Code and data for "Not flawless, but good enough: Vision-language models and the invisible errors of machine transcription"

Code and data accompanying the article "Not flawless, but good enough: Vision-language models and the invisible errors of machine transcription". The deposit contains what the article's own recommendations ask a transcription project to publish: the per-line output of ten transcription systems over the IAM handwriting database, the scored records behind the results table, the prompts given to each system, run manifests recording model identifiers and configuration, and the unscreened values and bootstrap confidence intervals referred to in the footnotes. Contents:- README.md — full description and reproduction instructions- prompts.md — the instruction given to each system- table1_full.csv — every column of the results table, plus unscreened smooth shares, 95% bootstrap intervals, clean-looking line counts, and the per-system coverage breakdown- second_auditor_uniform.txt — output of the second-auditor robustness check- fig2_censure_line.png — the line image reproduced in the article- code.zip — the analysis layer and the scripts that produced the runs- data_model_outputs.zip — one JSONL per system: reference transcription, model output, per-line CER and WER- data_scored.zip — scored records: every substitution with its class, direction score, context length and anchor confidence, plus the manifest- data_backfill.zip — the second pass over lines initially left empty Every number and figure in the article can be reproduced from data_scored.zip using code/analysis.py. The IAM line images and reference transcriptions are not redistributed here. They are available as the Teklia/IAM-line distribution on Hugging Face.