Opening-Step Prediction / V75–V75.2
Forty opening steps improved prediction, but V75, V75.1, and V75.2 failed their locked gates. V75.1 reached 89% versus 91.5% for boosting on the first development set; on fresh runs the neural and tree models tied at 89.75% and 89.50%. Sealed runs stayed unopened.
CONTROL COMPARISON
Predict the outcome of a 200-step synthetic training run after observing forty steps, using opening losses, configuration fields, and the frozen composition estimate. Compare small residual neural heads and a default gradient-boosting owned module. The V75.2 comparison uses 400 newly executed development runs.
All forty-step versions failed. V75.1 missed the 90% floor and parity-with-boosting requirement; V75.2 missed 90% on fresh data. The fresh comparison corrects the apparent earlier tree advantage: neural and tree results are tied within sampling variation. Sealed configurations remained unexecuted.
- V75 · first development set
- 86.75%
- V75.1 · first development set
- 89.00%
- Boosting · first development set
- 91.50%
- V75.1 · fresh development set
- 89.75%
- V75.2 tree · fresh development set
- 89.50%
- Frozen composition · fresh development
- 83.00%
Question and method
Predict the outcome of a 200-step synthetic training run after observing forty steps, using opening losses, configuration fields, and the frozen composition estimate. Compare small residual neural heads and a default gradient-boosting owned module. The V75.2 comparison uses 400 newly executed development runs.
Recorded results
| Condition | Recorded result |
|---|---|
| V75 · first development set | 86.75% |
| V75.1 · first development set | 89.00% |
| Boosting · first development set | 91.50% |
| V75.1 · fresh development set | 89.75% |
| V75.2 tree · fresh development set | 89.50% |
| Frozen composition · fresh development | 83.00% |
Comparative standing and locked gates
All forty-step versions failed. V75.1 missed the 90% floor and parity-with-boosting requirement; V75.2 missed 90% on fresh data. The fresh comparison corrects the apparent earlier tree advantage: neural and tree results are tied within sampling variation. Sealed configurations remained unexecuted.
Interpretation and limitations
Observing twenty percent of execution buys information but does not qualify a cheaper prediction policy. V75.3 tests longer openings. Earlier gates remain failed even though later sealed measurements show higher forty-step scores.
SOURCE PROVENANCE
EMMA V75 and V75.1: judging a training run again after its opening steps
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
7cf0ce54ad5ff82af7f2690f5cfb719493c0d4c7ee8a61e35e2accb96b984260Public 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.