Verified Live Learning / V79–V79.2
V79.2’s revised promotion policy passed all eleven gates. A 305-parameter live unit reached 79.4% versus 72.1% frozen accuracy on 800 sealed runs, while a shuffled-outcome unit decayed. Earlier V79 and V79.1 remain failed because their promotion controls accepted or mismeasured base-rate learning.
CONTROL COMPARISON
Run text reading, prediction, independent execution, and verified feedback inside the prototype. The temporary unit learns during operation. V79.2 uses a fresh stream and requires validation improvement over the best of the frozen estimate, fitted base-rate shift, and majority outcome. It opens the previously untouched 800-run sealed set after validation.
V79.2 passed all eleven gates and is qualified within the tested regime. Older answers, the earlier promoted unit, and frozen tensors remained unchanged; fresh-process restoration passed. Tree preflight figures of 73.0–74.3% are cross-validation references, not matched sealed controls. The fitted-shift reference was not computed on sealed data, a recorded reporting omission.
- V79 · validation
- Real unit 84.6%; noise unit also promoted; gate failed
- V79.1 · unchanged real unit
- Noise/base-rate control gate failed
- V79.2 live unit · sealed
- 79.4%; calibration error 0.025
- V79.2 frozen control · sealed
- 72.1%; calibration error 0.142
- Sealed majority outcome
- 52.7%
- V79.2 shuffled-outcome unit
- Not promoted; decayed at twelve-block limit
Question and method
Run text reading, prediction, independent execution, and verified feedback inside the prototype. The temporary unit learns during operation. V79.2 uses a fresh stream and requires validation improvement over the best of the frozen estimate, fitted base-rate shift, and majority outcome. It opens the previously untouched 800-run sealed set after validation.
Recorded results
| Condition | Recorded result |
|---|---|
| V79 · validation | Real unit 84.6%; noise unit also promoted; gate failed |
| V79.1 · unchanged real unit | Noise/base-rate control gate failed |
| V79.2 live unit · sealed | 79.4%; calibration error 0.025 |
| V79.2 frozen control · sealed | 72.1%; calibration error 0.142 |
| Sealed majority outcome | 52.7% |
| V79.2 shuffled-outcome unit | Not promoted; decayed at twelve-block limit |
Comparative standing and locked gates
V79.2 passed all eleven gates and is qualified within the tested regime. Older answers, the earlier promoted unit, and frozen tensors remained unchanged; fresh-process restoration passed. Tree preflight figures of 73.0–74.3% are cross-validation references, not matched sealed controls. The fitted-shift reference was not computed on sealed data, a recorded reporting omission.
Interpretation and limitations
A hand-declared scope, one new regime beyond V69.1, and one seed. Qualification belongs to the mechanism on this regime; the artifact stays in its run folder and attachment to the default prototype is separate. The two earlier gate failures are preserved. No unrestricted continual-learning claim follows.
SOURCE PROVENANCE
V79 and V79.1: learning during operation inside the running prototype
LABORATORY REPORT / 2026-10-09SOURCE CHECKSUM / SHA-256
30066b663facb3dc56c3fde696069fa97c6625e039d717bcc5330ea8e0a5f7afPublic journal edition reviewed 2026-10-10. 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.