← Back to NewBrain

Public reviewed research · October 27, 2026 (UTC)

NewBrain NB15 and NB16: two tested transfer ideas fail

These are separate 8,343-parameter synthetic reference studies, not the accepted NewBrain checkpoint. Accepted Original12 remains 2/12, with the prior trained 0/12 failure retained. Positive NB14 value weighting is preserved as a different study; cross-study protocols and worlds must not be pooled.

NB15 · New-question form teaching

Direct-only versus varied-form teaching matched the foundation, architecture, value-token weighting, optimizer, old replay, training updates and presentation budgets. The NEW third question form was reserved outside teaching. NEW alternate questions were taught in the varied arm, so they were not the main held-out measure.

ConditionNEW direct /24NEW alternate /24NEW third held-out /24OLD /72
Frozen00047
Direct-only1912954
Varied1413755

Varied-minus-direct primary transfer was −2/24, with zero positive worlds out of six. The fixed support gate failed. Across 36,864 updates and 294,912 examples, 18 saved states reproduced all 432 generated answers after separate-process reload. Nine original and nine restarted answers were incorrect immediate-EOS empty third-form outputs. Independent saved-evidence review was not independent retraining.

Source: public NB15 verified status ↗

NB16 · Replaying more OLD question forms

DA replayed OLD direct/alternate forms; DAT also replayed OLD third-form questions. NEW direct teaching stayed matched and NEW third wording was not taught. DAT replaced some older repetitions and processed 8,192 more prefix tokens; FLOPs were not matched.

ConditionNEW direct /24NEW alternate unseen /24NEW third held-out /24OLD /72
Frozen00043
DA1910750
DAT167852

DAT gained three and lost two third-form answers, net +1/24, below the preregistered +6 and with only one positive world out of six (required four). NEW direct and alternate scores each declined by three; OLD retention was not loss-free. The study completed 36,864 updates and 294,912 presentations and matched 864 original-plus-restart true-EOS returns from 18 saved states in a distinct process with zero new training and no empty answers. Five new contract checks passed.

Source: public NB16 verified status ↗

Next research boundary

The recorded priority is to qualify the accepted NewBrain checkpoint's exact source, vocabulary and counters using separate public controls before transferring any learning intervention. Protected Original12 answers must not be used for teaching or tuning. Neither study tested Qwen, natural conversations, perception or avatar control.