Reviewed internal engineering evidence · October 30, 2026 (A98) · October 31 UTC follow-up entry checks (A101/A103)
Continuing-lesson saves survived restart; learning accuracy was not measured
Qualified software result: Two separately authorized native continuing-lesson updates were saved and reloaded without replacing the earlier practice. The final session was restored byte-for-byte in a distinct interpreter, and 22/22 no-mutation rejection checks passed. No new meaning, comprehension, retention or generalization score was established. Historical accepted Original12 remains 2/12.
Exactly what ran in A98
Three ordinary isolated interpreter invocations exited 0 (old producer, candidate, distinct reader). They form one bounded qualification, not three independent replications.
| Stage | Observed count/result | Limit |
|---|---|---|
| Preserved v1 practice | One native update, 14 generations, 15 completed operations; practice-cap stop. | No successful complete free-language output. |
| Continuing lessons | 2/2 authorized transitions completed exactly once, with save/reload after each. | Successful state transitions, not two scored word-learning gains. |
| Native counters | Updates, examples, prefix exposures and target exposures: [1,1,17,5] → [3,3,735,17]. | Not distinct conversations or proof of language understanding. |
| Defensive checks | 22/22 no-mutation rejections passed; one deliberately no-headroom branch retained its genuine native failure without replay. | Not host-atomic recovery or power-loss validation. |
| Distinct-process reader | Exact native model, memory and predecessor/continuation history restored; zero additional reader native operations. | No free-answer accuracy or generalization evaluation. |
This was an independently initialized native 29,977-scalar-parameter, 265-codec-entry experimental instance, not the protected Original12 checkpoint. Its 17 completed final positive-branch operations are distinct from counts of attempted operations across the deliberate failure branch.
October 31 UTC: subsequent entry checks stopped before execution
A101 found that a proposed diagnostic caller had assumed less saved-memory capacity than the existing session requires. It stopped at the read-only envelope check: no question was nominated, no model imported, no forward or teaching update ran, and nothing in saved memory or history was altered. That is a caller-budget mistake, not proof the brain lost its saved memory.
A103 then recorded unqualified preparation, an overlong literal command-line form and unresolved host/resource and custody accounting. The inventory made outside its approved scope was not promoted to verified evidence. New actual-prefix diagnostic Source99/Source102 code remains static-reviewed / UNRUN for this path. No new model response was scored. The mirrored records in three private research repositories are one research sequence, not three experiments.
Do not merge software success with earlier language scores
In separately published A87/A92 research, fixed familiar-word answers rose from 0/5 to 2/5 after 128 updates, but substantive held-out meaning/use stayed 0/6, a differently worded delayed question scored 0/1, and older correct missing-information controls regressed 2/2 to 0/2. Final accuracy was 3/20. Restart reproduced 20/20 complete answer records including wrong ones—not 20/20 correct. A94 previously stopped with a 501-token full context against a 256-token limit, without its planned continuation updates. A98 did not retroactively qualify A94.
Unverified questions and private evidence boundaries
A98 did not establish an improved teaching answer, general conversation, unknown-word use, long-term semantic memory, whole-run peak RAM, complete external physical I/O or power-loss-safe durability. A new experiment must freeze novel prompts and score taught recall, unfamiliar usage, older retention and unknown-information restraint without reusing consumed prompts. Raw saved state, personal identities, private lessons, model bytes, media and protected cases remain unpublished. Historical accepted Original12 remains 2/12, not tested here.