Reviewed November 18, 2026 · Executed, overall unqualified study
Mixed rehearsal recovered old synthetic answers, not reliable new binding
A fixed 2,048-update local CUDA synthetic continuation (step 12,560→14,608) recovered replayed old controls 1/16→16/16, but four of five frozen gate families failed. Known-order complete mappings moved 26/48→30/48, unseen order 10/48→11/48, and new wording with previously taught facts 1/32→3/32. Positive unseen-order binding remained 5/32; positive wording stayed 1/24. Replayed old controls represent reacquisition, not new-memory discovery. Accepted Original12 remains 2/12, not rerun.
Exact before-and-after scores
Strict complete-answer scoring requires the exact answer, newline and end marker. The same 208 evaluation cases were measured before and after; not every correct unknown reply is successful factual copying.
| Evaluation cohort | Before | After |
|---|---|---|
| Replayed old controls (consumed) | 1/16 | 16/16 |
| New complete contexts, known order | 26/48 | 30/48 |
| New complete contexts, unseen order | 10/48 | 11/48 |
| New wording about old facts | 1/32 | 3/32 |
| Consumed binding retention | 41/64 | 45/64 |
Known-order positive factual binding rose 10/32→15/32, while correct unknown responses regressed 16/16→15/16. Seven-fact positive copying stayed 5/16; eight-fact positives rose 5/16→10/16. The unseen-order improvement was an unknown reply: positive copying remained 5/32 and complete three-condition matches 0/16. New-wording positives were unchanged at 1/24. Four of five frozen families therefore failed; the model was not adopted.
Execution and audit receipts
The preserved experimental state continued for exactly 2,048 local CUDA updates. A fresh-process cold restore repeated 208/208 raw answers byte-for-byte, including errors, and repeated the state. The saved-artifact audit recorded 116,733 checks over 624 records. Repository CI recorded 118 core tests: 84 pass, 34 optional skips; nine additional combined-qualifier fixtures and source checks were saved. These checks do not measure conversational reliability or generalization.
Methods, negative gates and next steps
This was a single-arm, assistant-authored synthetic before/after experiment; no matched no-replay causal arm or human-independent assessment was executed. Replayed examples were consumed in training; 32 novel phrases refer to old teaching facts, and 96 new contexts use familiar pair labels. All 208 evaluation cases are now consumed. Review did not independently rerun tokenizer encoding, Adam or backend numerical training; sampler/CPU/CUDA RNG was intentionally reseeded rather than replayed from parent state.
The next planned evidence would require separately frozen unseen contexts, positive fact-order copying and new-wording tests plus a matched control. This run does not qualify durable lifelong memory, actual-answer multi-turn conversations, live people, emotion, physiology, Qwen replacement or a trained 9B model. Accepted Original12 remains 2/12, not rerun. The experimental checkpoint was preserved but not adopted; no cloud training session or public model deployment occurred.