Randomized-Format Field Reader / V71
Randomized training formats lifted the reader to 97.87% field accuracy and 78.97% completely correct texts on four sealed-only formats. Downstream accuracy was 85.41% versus 87.38% from true fields. Locked development and sealed gates passed within this configuration family.
CONTROL COMPARISON
A 101,386-parameter head reads frozen ModernBERT-base states and copies numeric values from source positions. Randomized training adds variation in names, order, syntax, and presence wording. Eight teacher formats are independently written; four are development formats and four first appear in sealed scoring. The sealed phase executes 800 configurations, with reported headline results on the four sealed-only formats.
Both development and sealed gates passed: at least 95% fields, 70% fully correct texts, downstream within five points of true fields, and five points over V70 field accuracy. Randomized training is stronger than the ablation and V70 on the matched set. The causal comparison arm failed its fit check and is informational only.
- V71 · sealed-only formats
- 97.87% fields; 78.97% all eleven right; 85.41% downstream
- V70 unchanged · same formats
- 76.27% fields; 4.72% all right; 45.28% downstream
- Number-only pointer ablation
- 79.28% fields; 7.75% all right
- True fields · same runs
- 87.38% downstream
Question and method
A 101,386-parameter head reads frozen ModernBERT-base states and copies numeric values from source positions. Randomized training adds variation in names, order, syntax, and presence wording. Eight teacher formats are independently written; four are development formats and four first appear in sealed scoring. The sealed phase executes 800 configurations, with reported headline results on the four sealed-only formats.
Recorded results
| Condition | Recorded result |
|---|---|
| V71 · sealed-only formats | 97.87% fields; 78.97% all eleven right; 85.41% downstream |
| V70 unchanged · same formats | 76.27% fields; 4.72% all right; 45.28% downstream |
| Number-only pointer ablation | 79.28% fields; 7.75% all right |
| True fields · same runs | 87.38% downstream |
Comparative standing and locked gates
Both development and sealed gates passed: at least 95% fields, 70% fully correct texts, downstream within five points of true fields, and five points over V70 field accuracy. Randomized training is stronger than the ablation and V70 on the matched set. The causal comparison arm failed its fit check and is informational only.
Interpretation and limitations
One configuration family with eleven fixed facts and teacher-authored templates. Keras, tables, and shell variables remain weaker. This does not establish arbitrary real-file reading, missing-field handling, units, contradictions, or free prose. Later versions and format audits provide broader tests.
SOURCE PROVENANCE
EMMA V71: the field pointer reader with randomized training formats
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
11d1f89c04cb1d988892fba109d5ff409fadfe63eaa808690e812afb9af55e91Public 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.