Positional Rewrite and Transfer Audit / V76
An exact rewrite improved complete reading from 71.39% to 77.97% on a third independent format set, but failed sealed field and record floors. It remains a retained improvement, outside the prototype default. The audit corrects the broader standing of the V72 reader.
CONTROL COMPARISON
Place a deterministic layout rewrite before the frozen V72 reader, preserving original value positions. Score eight new formats on 800 newly executed configurations, yielding 6,400 texts. No model is trained; the rewrite and protocol are frozen before scoring.
Development passed; sealed qualification failed despite a 6.58-point gain and passing downstream and improvement conditions. The rewrite is retained best for positional handling but unqualified. V72’s original 91.52% on another format set must be reported alongside the later 71.39%, rather than generalized to all unseen formats.
- Reader with rewrite · sealed
- 77.97% texts fully right; 97.47% fields
- V72 without rewrite · same texts
- 71.39% texts fully right; 94.79% fields
- With rewrite · downstream
- 84.64% versus 86.62% from true fields
- Required sealed floors
- 92% complete texts; 98.5% fields
Question and method
Place a deterministic layout rewrite before the frozen V72 reader, preserving original value positions. Score eight new formats on 800 newly executed configurations, yielding 6,400 texts. No model is trained; the rewrite and protocol are frozen before scoring.
Recorded results
| Condition | Recorded result |
|---|---|
| Reader with rewrite · sealed | 77.97% texts fully right; 97.47% fields |
| V72 without rewrite · same texts | 71.39% texts fully right; 94.79% fields |
| With rewrite · downstream | 84.64% versus 86.62% from true fields |
| Required sealed floors | 92% complete texts; 98.5% fields |
Comparative standing and locked gates
Development passed; sealed qualification failed despite a 6.58-point gain and passing downstream and improvement conditions. The rewrite is retained best for positional handling but unqualified. V72’s original 91.52% on another format set must be reported alongside the later 71.39%, rather than generalized to all unseen formats.
Interpretation and limitations
The rewrite misses a tuple variant and does not fix unfamiliar category names. Same-reader scores vary from 29% to 100% by format. A wider format distribution is needed; this sealed set is now spent development data.
SOURCE PROVENANCE
EMMA V76: an exact rewrite of positional layouts in front of the V72 reader
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
87565a38fef900c6cd91df81fdee7f0908b2e74a7ec21d04acfbb9253c046359Public 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.