d43f3cefb3",
+  "publication.json": "2cd97988f699fbf0feb7fa422eff64236f805fbc3ea0375132e8f7093b0a32e0"
+}

data/recognizer-data-v3-line-review-v1-20261005/publication.json

@@ -0,0 +1,25 @@
+{
+  "schema": "slayer-recognizer-data-v3-result-publication-v1",
+  "visibility": "public",
+  "code_revision": "b2d4b02371690da2de408b449d5c50fbb749f1d5",
+  "code_included": false,
+  "weights_included": false,
+  "sota_claim": false,
+  "training_examples_created": 0,
+  "gold_labels_created": 0,
+  "source_archive_sha256": "e7f54254c33eeadaf6c873539d2657ca751909cb2f0cb66c4b815bf25ea27811",
+  "focused_CPU_tests_passed": 142,
+  "scope": "64 training-source line crops and 35 frozen source regions for human review; not approved labels",
+  "scans_included": true,
+  "scan_license": "CC-BY-3.0",
+  "source_attribution": "IMPACT/PSNC Polish historical print corpus, source IDs/revisions/license retained in manifests",
+  "source_dataset_revision": "c7cb156fb95d2880699c33725bbaf1fbc1008fea",
+  "bundle_sha256": "2c514d52bf1254aa6b7e3318e6aa2f6c15b24fa4f229a16865ceaff7c61cdc1b",
+  "review_manifest_sha256": "ae9a4dcb667b753bb6592a3ee86a1cd54faf5a2839348e4027d44dc4a7fbe0c0",
+  "review_patch_schema": "slayer-recognizer-line-review-v1",
+  "files": {
+    "recognizer-data-v3-line-review-data.zip": "2c514d52bf1254aa6b7e3318e6aa2f6c15b24fa4f229a16865ceaff7c61cdc1b",
+    "review-report.json": "5ad210025d179d4239e063e19e04917f6f48534578625b499b199f45181b1ab0",
+    "README.md": "e223bcc78af9ce0c367bd52b990bc618e5d3e3d6bc733b487ae419d43f3cefb3"
+  }
+}

data/recognizer-data-v3-line-review-v1-20261005/recognizer-data-v3-line-review-data.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c514d52bf1254aa6b7e3318e6aa2f6c15b24fa4f229a16865ceaff7c61cdc1b
+size 25480953

data/recognizer-data-v3-line-review-v1-20261005/review-report.json

@@ -0,0 +1,24 @@
+{
+  "schema": "slayer-recognizer-line-review-preparation-v1",
+  "lines": 64,
+  "source_regions": 35,
+  "source_archive_sha256": "e7f54254c33eeadaf6c873539d2657ca751909cb2f0cb66c4b815bf25ea27811",
+  "original_manifest_sha256": "e89e83506f8179c4a10a76bd29bbf18bbce84cbd8211076307fbcd3cdfcff158",
+  "review_manifest_sha256": "ae9a4dcb667b753bb6592a3ee86a1cd54faf5a2839348e4027d44dc4a7fbe0c0",
+  "review_summary": {
+    "manifest_sha256": "ae9a4dcb667b753bb6592a3ee86a1cd54faf5a2839348e4027d44dc4a7fbe0c0",
+    "pages": 64,
+    "flagged_pages": 0,
+    "issues": 0
+  },
+  "review_patch_schema": "slayer-recognizer-line-review-v1",
+  "geometry_decisions": [
+    "unreviewed",
+    "complete-line",
+    "reject-crop"
+  ],
+  "original_line_labels_changed": false,
+  "training_examples_created": 0,
+  "gold_labels_created": 0,
+  "claim_boundary": "Human review required for both geometry and diplomatic text. Complete-line is separate from text verification. No automatic training promotion."
+}