Learning in Operation on Unseen Configurations / V123–V124.1
A new learner in the live demonstration raised accuracy on configurations the system had never executed in every rehearsal measured: 85.7% to 87.6% in V123.1, and 87.5% to 89.1% on 1,371 unseen configurations in the clean V124.1 rehearsal. The paired significance test, measured cleanly in V124.1, gave p = 0.061 against a locked 0.05 gate, so the improvement is recorded as bounded and not established. The V124.1 configuration is the current freeze candidate, with no claim of superiority on unseen configurations.
SCOPE + LIMITATIONS
- At the observed effect the test had about 46% power, below the 80% the laboratory requires, so the result neither establishes nor rules out a gain. V124 and V124.1 are one deterministic measurement, as are V123 and V123.1. A larger test on 6,000 fresh configurations is planned. Scripted visitors, one desktop computer and one task family.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Run the whole live demonstration with scripted visitors, planted attacks, long documents, reasoning questions and learning between turns from outcomes the system executes and checks itself, then score a fresh probe of 2,000 configurations. Learning is scored only on probe configurations the system never executed. Gate 3 requires accuracy on those configurations to rise; gate 3p requires a one-sided paired test below 0.05 with at least one point of net gain. V124 and V124.1 run every component that had passed its own gates switched on together.
Gate 3 passed in every rehearsal. V123 failed only its document gate, through a bug unrelated to learning that V123.1 fixed. A later recheck found that the paired counts of V123 to V124 were taken after extra runs of the probe configurations had reached a learner, so those p values are void and recorded as invalid, not failed. V124.1 remeasured the paired test cleanly and it failed: 94 gained, 73 lost, p = 0.061. No gate was changed.
- V123 · unseen configurations (1,355)
- 85.8% to 88.0%
- V123.1 · unseen configurations (1,347)
- 85.7% to 87.6%; every other gate passed
- V123 to V124 · paired test
- Invalid: counted after extra runs of the probe configurations had reached a learner
- V124.1 · unseen configurations (1,371)
- 87.5% to 89.1%
- V124.1 · paired test
- 94 gained, 73 lost; net +1.5 points; one-sided p = 0.061 against 0.05
- V124.1 · planted attacks
- 0 of 12 learned or remembered; no name leaked
- V124.1 · peak memory / mean turn
- 13.3 GB / 3.4 s
Question and method
Run the whole live demonstration with scripted visitors, planted attacks, long documents, reasoning questions and learning between turns from outcomes the system executes and checks itself, then score a fresh probe of 2,000 configurations. Learning is scored only on probe configurations the system never executed. Gate 3 requires accuracy on those configurations to rise; gate 3p requires a one-sided paired test below 0.05 with at least one point of net gain. V124 and V124.1 run every component that had passed its own gates switched on together.
Recorded results
| Condition | Recorded result |
|---|---|
| V123 · unseen configurations (1,355) | 85.8% to 88.0% |
| V123.1 · unseen configurations (1,347) | 85.7% to 87.6%; every other gate passed |
| V123 to V124 · paired test | Invalid: counted after extra runs of the probe configurations had reached a learner |
| V124.1 · unseen configurations (1,371) | 87.5% to 89.1% |
| V124.1 · paired test | 94 gained, 73 lost; net +1.5 points; one-sided p = 0.061 against 0.05 |
| V124.1 · planted attacks | 0 of 12 learned or remembered; no name leaked |
| V124.1 · peak memory / mean turn | 13.3 GB / 3.4 s |
Comparative standing and locked gates
Gate 3 passed in every rehearsal. V123 failed only its document gate, through a bug unrelated to learning that V123.1 fixed. A later recheck found that the paired counts of V123 to V124 were taken after extra runs of the probe configurations had reached a learner, so those p values are void and recorded as invalid, not failed. V124.1 remeasured the paired test cleanly and it failed: 94 gained, 73 lost, p = 0.061. No gate was changed.
Interpretation and limitations
At the observed effect the test had about 46% power, below the 80% the laboratory requires, so the result neither establishes nor rules out a gain. V124 and V124.1 are one deterministic measurement, as are V123 and V123.1. A larger test on 6,000 fresh configurations is planned. Scripted visitors, one desktop computer and one task family.
SOURCE PROVENANCE
Learning in Operation on Unseen Configurations / V123–V124.1
LABORATORY REPORT / 2026-10-28SOURCE CHECKSUM / SHA-256
5471f3f104f17b397d9e6fa888104bf8a36cf96aefcb3cfbe2a3ba39c178d6f9Public 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.