Reviewed internal experiment · October 30, 2026 · Separate open-language development study
A saved word lesson survived restart, but NewBrain did not learn the complete word meaning
Evidence level: a real, bounded word-lesson experiment with independent review. Memory and model bundles round-tripped, while meaning, rephrasing and application stayed 0/3 correct. This is not a successful general-conversation or word-understanding result. The historical Original12 remains 2/12 and was not re-evaluated.
Model and teaching conditions
The preserved original 11,914-parameter / 42-token checkpoint was explicitly migrated without training into a separate UTF-8 language extension with 300 vocabulary entries and 32,812 scalar parameters. Existing parameter blocks and counters were preserved, but tokenization, response distributions and framing changed; the old and new versions are not proven behaviorally equivalent. The original checkpoint and its complete saved state remained byte-identical.
One authorized, independently sourced common-word lesson consisted of a short question, an explanation and a correction. It was rehearsed with 16 fixed full-parameter updates. These are neural rehearsals of one lesson, not 16 conversations. The test used the native NewBrain core, not an external conversational answer engine or protected benchmark answers.
Measured complete-answer results
| Separate evaluation | Before lesson | After lesson |
|---|---|---|
| Meaning, rephrasing and application (three prompts) | 0/3 | 0/3 |
| Missing-information control | 0/1 | 1/1 |
| Application with explicitly included saved source-memory context | Not reported | 0/1 |
The three post-training factual answers consisted only of the definition's initial article followed by a genuine end-of-sequence token. That article appeared even before teaching; the result is not evidence of an acquired first word. Teacher-forced loss fell from 10.40055 to 7.34226 without producing a complete correct definition.
Saved state and exact restart
Both ordinary direct CPU invocations exited 0 (reported inner times 0.2798498 and 0.1496528 seconds). Ten primary answer records, including the initial interaction, and nine post-restart records yielded 19 EOS-terminated answers. The nine fixed post-training evaluation answers and full generation records matched exactly in a distinct process restart with no additional training. Saved bundles round-tripped. The source explanation, correction history and 16 confirmed teaching targets remained available after restart.
Weights changed in existing neural arrays, but the parameter count and array shapes did not grow during teaching. A previously zero routing projection became nonzero; its routed-residual norm remained about 9.38×10⁻¹⁰ to 4.39×10⁻⁹. This does not establish that routing caused a useful response or that neurons or structure were added.
What this does and does not prove
The experiment establishes a narrow working lesson/save/restart pathway and one improved missing-information control, not successful semantic learning or understanding. It does not qualify ordinary open-ended conversation, unfamiliar combinations, long-term transfer, live user identity authentication, whole physical I/O or power-loss persistence. All original states, consumed inputs and previous negative results were preserved. The accepted Original12 score is still 2/12; that benchmark was not rerun in this experiment.
A sensible next study is a prospectively registered gradual communication curriculum. Score single-word production, correct usage, meaning comprehension, retention after restart and fresh wording/generalization separately. Distinguish recalling a stored explanation from neural-weight learning. No private lesson text, voice, identities, state arrays or scoring keys are published here.