← NewBrain project

Public research record · October 25, 2026

Five-world replay comparison: retention improved, composition worsened

This completed synthetic association study is separate from the accepted NewBrain Original12 learner. It tests a small retention/generalization question, not broad conversational intelligence or Qwen replacement.

Primary balanced replay: acquisition 13/20, held-out question-form composition 7/20, old cases 38/40, unseen abstention 0/10.

Primary blocked old-then-new: acquisition 13/20, composition 12/20, old cases 14/40, unseen abstention 0/10.

Accepted Original12: 2/12 unchanged. Previously failed trained 0/12 was not promoted.

Method

Five fresh world/seed pairs (seeds 2026102507 through 2026102511) each used a 34-token fixed vocabulary. Each world received 4,096 foundation updates; three independently reloaded arms then ran 1,024 updates. Both primary arms saw identical old/new example multisets and total work. The new-only arm was auxiliary with twice the new example dose. World D's foundation reached 6/8 old answers and was retained, rather than being discarded from the denominator.

Total observed work: 35,840 updates; 286,720 examples; 1,720,320 prefix tokens; 1,146,880 target tokens. A distinct ordinary Python process reloaded all 15 arm states and reproduced all 270 complete answers with zero training updates. Saved-data reviews checked receipt/hash chains, states and 630 complete answer outputs for consistency; they did not independently rerun every SGD calculation.

Per-world primary old-case and composition scores

WorldInitial old /8Balanced old /8Blocked old /8Balanced composition /4Blocked composition /4
A88202
B88434
C88423
D66212
E88211

The 40 old-case aggregate is not 40 previously correct answers: the initial foundations answered 38/40. Conditional preservation among the 38 previously correct answers was 38/38 for balanced replay and 14/38 for blocked training. No world was dropped.

Interpretation and caveats

Balanced replay protected old answers but did worse on unfamiliar alternate-form questions. Neither condition handled unseen abstention. This is a tradeoff, not an unconditional learning improvement.

Five world/seed pairs are descriptive, not a statistically established general effect. Seed and world varied together, and presentation order and within-update gradient mixing also differed together, so their individual causal contributions cannot be isolated. The study does not establish open-ended language competence, real-world transfer, independent external replication or any accepted Original12 gain.

The detailed source, numeric receipts, private reviews and checksum archive are retained under the NewBrain research project's research/retention-replication1207/ folder. This public record excludes private files.