Training-Outcome MicroModel Composition / V67
A 3,374-parameter composition predicted actual small training-run outcomes with 85.50% accuracy on 800 sealed cases. A flat neural control reached 86.37%; added strong tree controls reached 84.50–85.12%. No advantage over the strongest controls is established. Candidates remain unpromoted.
CONTROL COMPARISON
800 sealed numeric training-run configurations. Strong tree controls were added post hoc after opening.
No measurable modular advantage over the strongest controls is established. The later tree controls are not a fresh sealed qualification; candidates remain unpromoted.
- Learned composition
- 85.50%
- Flat neural control
- 86.37%
- Extra Trees · post hoc
- 85.12%
- Random forest · post hoc
- 84.87%
- Gradient boosting · post hoc
- 84.50%
Task and protocol
Predict divergence or stalling before executing a small 200-step training run. Labels come from actual CPU training using sixteen numeric configuration fields. Two independently owned 1,666-parameter MicroModels and a 42-parameter combiner total 3,374 parameters. Data comprise 3,600 training, 400 development, 400 calibration, and 800 sealed configurations, with nine fit budgets per arm.
Results
| Sealed condition | Accuracy |
|---|---|
| Learned composition | 85.50% |
| Exact composition | 85.25% |
| Flat neural control | 86.37% |
| Extra Trees, added post hoc | 85.12% |
| Random forest, added post hoc | 84.87% |
| Gradient boosting, added post hoc | 84.50% |
The learned composition’s calibration error was 0.047. All original gates passed, but candidates remained unpromoted. Fresh-process restoration reproduced all 800 probabilities; 223 protected weight artifacts remained unchanged.
Interpretation and limitations
The later V67.1 tree comparisons were added after the sealed set opened. They are post hoc controls, not a new qualification. Their results are statistically tied with the composition within the reported sampling uncertainty of approximately 2.4 percentage points. No modular accuracy advantage over the flat model or strongest trees is established. This numeric synthetic task does not qualify a text reader, a Foundation, or real-world training prediction.
SOURCE PROVENANCE
EMMA V67: MicroModels that predict whether a training run will fail
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
6b350bef8ebe4cd1be3543bc8d71c1be0d296a58e8b20b28b624c002110be90ePublic journal edition reviewed 2026-10-08. 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.