{
  "sample": "S3HT-62RW-C1Y, one page from Ledin (2026)'s final test set",
  "selection": "Selected by the author as the best-performing of 14 final test pages for the blur-trained adapter; not a representative sample",
  "source": "Figure 5.16, PDF page 64, extracted losslessly with Poppler pdfimages",
  "source_preprocessing": "Author's median-blurred image; 2037 x 1024 grayscale. Convert to RGB PNG without rescaling.",
  "reference": "Figure A.31, PDF page 92. Figure A.30 is a model prediction and is not used as truth.",
  "reference_extraction": "Poppler pdftotext -layout; select numbered rows between captions A.30 and A.31; replace TeX low double-quote glyph with two comma delimiters; preserve other characters",
  "models": [
    "datalab-to/chandra",
    "Saucystream/fine_tuned_chandra",
    "datalab-to/chandra-ocr-2",
    "Saucystream/fine_tuned_chandra_blur"
  ],
  "inference": "Same RGB PNG input, BF16, greedy temperature 0, maximum 8192 output tokens; no reference labels in prompts",
  "prompts": "Both base models use chandra-ocr 0.2 OCR_PROMPT and output HTML. The published adapter uses its author's JSON prompt. This compares extraction workflows.",
  "adapter": "Primary comparison uses fine_tuned_chandra, trained on untransformed pages. The secondary run uses fine_tuned_chandra_blur, trained on median-blurred pages and used for the thesis's final evaluation. Every model receives the same blurred thesis figure. BF16 merged inference is not the author's quantized runtime.",
  "fields": [
    "full_name",
    "sex",
    "age",
    "occupation",
    "literacy",
    "property_owner"
  ],
  "reference_columns": [
    [
      2,
      3
    ],
    [
      4
    ],
    [
      5
    ],
    [
      9
    ],
    [
      10
    ],
    [
      12
    ]
  ],
  "normalization": "Unicode NFC, collapsed whitespace, empty string to *, integers to strings; retain case, accents, punctuation and ditto marks",
  "alignment": "Unique printed row number. Score fields independently; duplicate or absent row IDs and unavailable fields become NaN. Retain other rows. Exact hash-bound source-audited maps are a separate assisted result.",
  "denominator": "15 people x 6 fields = 90 cells; also report the 67 filled cells separately",
  "format": "Keep complete records, including complete JSON prefix objects if truncation occurs. Unreadable or absent fields are NaN; report syntax/termination separately. No complete structured record means 0% coverage and undefined reading accuracy.",
  "diagnostics": "Period-only, no/nó and accent-fold sensitivities operate on available cells. Exclude five flagged reference cells only in the separately labeled sensitivity.",
  "uncertain_cells": [
    [
      9,
      "full_name"
    ],
    [
      10,
      "full_name"
    ],
    [
      12,
      "full_name"
    ],
    [
      12,
      "occupation"
    ],
    [
      14,
      "full_name"
    ]
  ],
  "reference_uncertainty": "Retain the author's spellings even where handwriting is ambiguous; no model-assisted correction of reference labels",
  "review_amendment": "An independent visual review flagged five reference cells and missing accents. Sensitivities and the secondary blur run were added after the first inference request but before inspecting any model response.",
  "interpretation": "Agreement on one author-selected held-out page, not overall generalization performance or statistical significance",
  "distribution": "Download thesis, scan and labels on demand into ignored build/. Publish model predictions and aggregate/per-field counts only.",
  "html_projection": "Use printed-column positions, respecting explicit spans. The source-audited OCR2 holdout mapping shifts only left-page selected fields; automatic scores keep the original fixed map.",
  "parser_amendment": "The colspan rule was added after inspecting Chandra 1 output. It interprets explicit HTML cell spans; it does not infer missing columns or move field contents.",
  "layout_diagnostic": "The previously exploratory source-checked OCR2 map is now included only in explicitly assisted results. See ../../layout_audit.json for the exact mapping and hash binding. Automatic coverage and exact yield remain separately reported.",
  "layout_amendment_utc": "2026-10-12T23:58:50.979248+00:00",
  "layout_amendment": "After inspecting all model outputs, added emitted/aligned row counts and a separate post-hoc OCR 2 layout diagnostic. Primary scores and reference labels are unchanged.",
  "freeze_record": {
    "frozen_utc": "2026-10-12T23:42:58.321704+00:00",
    "protocol_sha256": "e26e7b4066cf7682a66c0e64fbefd30ca6cefdd5dae6d471e88bfd311cc2388e",
    "reference_sha256": "5f6e5f20fc238a770323a88fd4dfd84157903d8c6f6643b3ba66f3e2de090365",
    "reference_people": 15,
    "reference_widths": [
      20
    ],
    "filled_cells": 67,
    "model_runs_started": false
  },
  "reference_review_record": {
    "amended_utc": "2026-10-12T23:44:26.005432+00:00",
    "protocol_sha256": "69e5c559ffd0d29efb4c076935eea97d844f75e1fd37be2b0934f0e4e11da558",
    "reason": "Independent visual review of reference; no model response inspected",
    "reference_cells": [
      [
        9,
        "full_name"
      ],
      [
        10,
        "full_name"
      ],
      [
        12,
        "full_name"
      ],
      [
        12,
        "occupation"
      ],
      [
        14,
        "full_name"
      ]
    ]
  },
  "partial_scoring": {
    "normalization": "NFC and collapsed whitespace; preserve case, punctuation, accents, ditto marks and all predicted characters. Paris ! and ? annotation markers become spaces; Argentina * is an extracted blank.",
    "missing": "NaN means unavailable extraction/alignment, not a readable blank. Exclude NaN from conditional accuracy; count it in missing coverage. Incorrect and extracted-blank cells remain scored.",
    "character_accuracy": "1 - total Levenshtein edits / available filled reference characters, calculated within cells. Insertions, deletions and substitutions cost 1, with no clipping; accuracy can be negative. Blank reference cells have a separate exact-match measure.",
    "digit_accuracy": "Same edit formula on age or birth-year cells whose normalized reference consists entirely of ASCII digits. Keep all prediction characters, including stray periods and letters. Report the available digit denominator and digit coverage.",
    "coverage": "Available cells / all reference cells, with filled-cell and reference-character coverage separately. Raw automatic coverage is recorded before any hash-bound source audit.",
    "exact_yield": "Exactly recovered cells / all reference cells. Unavailable cells contribute zero correct cells; this measures extraction plus reading. Report automatic and source-audited yield separately.",
    "rejected_fields": "--reject-field FIELD sets that column to NaN, including otherwise readable blanks, while preserving other columns. Repeat the flag for multiple columns. Coverage and rejection metadata expose the choice.",
    "alignment_audit": "layout_audit.json binds source-checked structural maps to exact raw output and reference hashes. No text similarity alignment or label copying. --automatic-only disables these maps; future outputs require their own audit.",
    "amendment": "2026-10-13: scoring revised after viewing outputs at the user request. Raw model outputs and references unchanged. Conditional scores with source assistance do not measure autonomous layout recovery."
  },
  "metrics": [
    "conditional character and digit accuracy",
    "available cell and character coverage",
    "automatic and source-audited exact recovery over all reference cells",
    "conditional exact-cell disagreements",
    "filled and blank agreement",
    "automatic and assisted row counts"
  ]
}