Wide-Lexicon Field Reader / V77
V77 improved on the matched V72 reader but failed its sealed gates: 85.2% of texts completely right against a 95% floor, with a worst format of 14% against 80%. It is retained best, unpromoted; the prototype keeps V72.
CONTROL COMPARISON
Train a wider-lexicon reader behind positional rewriting, using diagnosed open data. Freeze the version and score the sixteen reserved format-pool templates once, each on the same 400 configurations as the open half. Compare V72 behind the same rewrite after the reserved set has opened.
Fit and development gates passed, but sealed mean and worst-format gates failed. The candidate improves matched average reading yet trails its required reliability. V77 is retained best, not promoted or qualified. The V72 sealed reference was taken post hoc and is not an independent new qualification.
- V77 · reserved half
- 85.2% fully right; worst format 14.0%
- V72 with rewrite · matched post hoc control
- 83.0% fully right; worst 13.8%
- V77 · open development half
- 99.8% fully right; worst 98.2%
- V77 · downstream sealed
- 81.6% versus 83.0% from true fields
- Untrained category-word probe · ungated
- 30.6% records fully right
Question and method
Train a wider-lexicon reader behind positional rewriting, using diagnosed open data. Freeze the version and score the sixteen reserved format-pool templates once, each on the same 400 configurations as the open half. Compare V72 behind the same rewrite after the reserved set has opened.
Recorded results
| Condition | Recorded result |
|---|---|
| V77 · reserved half | 85.2% fully right; worst format 14.0% |
| V72 with rewrite · matched post hoc control | 83.0% fully right; worst 13.8% |
| V77 · open development half | 99.8% fully right; worst 98.2% |
| V77 · downstream sealed | 81.6% versus 83.0% from true fields |
| Untrained category-word probe · ungated | 30.6% records fully right |
Comparative standing and locked gates
Fit and development gates passed, but sealed mean and worst-format gates failed. The candidate improves matched average reading yet trails its required reliability. V77 is retained best, not promoted or qualified. The V72 sealed reference was taken post hoc and is not an independent new qualification.
Interpretation and limitations
Eleven of sixteen reserved formats were perfect, but vertical tables and normalization wording failed badly. Open scores overstated sealed complete reading by approximately fourteen points. One teacher authored the templates and lexicon; configuration identities are shared across open and reserved halves. The reserved pool is now spent and future qualification requires fresh formats.
SOURCE PROVENANCE
V77: reader with a wide lexicon, behind the positional rewrite
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
9b1c004d4c6780367ec18e7c353d66f00beec026bb194e6c4bd161cbf1655191Public 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.