NewBrain / Reviewed November 18, 2026 / Closed negative learning study
Fixed copy-learning continuation: training completed; supported copying failed
Two fixed local CPU neural-learning arms each completed 3,072 updates from the same preserved parent state. The A/w curriculum and wording-only control both failed strict supported factual-value copying at 0/64. The experimental lane is not a qualified conversational model.
Exact outcome denominators
| Measure | Initial | A/w arm | Wording-only arm |
|---|---|---|---|
| Primary exact | 3/96 | 3/96 | 2/96 |
| Supported copying | 0/64 | 0/64 | 0/64 |
| Wording | 0/32 | 0/32 | 0/32 |
| Distractor | 0/16 | 0/16 | 0/16 |
| Combination | 0/16 | 0/16 | 0/16 |
| Multi-turn turns | 0/16 | 0/16 | 0/16 |
| Consumed retention | 4/16 | 6/16 | 6/16 |
Acquisition was 3/192 in A/w and 2/192 in wording-only. Both strict copying gates failed. The narrow consumed-retention rise from 4/16 to 6/16 does not establish reliable learning or improvement over earlier studies with different starting baselines.
Verified engineering versus failed capability
The original attempt failed its text-token admission guard before any update. Its failure was retained. A separate successor corrected the guard for legal one-byte token payloads and completed the frozen planned data, equations and step counts. Its 48 contract/separation guards passed; 192 training rows were admitted. Saved-artifact review logged 233,205 checks. Complete 192-decision evaluations reproduced byte-exactly in fresh operating-system child processes, but acquisition was not independently cold-repeated. These are engineering/integrity observations, not neural-learning qualifications.
Limits, privacy and next gate
The evaluated cases were previously exposed by the original baseline program; they were reused rather than collected as virgin or human-independent evaluations. The 34,924-parameter Native copy lane is separate from NewBrainLM, with no trained 9B checkpoint. The accepted Original12 remains 2/12, not rerun. No adoption, live companion, Qwen replacement, audio or vision result, or external GPU provider use was established.
The proposed next step is a bounded review of a separate transformer training/evaluation pipeline and later independent controlled counterfactual tests. This proposal has not demonstrated a root cause or an executed remedy. Only reviewed aggregate measurements appear here: no raw research files, private records, prompts, checkpoints, synthetic fixtures, unlicensed assets or rejected images are distributed.
NewBrain research status · Dated update · Lab timeline · Ask Iris