Codariva — v1 timing supplement 1 — 15 October 2026

This is a bounded supplement to the unchanged turkce-belge-v1 dataset and initial results. It adds six timed rounds per document/model (288 executions), not 288 different documents. All 24 inputs use the same single-column template; no claim of representative Turkish document accuracy is made. No customer data is included.

The protocol was saved before execution. The run started 2026-10-15T10:05:31.005Z and ended 10:06:30.600Z (12:05–12:06 Europe/Rome). Tesseract.js 6.0.1, Node 24.17.0, Apple M4, 10 logical CPUs, 16 GiB memory. Original model and image hashes are checked. Both models are initialized and warmed once before measurement, then retained. Document order reverses each round; each document/model is first exactly three times and second exactly three times. No timed retries.

Timer: OCR recognition plus unchanged v1 extraction. Excluded: image file read, model download/load/init, warm-up, hashing, scoring and writing results. The published v1 baseline remains a separate single-pass result. Re-running overwrites this supplement's results.json and summary.json; preserve copies first.

Reproduce in the two sibling directories supplied in the archive:
  cd turkce-belge-v1
  npm ci --ignore-scripts
  mkdir -p models/fast models/best
  curl -L https://raw.githubusercontent.com/tesseract-ocr/tessdata_fast/4.1.0/tur.traineddata -o models/fast/tur.traineddata
  curl -L https://raw.githubusercontent.com/tesseract-ocr/tessdata_best/4.1.0/tur.traineddata -o models/best/tur.traineddata
  cd ../turkce-belge-v1-timing
  TESSERACT_MODULE=../turkce-belge-v1/node_modules/tesseract.js BENCH_MODELS=../turkce-belge-v1/models node run.cjs
  python3 audit.py

Results (milliseconds; 144 observations/model):
  fast: median 141.091; nearest-rank p95 153.292
  best: median 272.474; nearest-rank p95 285.424
  fast first/second-position median: 140.479 / 141.199
  best first/second-position median: 272.825 / 272.283
No large position effect appeared in this bounded run; this does not prove absence of order effects generally. Across round starts the one-minute OS load average fell from 12.912 to 7.106. Whole-device busy share during rounds was 16.1–18.9%, including OCR and unrelated background applications. This was NOT an idle-device benchmark. No build or other benchmark was deliberately run concurrently. These observations do not isolate a causal load effect or establish an SLA. Six repeated measurements of one image are not independent documents.

audit.py independently recalculates 168 expected fields using Decimal arithmetic, checks input/model hashes, recalculates all 336 baseline and 2016 repeat field scores, verifies order balance and recomputes medians/p95. All repeat field values match v1. audit.json preserves all eleven baseline missing/wrong fields. This is code-assisted verification by OpenAI Codex, not named human expert review.

Failure interpretation (raw OCR is retained in both result files):
- fast, two rotated images: digits of the date are present but 'Tarih;' replaces the required 'Tarih:' label, so extraction is missing.
- best, four low-resolution images: 'KOV' replaces 'KDV'; the amount is present but its label no longer matches.
- best, four low-resolution dates: comma/dot changes remain wrong under frozen exact-match scoring.
- best, SIP-rotated-1: 'SsIP-2026-0013' adds a letter; a separate '1734 ,25' amount is correctly normalized to 1734.25.
Thus this measures OCR plus extraction, not character accuracy alone. The parser was not tuned to repair these test cases after seeing results. A different parser might change outcomes and requires a separately versioned experiment.

No external API charge was incurred. Hardware, electricity, development, maintenance and human correction are unpriced; total cost is not zero. Do not infer production model superiority, general accuracy, cost savings or real-world performance from this experiment. Complex tables, handwriting and varied layouts are outside this dataset.

Licensing follows the original package: original code MIT; synthetic inputs and generated result data CC0-1.0. No model weights, font software or node_modules are redistributed. Official Tesseract repositories and Tesseract.js retain their own licenses. See ../turkce-belge-v1/LICENSE.txt.
