Why Learning Showed No Gain on Unseen Configurations / Study
Replaying each rehearsal’s own stream of 1,640 verified outcomes showed that the demonstration’s continuous learner improves estimates near configurations it has executed, but moves accuracy on distant, unseen configurations by only a few cases either way under every setting tried. The learning gate as written could not separate that from noise. The finding led to a learner that generalises (V123).
SCOPE + LIMITATIONS
- Two streams, five settings, and offline replay against the frozen prototype.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Replay the verified streams of two rehearsals offline, in order, under the demonstration’s settings and four alternatives. Measure accuracy every 400 outcomes on probe configurations absent from the stream: 239 and 244 cases.
Development analysis with no gate. It proposed splitting the learning gate into near-seen and unseen parts and giving it statistical power; a larger, paired test was used from V123.
- Replay against the rehearsals
- Reproduced: 82.8% and 84.4% final
- Every setting · unseen configurations
- Within about 2 points (about 5 of 240 cases), direction changing along the stream
- Earlier pass and later fail
- The same behaviour seen through two samples
Question and method
Replay the verified streams of two rehearsals offline, in order, under the demonstration’s settings and four alternatives. Measure accuracy every 400 outcomes on probe configurations absent from the stream: 239 and 244 cases.
Recorded results
| Condition | Recorded result |
|---|---|
| Replay against the rehearsals | Reproduced: 82.8% and 84.4% final |
| Every setting · unseen configurations | Within about 2 points (about 5 of 240 cases), direction changing along the stream |
| Earlier pass and later fail | The same behaviour seen through two samples |
Comparative standing and locked gates
Development analysis with no gate. It proposed splitting the learning gate into near-seen and unseen parts and giving it statistical power; a larger, paired test was used from V123.
Interpretation and limitations
Two streams, five settings, and offline replay against the frozen prototype.
SOURCE PROVENANCE
Why Learning Showed No Gain on Unseen Configurations / Study
LABORATORY REPORT / 2026-10-27SOURCE CHECKSUM / SHA-256
532a30167c246478cbc4a5263c8cc9a07d730305b8043ddfd4379141dc736c42Public 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.