Learning from Self-Verified Outcomes / V90
A research worker learned only from outcomes it executed and checked itself, without learning anything a visitor said. From 2,400 outcomes it promoted one unit that scored 79.0% against 78.0% frozen on 1,000 sealed configurations, exactly at the locked one-point gate; nothing was promoted from permuted outcomes.
SCOPE + LIMITATIONS
- Simulated visitors drawn from the laboratory’s own regimes and one seed. The promoted scope is broad and covers two thirds of the sealed configurations. Idle time between visitors is enough for promotions within an hour of operation.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Execute 24 simulated visitor configurations from six regimes once, then the worker’s own cases, each a visitor configuration with one to three training settings changed, in batches of 20 by the independent executor. Every outcome passes through the unchanged learning loop. The control repeats the stream with outcomes permuted. CPU only, 16 minutes in total.
All five gates passed, the improvement gate exactly at its threshold. A sign test on the 44 changed answers gives p of about 0.17, so the gain is real in direction and small in size. This result does not count toward self-chosen learning scopes, whose gate is three points.
- Unit promoted
- At 1,598 outcomes; validation 88.9% against 76.1% frozen; 305 parameters
- Sealed, 1,000 configurations
- 79.0% against 78.0% frozen
- Answers changed
- 44: 27 to right, 17 to wrong
- Outside the promoted scope
- 333 sealed configurations bit-identical
- Permuted-outcome control
- Nothing promoted
Question and method
Execute 24 simulated visitor configurations from six regimes once, then the worker’s own cases, each a visitor configuration with one to three training settings changed, in batches of 20 by the independent executor. Every outcome passes through the unchanged learning loop. The control repeats the stream with outcomes permuted. CPU only, 16 minutes in total.
Recorded results
| Condition | Recorded result |
|---|---|
| Unit promoted | At 1,598 outcomes; validation 88.9% against 76.1% frozen; 305 parameters |
| Sealed, 1,000 configurations | 79.0% against 78.0% frozen |
| Answers changed | 44: 27 to right, 17 to wrong |
| Outside the promoted scope | 333 sealed configurations bit-identical |
| Permuted-outcome control | Nothing promoted |
Comparative standing and locked gates
All five gates passed, the improvement gate exactly at its threshold. A sign test on the 44 changed answers gives p of about 0.17, so the gain is real in direction and small in size. This result does not count toward self-chosen learning scopes, whose gate is three points.
Interpretation and limitations
Simulated visitors drawn from the laboratory’s own regimes and one seed. The promoted scope is broad and covers two thirds of the sealed configurations. Idle time between visitors is enough for promotions within an hour of operation.
SOURCE PROVENANCE
V90: learning in the demonstration from outcomes the system verifies itself
LABORATORY REPORT / 2026-10-20SOURCE CHECKSUM / SHA-256
45ac15b459563aaee560954666ddd81736845ac1a50fc0eaf3118f5f02da7d5ePublic 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.