Continuous Learning from Every Verified Outcome / V108–V108.1
A component that learns from every verified outcome, with constant memory and no training run, raised sealed accuracy from 78.2% to 82.9% on 3,000 configurations, 2.1 points above an online-learning control, with no forgetting of earlier regimes. V108 first gained 3.1 points but forgot 2.1 points of earlier regimes and failed; V108.1 passed every gate.
SCOPE + LIMITATIONS
- One task family and the estimate kind. In the demonstration its effect was local to configurations near executed ones, which led to a learner that generalises (V123).
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Feed a fresh stream of 3,000 verified outcomes, each answered before it is learned, then score once on a fresh sealed set of 3,000 across six regimes. Settings are chosen on the stream only; V108.1 chooses them under a retention check on early regimes. Retention is measured on a regime-ordered stream.
V108 passed the gain, control and cost gates and failed retention. V108.1 passed all four: +4.7 points over the core, +2.1 over the control, under 1 ms per update, and no forgetting. It went into the demonstration.
- V108 · sealed
- 78.1% to 81.2%; control 79.9%
- V108 · retention of early regimes
- -2.1 points; limit -1
- V108.1 · sealed
- 78.2% to 82.9%; control 80.8%
- V108.1 · retention
- 80.0% to 80.3%
- V108.1 · update cost
- 0.75 ms per outcome; constant memory
- V108.1 · regimes
- None below the core
Question and method
Feed a fresh stream of 3,000 verified outcomes, each answered before it is learned, then score once on a fresh sealed set of 3,000 across six regimes. Settings are chosen on the stream only; V108.1 chooses them under a retention check on early regimes. Retention is measured on a regime-ordered stream.
Recorded results
| Condition | Recorded result |
|---|---|
| V108 · sealed | 78.1% to 81.2%; control 79.9% |
| V108 · retention of early regimes | -2.1 points; limit -1 |
| V108.1 · sealed | 78.2% to 82.9%; control 80.8% |
| V108.1 · retention | 80.0% to 80.3% |
| V108.1 · update cost | 0.75 ms per outcome; constant memory |
| V108.1 · regimes | None below the core |
Comparative standing and locked gates
V108 passed the gain, control and cost gates and failed retention. V108.1 passed all four: +4.7 points over the core, +2.1 over the control, under 1 ms per update, and no forgetting. It went into the demonstration.
Interpretation and limitations
One task family and the estimate kind. In the demonstration its effect was local to configurations near executed ones, which led to a learner that generalises (V123).
SOURCE PROVENANCE
Continuous Learning from Every Verified Outcome / V108–V108.1
LABORATORY REPORT / 2026-10-25SOURCE CHECKSUM / SHA-256
9e4197d040d26b025fd71e69fe4ab441d740bb804cc082b069da0ccf2ec2ecd0Public journal edition reviewed 2026-10-28. 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.