The fifteen problems · 03 of 15 · extraction

A deterministic one-character corruption, caught only by exact bytes

During the 2 August extraction run one Amharic document failed four independent model draws with the same one-character corruption. The byte-exact span contract rejected every draw; a normalizing or fuzzy citation check would have passed all four. The four raw responses were preserved unrepaired, because repairing them would falsify the record.

The single most important language finding in the programme.

✓ The source read

የደረሰበ

× The model claimed, four draws in a row

የደረሰበ

The final syllable changed to . Rendered from system fonts; three bytes per code point in UTF-8. Awaiting native-speaker review.

The extraction contract required every assertion to cite one contiguous verbatim span from the supplied segment with exact byte offsets and exact text. The source read የደረሰበው and the model claimed የደረሰበት. Four independent draws, the same single-character change every time.

× What it broke

Nothing shipped wrong, which is the point; but the all-or-nothing document function discarded the invalid raw response, so rejections were uninspectable until append-only rejected-run records were added.

✓ What it established

The defect is deterministic, so more draws do not fix it, and it is a different failure class from stochastic non-convergence. The join surface of the whole design rests on byte anchors; repairing offsets to make a document pass would have quietly damaged it.

oo-ld-corpus (extraction run 2026-08-02) · u--ra9 · retrieved 2026-08-21

Ask your AI about this page

Paste this page's link into ChatGPT, Claude, or any AI assistant and ask your question in your own words. Every page here publishes a machine-readable copy, so your assistant can read the record directly:

https://ethiopia-build.stoagen.com/problems/03/

Published . Last updated . Times come from this page's revision history and can be checked against it.