Keep the record.
Definitions, source boundaries and the runnable evidence behind the example.
Definitions close at hand
- Independent unit
- The entity whose new occurrence the evaluation treats as new evidence; document, writer, site or event may be appropriate under different claims.
- CER
- Minimum character edit count divided by nonempty reference length. Our characters are raw Unicode code points.
- Percentage point
- An absolute difference on a percentage scale: 40% minus 30% is +10 percentage points.
- Paired difference
- Candidate CER minus baseline CER for the same document. Negative favors the candidate.
- Empirical bootstrap
- Resample the observed units with replacement. Its assumptions and the chosen statistic determine the question it approximates.
- Development versus test
- Development guides choices. A protected test evaluates the frozen procedure. Inspecting a test to revise the model consumes that boundary.
- Unresolved here
- Any actual OCR performance, power-system usefulness, real-population uncertainty, or measured learning effectiveness.
Read the evidence and reproduce the lab
Primary publications and official documentation were inspected on 4 September 2026. Links near the prose identify supporting claims. The source notes distinguish published indexed text from readable author preprints where direct publication pages were blocked. No source provides the toy’s numeric outcomes.
From the full project folder, run node lab/run.mjs to print both results; run node lab/check.mjs to verify the recorded output. The website’s downloadable CLI copy uses the same model. No data download, random source, learned neural weights or external inference call is used.
The bootstrap enumerates six-to-the-six ordered resamples. It computes the equal-document paired mean, using linear interpolation at index (M−1)p for sorted bootstrap means. Values shown as percentages are rounded to two decimals. The model supports none or NFC normalization explicitly; all fixture scores use none.