Wider-Format Reader and Joint Decoding / V72
On eight new formats and 800 executed configurations, V72 read all eleven fields correctly on 91.52% of texts and reached 98.42% field accuracy. Locked gates passed. Later audits found 71.39% complete reading on another format set, so the original pass does not establish uniform transfer.
CONTROL COMPARISON
Retain the 101,386-parameter head and frozen ModernBERT encoder; widen the format generator to 14,400 training texts and jointly decode seven different numeric source values. Eight new sealed formats produce 6,400 texts. Earlier formats serve as regression checks. The designer knew sealed format kinds, but not their content, before extending training.
Development and sealed gates passed: at least 98% fields, 85% fully correct texts, downstream within three points of true fields, and three points over V71 complete reading. Six of eight sealed formats had no errors. These are internal matched controls, not an established outside benchmark.
- V72 · new sealed formats
- 91.52% texts fully right; 98.42% fields; 82.67% downstream
- V71 with joint decoding only
- 88.06% texts fully right
- V71 unchanged
- 79.89% texts fully right
- True-field control
- 83.62% downstream
- Later third format set · V76 audit
- 71.39% texts fully right without rewrite
Question and method
Retain the 101,386-parameter head and frozen ModernBERT encoder; widen the format generator to 14,400 training texts and jointly decode seven different numeric source values. Eight new sealed formats produce 6,400 texts. Earlier formats serve as regression checks. The designer knew sealed format kinds, but not their content, before extending training.
Recorded results
| Condition | Recorded result |
|---|---|
| V72 · new sealed formats | 91.52% texts fully right; 98.42% fields; 82.67% downstream |
| V71 with joint decoding only | 88.06% texts fully right |
| V71 unchanged | 79.89% texts fully right |
| True-field control | 83.62% downstream |
| Later third format set · V76 audit | 71.39% texts fully right without rewrite |
Comparative standing and locked gates
Development and sealed gates passed: at least 98% fields, 85% fully correct texts, downstream within three points of true fields, and three points over V71 complete reading. Six of eight sealed formats had no errors. These are internal matched controls, not an established outside benchmark.
Interpretation and limitations
CSV and SQL positional layouts remained at 67.25% and 64.88% complete reading. Subsequent V76 and V77 records correct the broader standing: unseen formats are heterogeneous, and one passing set is insufficient to establish general reliability. V72 remains the qualified prototype reader while later higher-scoring candidates remain unqualified.
SOURCE PROVENANCE
EMMA V72: the field pointer reader with a wider format generator and joint decoding
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
536b360c3ef1fa3875a75f3d899ecdd703fcca32c529ab2546cf5551f0582096Public 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.