← NewBrain research overview

Reviewed research note · October 30, 2026 · Software qualification, not language-learning evaluation

NewBrain's separately versioned open-language extension passed software checks

Evidence level: one direct software qualification completed with exit code 0. This is not a functioning general-language conversational demonstration, a trained replacement for the original core, or an improvement in the accepted Original12 result of 2/12.

What was implemented and tested

Source72 preserves the original experimental recurrent/routing architecture as a lineage while adding a separately versioned text-and-memory path. The new path includes UTF-8 byte fallback, an expandable vocabulary and dynamic input/output dimensions; typed ordinary-message input; attributed, owner-scoped conversation and correction records; explicit confirmed teaching calls; configurable context, reply and resource limits; explicit end-of-sequence, length and invalid-UTF-8 states; and versioned save/restart interfaces.

On October 30, one ordinary direct software invocation completed with exit code 0 in 0.238731 seconds. Its six software assertion groups and a separate source/backend guard returned normally. These are groups, not a published count of individual test cases. An independent review confirmed the recorded outcome and matching selected-runtime manifests. Mirrored copies in three research repositories are one qualification, not independent experiments.

The earlier fixed 42-token, 11,914-parameter original decoder rejected a separate plain-language greeting before generation. That original-core observation and this new extension's software checks concern different paths; one does not turn the other into a successful conversational run.

What has not been demonstrated

No real conversation or unfamiliar-word learning was measured. No old research checkpoint was migrated and qualified under the new format; no neural meaning, factual generalization, full save-and-restart conversation, live avatar integration, user authentication or durable host storage was established. The extension's byte coverage describes which text can be represented, not what the network understands.

A byte-capable migration would have 300 output classes before optional lexical pieces, rather than the older 42-token vocabulary; this is an interface design property, not evidence of a successfully migrated historical model. The accepted protected Original12 remains 2/12, with no newly adopted checkpoint.

Whole-system physical I/O, peak RAM and full outer execution timing were not independently qualified. Source-only design records written before the software invocation remain historical unrun records; they should not be mistaken for additional executed experiments.

Next measurement

A separately authorized test should teach one real word or correction, save the model and attributed memory, restart in a distinct process and ask an unfamiliar question. It must distinguish (1) stored explanation retrieval, (2) the neural model's unaided answer and (3) an explicitly memory-assisted answer. Earlier protected questions should not be used to claim fresh generalization or to tune the accepted Original12 benchmark.

This public page summarizes reviewed evidence without exposing private source repositories, protected benchmarks, raw conversations, research checkpoints or unpublished media.