Sampling Path for Proposed Changes / V93
For 400 failing training runs the proposal stem sampled 16 candidate changes, ranked them with the frozen composition, and its top proposal fixed 74.3% of runs when executed, against 28.5% for a random candidate from the same set. All four gates passed.
SCOPE + LIMITATIONS
- One family of tasks, 16 candidates, one sampling seed per run. Fixed means progress within 200 steps. In the demonstration, questions about what change would help are answered through this path.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Draw and execute 1,000 fresh configurations and keep the first 400 that diverged or stalled. For each, sample 16 candidate changes to training settings with a recorded seed through the staged path, judge them with the frozen composition, and execute the top proposal. The control executes a random candidate from the same 16. A proposal becomes a committed fix only after it is executed and makes progress.
At least 50% fixed, at least 15 points above random (+45.8), problem and architecture unchanged, and the commit rule on 40 cases checked through the verified path all passed. The frozen composition plays the role of a surrogate model in model-based configuration search.
- Top proposal
- 297 of 400 failing runs fixed · 74.3%
- Random candidate from the same set
- 28.5%
- Runs with an improving candidate found
- 400 of 400
- Problem and architecture unchanged
- 100%
Question and method
Draw and execute 1,000 fresh configurations and keep the first 400 that diverged or stalled. For each, sample 16 candidate changes to training settings with a recorded seed through the staged path, judge them with the frozen composition, and execute the top proposal. The control executes a random candidate from the same 16. A proposal becomes a committed fix only after it is executed and makes progress.
Recorded results
| Condition | Recorded result |
|---|---|
| Top proposal | 297 of 400 failing runs fixed · 74.3% |
| Random candidate from the same set | 28.5% |
| Runs with an improving candidate found | 400 of 400 |
| Problem and architecture unchanged | 100% |
Comparative standing and locked gates
At least 50% fixed, at least 15 points above random (+45.8), problem and architecture unchanged, and the commit rule on 40 cases checked through the verified path all passed. The frozen composition plays the role of a surrogate model in model-based configuration search.
Interpretation and limitations
One family of tasks, 16 candidates, one sampling seed per run. Fixed means progress within 200 steps. In the demonstration, questions about what change would help are answered through this path.
SOURCE PROVENANCE
V93: the sampling path qualified: proposed changes fix failing runs when executed
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
ae2e27c2bb45c4398aedd41bf80dd467162bf8bdac1a0cc7df0555a4e40f3340Public journal edition reviewed 2026-10-22. 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.