← Development updates · NewBrain status · Ask Iris
Reviewed public aggregate · Later November 2, 2026
Content-state readout: the completed learning qualification failed
A fresh independently authored fixed local study completed 216 genuine teaching updates and 72 full-generation calls. Study integrity checks passed; the learning criteria failed. Accepted protected Original12 remains 2/12, not rerun.
Complete-answer outcomes
| Outcome | Before training | Immediately after | Fresh restart |
|---|---|---|---|
| Directly taught facts | 0/6 | 0/6 | 0/6 |
| Untrained paraphrases | 0/6 | 0/6 | 0/6 |
| Unfamiliar two-fact application | 0/6 | 0/6 | 0/6 |
| Correct unsupported-question abstention | 0/6 | 0/6 | 0/6 |
There were six questions per group, four groups per stage and three stages, totaling 72 generations. Strict source/input/artifact/state/order/custody/process/restart timing checks passed before one sealed scoring pass. Teaching really changed parameters and optimizer state, but no group met the answer criterion. All 24/24 immediate-to-delayed complete answer pairs matched byte-for-byte but remained incorrect: exact persistence of wrong output is not useful learning.
Private source questions, target answers, teacher identities, states, histories and raw custody are withheld. This website report reviews saved public aggregates; the private cases were not independently rerun here.
What the mechanism review establishes—and does not
The optional readout averages delivered recurrent content-position states instead of using only the terminal state. A separate static source-path audit matched 14 pinned function AST records but did not run a learner or demonstrate causal benefit. Episodic retrieval was excluded from this fixed study; persistent cross-question neural working memory was not established.
The earlier November 2 source-only review said this candidate was unrun at that time. It later ran and failed. A preceding attention-policy study scored 2/6 unsupported-question outcomes on different fresh cases; comparing its score to 0/6 here does not prove regression or improvement.
Cases are consumed, with no tuning or replay. Future comparisons require independently frozen new cases. No accepted new checkpoint, demonstrated Qwen replacement, language generalization or Original12 improvement has been established.