Reviewed internal experimental evidence · October 30, 2026 · A87 primary study / A92 read-only restart
Two familiar-word answers improved and survived restart, but broader word learning failed
Qualified finding: after 128 neural training updates on one previously taught term, five fixed familiar-word prompts improved from 0/5 to 2/5 correct. A separate process reproduced 20/20 complete generation records exactly, including the mistakes. All six substantive held-out word/meaning/use/context probes failed (0/6), an altered-wording delayed probe failed (0/1), and two previously correct unknown-answer controls regressed (2/2 to 0/2). Final response accuracy was 3/20, including one correct held-out missing-information answer. Full predefined study success was FALSE. Accepted Original12 remains 2/12, not rerun.
Experimental setup
A87 used a separately versioned 300-vocabulary-entry, 32,812-scalar-parameter UTF-8 extension of NewBrain's native recurrent/routing research lineage. This was further practice on a previously attempted term, not first-ever vocabulary learning or 128 distinct conversations. The ordinary primary process completed 128 fixed additional neural updates and 51 free-generation calls; architecture and parameter inventory stayed fixed while weights changed. The original 11,914-parameter/42-token historical checkpoint was preserved separately.
Native scoring separated free answer generation from saved explanations: no retrieved teaching answer or conversation history was inserted into query-time neural input. Private lesson text, prompts and model arrays are not published.
Measured results, including negative findings
| Evaluation | Result | What it supports |
|---|---|---|
| Same five fixed familiar probes | 0/5 → 2/5 | Exact taught-prompt recall and one familiar-use wording; three other familiar probes remained incorrect. |
| Six substantive held-out form/meaning/use/context probes | 0/6 | No qualified fresh semantic transfer. |
| Differently worded delayed prompt | 0/1 | Failure cannot isolate elapsed-time forgetting from changed wording. |
| Two formerly correct known missing-information controls | 2/2 → 0/2 | Regression in uncertainty handling. |
| All final A92 replay answers | 3/20 correct | Includes one held-out correct missing-information answer; not an Original12 result. |
No new vocabulary, newly grown neurons, learned code module, general word meaning or adopted checkpoint was established. The same results mirrored across three private research repositories represent one study, not independent replications.
A92: exact answer persistence, not 20/20 accuracy
One fresh ordinary read-only process restored the saved post-evaluation A87 state and replayed the same 20 final probes with zero additional training updates. All 20/20 complete native generation dictionaries matched exactly, including answer bytes, token IDs, EOS/UTF-8 fields and native metadata. Every replay response ended at true EOS. The model parameter hash, learning counters, memory, owner and limits remained unchanged, preserving both the narrow gains and the incorrect results.
Earlier verification preparations failed because of a normalized literal file path or mismatched runtime-inventory scope; those failures remain recorded and are not retroactively successful. Scoped pre/post runtime inventories agreed on 4,765 files / 120,382,893 bytes, but full physical I/O, peak RAM, complete OS-library closure and power-loss durability remain unverified.
Related software and prospective work
A separate developmental-practice execution passed eight named software assertion functions without a successful new language-learning score. Later Source93 continuing-lesson journaling and explicit authorization code passed bounded static review, but its new parse, import, runtime, compatibility and learning checks are all UNRUN. It is source implementation only, not another completed learning experiment.
A future study must prospectively freeze independent unfamiliar-word tests, meaning/use probes and uncertainty-retention controls without tuning on already consumed held-outs.
Ethical and evidence boundaries
These reviewed observations do not establish broad conversation, semantic understanding, human developmental age, autonomous embodiment or clinical safety. The historical accepted Original12 still scores 2/12 and was not evaluated or replaced in A87/A92. No private teaching content, subject identity, media, secret token or internal checkpoint is being published. Kira Labs reports internal experimental observations with their negative outcomes and denominators intact.