Reviewed public research · October 31, 2026
NewBrain native API agreement: one unscored two-restore test
Unscored software consistency: one ordinary two-restore comparison, not a teaching, semantic or Original12 experiment.
What completed
Restore A called talk; restore B called model.generate using A's exact effective message sequence in the same expanded native development lineage. Both returned identical complete native reports and true EOS. There were zero updates; model parameters, counters, source and original bundle remained unchanged.
One earlier attempt failed
Before the valid pair, an ordinary call failed before generation because a JSON list was supplied instead of the required tuple and the root codec rejected a live-model object reference. The failed attempt was preserved, not relabeled as a valid generation, and its consumed question was not simply retried.
Interpretation and boundaries
Call-path agreement does not establish reasoning, language learning, held-out accuracy, checkpoint adoption or an Original12 improvement. Accepted Original12 remains 2/12, not rerun. The mirrored copies across three repositories describe one experiment, not three replications. Private inputs, outputs, names and retained model states are excluded. Reviewed GitHub source status ↗.