Kira Labs / NewBrain · November 17–18 Eastern / November 18 UTC
Native copying studies completed; the learning targets failed.
This is a reviewed account of saved experimental results, not a product or public model release. No private prompts, model states, personal records, research source, character media or original outputs are published.
Earlier fixed 640-update trial
Two real local CPU learning arms each completed 640 updates. On fresh exact replies, the original baseline and both trained arms scored 0/64. Both trained arms scored 0/16 on the copying assay. On consumed training examples, the bridge-only branch scored 0/64 and ordinary continuation 12/64, which is not held-out success. Retention fell from the earlier original 12/16 to 9/16 (bridge-only) and 1/16 (ordinary continuation). Both trained branches failed all four predeclared usefulness criteria; neither was adopted.
Paired versus rehearsal: two further fixed 960-update arms
Both arms resumed the failed 640-update bridge-only state, retaining its optimizer lineage. Each ran 960 additional CPU updates. Both failed all five copy-assay conditions, including the separately authored wording check:
| Measure | Paired curriculum | Rehearsal control |
|---|---|---|
| Primary, all conditions | 0/96 | 0/96 |
| Supported facts | 0/64 | 0/64 |
| Unsupported facts | 0/32 | 0/32 |
| Separate wording check | 0/32 | 0/32 |
| Previously consumed retention | 4/16 | 4/16 |
The shared failed parent scored 9/16 on the same retention controls. Paired training acquisition was 10/96 consumed cases, all exact “unknown” abstentions; there was no positive factual-payload copying. Rehearsal training acquisition was 0/64. The earlier original S3 retention figure of 12/16 is a separate baseline, not the parent for this later comparison. Primary cases were correlated views, not 96 independent facts.
What was and was not verified
Software validation recorded 118 tests run (84 passed, 34 skipped), with 12/12 packaged trainer and 15/15 evaluator checks. Separate bridge optimizer mechanics had passed seven synthetic transitions and 47 rejection observations. Saved-artifact reviews found no covered inconsistency and fresh-process restoration reproduced even the erroneous replies. These are bounded engineering/persistence checks, not neural-learning success. The reviews were AI author/domain-aware, not human-independent, and all scored study cases are now consumed. The two curricula changed multiple factors together, so no causal explanation of the failure was isolated.
A future copy-head proposal remains source-only and unrun; no state was adopted into a live Kira World person, Qwen replacement or approved model. Accepted Original12 remains 2/12, not rerun. Original Evelyn head, fictional-character reconstruction, anatomy/physiology research and Patient-Horizon have independent qualification criteria.