Add frozen 34-page eight-family printed replay expansion

data/printed-replay-source-expansion-v1-20261007/README.md

@@ -0,0 +1,86 @@
+# Printed replay expansion V1
+
+## Scope
+
+A frozen input contains **34 source pages from eight new work families**:
+24 candidate pages from six families and 10 probe pages from two other families.
+Scans include historical essays, memoir, regional geography, clipped newspaper
+articles and fiction. This improves source diversity, not representativeness
+for modern administrative documents. Newspaper articles are clippings, not
+complete newspaper-layout benchmark pages.
+
+All retained source revisions have Wikisource quality 4. Scans were checked
+against pinned Commons SHA1 and public-domain license metadata; source
+transcriptions retain CC BY-SA 4.0 attribution and revision links.
+No project human review, training eligibility or SOTA claim follows from this.
+Publication dates unavailable in source metadata remain `null`.
+
+| Family | Domain | Split | Retained pages |
+| --- | --- | --- | ---: |
+| Schneider: Babiagora | Regional geography | Candidate | 2 |
+| Witkiewicz: Teatr | Cultural essay | Candidate | 5 |
+| May: Nad Rio de la Plata | Adventure fiction | Candidate | 8 |
+| Pamietniki lekarzy (1939) | Memoir | Candidate | 2 |
+| Prokesch: Nowa Reforma (1923) | Newspaper article | Candidate | 4 |
+| Zawadzki: Kurjer Warszawski (1904) | Newspaper article | Candidate | 3 |
+| Andersen: Basnie (1929) | Literary fiction | Probe only | 5 |
+| Boccaccio: Dekameron | Literary fiction | Probe only | 5 |
+
+No prior Zeromski/Prus pilot pages or protected work families are included.
+The entire Andersen and Boccaccio families remain probe-only; other editions
+must not be added to training as aliases. Model pretraining exposure is unknown,
+so this is a diagnostic work-held-out probe, not a certified untouched test.
+
+## Discovery Evidence
+
+The first bounded selection requested 29 quality-4 pages, but 27 were short title
+pages and only two candidate pages survived. It had no probe and was rejected.
+The selection and raw results are preserved, not silently replaced.
+
+The final selection used four bounded category metadata batches of 500 pages,
+requiring quality-4 membership and revision size at least 1800 bytes before
+selecting works. Size is only a discovery signal, not proof of useful text.
+Of 44 requested pages, full rendered-body extraction retained 34 and rejected
+10 with footnotes/complex markup. Rejected snapshots and the discovery receipts
+are included in the frozen package; they are not input rows for OCR.
+
+Archive SHA256:
+`95661ff30f5b904f67896e8079d8332d4665cdf9b9b3d579bc4a2e3f6d548769`.
+The source package is independently checksum-verified and the selected image
+previews were inspected for all eight families. Preview JPEGs are not native OCR
+inputs: the notebook retrieves original DjVu files and verifies their SHA1.
+
+## Run One Notebook
+
+Use `training/colab_printed_replay_expansion_v1.ipynb` on **CPU** and run all cells.
+No files, tokens or GPU are needed. The notebook downloads the complete frozen
+source package from a pinned public HF revision and verifies its SHA256.
+Code and work selection are pinned separately to complete revision/checksum
+receipts. Original DjVu downloads total about 152 MB, based on the verified
+Commons sizes; each download is bounded at 100 MB.
+
+The miner uses V2 exact-line rules unchanged: Tesseract Polish, native page
+geometry, minimum word confidence 90, unique exact source anchors, line order
+and geometry checks. No fuzzy matching, spelling modernization or confidence
+relaxation is introduced to increase the yield. The only change is the source
+pool; this isolates whether data diversity and scan quality improve acquisition.
+
+Download **`printed-replay-expansion-v1-evidence.zip`** and return it for audit.
+It includes native PNGs, complete word TSV, crop/text pairs, quarantined lines,
+source receipts, environment versions and checksums. If automatic download is
+blocked, use the Colab Files panel and download the ZIP inside the timestamped
+`printed-replay-expansion-v1-...` directory.
+
+The notebook does **not train** or change the baseline. No expected candidate
+count is promised before OCR runs. If acquisition remains sparse, the next change
+must target segmentation/teacher agreement rather than training on a tiny pool.
+
+## Validation Boundary
+
+The source acquisition and package verification completed locally. Regression
+tests cover work/scan alias separation, frozen input checks, CPU-only notebook
+generation, original V1/V2 behavior and crop audits. Notebook schema and Python
+cell syntax are validated. Native DjVu decoding/Tesseract and Colab end-to-end
+execution are **not verified locally**, because those Linux executables and the
+Colab runtime are unavailable on this laptop. A returned evidence ZIP is the
+required execution check, not notebook generation alone.

data/printed-replay-source-expansion-v1-20261007/printed-replay-source-expansion-v1.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95661ff30f5b904f67896e8079d8332d4665cdf9b9b3d579bc4a2e3f6d548769
+size 16020099

data/printed-replay-source-expansion-v1-20261007/publication.json

@@ -0,0 +1,28 @@
+{
+  "schema": "slayer-printed-replay-expansion-publication-v1",
+  "code_revision": "c8ae68ae741d2c30d95317c2c0bb84d347d79a16",
+  "pages": 34,
+  "candidate_pages": 24,
+  "probe_pages": 10,
+  "work_families": 8,
+  "eligible_for_training": false,
+  "eligible_for_evaluation": false,
+  "files": {
+    "printed-replay-source-expansion-v1.zip": {
+      "sha256": "95661ff30f5b904f67896e8079d8332d4665cdf9b9b3d579bc4a2e3f6d548769",
+      "bytes": 16020099
+    },
+    "README.md": {
+      "sha256": "fb38cda85df3713463f957b8ac211f6a5a2a49fdac79f82d7fb0647ed4d62eae",
+      "bytes": 4763
+    },
+    "source-report.json": {
+      "sha256": "0bfa74eeab277a317274ac13727cf15b76ddeefe35156b2fbcccab88aa800534",
+      "bytes": 376
+    },
+    "preview.png": {
+      "sha256": "023ff17cd620c8e5bf54493d80a3713c1f822edadb9052b24d1a275bea0b98d2",
+      "bytes": 2618937
+    }
+  }
+}

data/printed-replay-source-expansion-v1-20261007/source-report.json

@@ -0,0 +1,15 @@
+{
+  "schema": "slayer-printed-replay-sources-v1",
+  "collected_at": "2026-10-07T13:56:28.462260+00:00",
+  "pages_requested": 44,
+  "pages": 34,
+  "work_families": 8,
+  "splits": {
+    "replay-candidate": 24,
+    "replay-probe": 10
+  },
+  "rejected": 10,
+  "ready_for_line_alignment": true,
+  "eligible_for_training": false,
+  "eligible_for_evaluation": false
+}