Reader Failure Taxonomy / V1
Error analysis separated numeric association mistakes from category-vocabulary failures. In the third open format set, rewriting reduced association errors from 2,379 to 371, but category errors rose from 1,288 to 1,411. No gate, training, or sealed qualification applies.
CONTROL COMPARISON
Inspect frozen V72 predictions on two already-open sets, each containing 6,400 texts. Classify wrong fields as another field’s value, a distractor, a missing value, or a wrong category. A 240-text follow-up checks momentum and presence wording.
This is diagnosis on spent data, not qualification. The rewrite fixes many label-to-value pairings but does not solve category wording and worsens that error count in one set. Correct number detection must not be mistaken for correct field association.
- Third set · numeric association errors
- 2,379 without rewrite; 371 with
- Third set · category errors
- 1,288 without rewrite; 1,411 with
- Open pool · association errors
- 289 without rewrite; 91 with
- Open pool · category errors
- 33 with or without rewrite
- Numeric distractor / missing errors
- Zero in both inspected sets
Question and method
Inspect frozen V72 predictions on two already-open sets, each containing 6,400 texts. Classify wrong fields as another field’s value, a distractor, a missing value, or a wrong category. A 240-text follow-up checks momentum and presence wording.
Recorded results
| Condition | Recorded result |
|---|---|
| Third set · numeric association errors | 2,379 without rewrite; 371 with |
| Third set · category errors | 1,288 without rewrite; 1,411 with |
| Open pool · association errors | 289 without rewrite; 91 with |
| Open pool · category errors | 33 with or without rewrite |
| Numeric distractor / missing errors | Zero in both inspected sets |
Comparative standing and locked gates
This is diagnosis on spent data, not qualification. The rewrite fixes many label-to-value pairings but does not solve category wording and worsens that error count in one set. Correct number detection must not be mistaken for correct field association.
Interpretation and limitations
Momentum aliases and bare presence flags account for many category failures. Wider vocabulary training and additional layout handling are proposals tested later in V77. The reserved formats were untouched at diagnosis time and subsequently opened by V77.
SOURCE PROVENANCE
Reader failure taxonomy v1
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
46fe3689f71a17d28be24b900790e1eabdac19743f5e8ddb6dc145ae3ce84514Public journal edition reviewed 2026-10-09. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.