# Evidence files: Beyond Character Error Rate (preprint)

Aggregate evaluation data behind the manuscript's tables and figures. The values were exported by
script (`paper/manuscript/scripts/export_web.py` in the author's private research repository) from
the audit's recomputed results. No page text, page image, file path or credential is included.

| File | Contents |
|---|---|
| `primary-results.csv` | Raw output (quote layer off). One row per evaluation set × system × metric, with the page count, the baseline and system values, the difference, and the paired bootstrap 95% interval where one exists. |
| `postprocessing-results.csv` | Page-mean DER for base vs. v0.1, with the quote layer off and on, over pages where DER is defined for both systems. |
| `release-gate-verdicts.csv` | The project's release-gate verdicts on the 160 gold pages, with the quote layer off and on. |
| `claims.csv` | The manuscript's principal claims, with dataset, page counts, metric, result, manuscript section, public artifact, reproducibility status and limitations. |
| `experiment-manifest.json` | Systems and revisions, shared training settings, inference settings, evaluation sets with page counts, metric definitions, bootstrap-seed sensitivity, provenance, and the known reproducibility gaps. |

**Reading the numbers**

- **`n_pages`:** for `cer_letters`, all pages scored. For `der_paired`, only pages where DER is
  defined for both systems, which can be fewer (DER is undefined when no aligned letter carries a
  mark in either output).
- **Intervals:** the bootstrap intervals describe variation across the evaluated pages. They are
  not estimates of uncertainty about other books, editions or publishers.
- **Reference status:** "silver" sets have references drafted by an LLM-based OCR pipeline and
  never checked by a person.

**What these files cannot support:** independent recomputation. The page-level predictions,
reference transcriptions and page images are not public; see the dossier's Availability section.