← NewBrain research overview

Research record · October 27–28, 2026 · Kira Labs

Learning in NewBrain's original core—and an Original12 test with no net gain

Two different experiments used the original 11,914-parameter, 42-token neural decoder. A fresh symbolic-learning task demonstrated narrow direct acquisition. A separate weighted intervention on Original12 changed which answers were correct but did not improve the benchmark score. Neither experiment adopted a new accepted checkpoint.

October 28 · Original12 weighted intervention

Measured partitionPreserved baselineNew weighted candidate
Complete correct answers2/122/12
Previously correct abstention controls2/20/2
Previously incorrect substantive questions now correct0/102/10

One fixed, preregistered 384-update intervention completed 2,304 teaching presentations, with a normalized emphasis on the first answer token, learning rate 0.01 and global gradient clip 5. Final state, logs and receipt were saved before any protected evaluation query was acquired; no intermediate scoring, dose search or checkpoint selection occurred.

Seven synthetic mathematical checks passed, covering all ten parameter families; the largest sampled weighted-coordinate derivative absolute error was 7.51×10⁻¹⁰. Twelve complete native answers reached true EOS. A distinct ordinary process restored the state without updates and reproduced the complete report. Independent saved-evidence and separate model-free score reviews confirmed the zero-net-gain and gain/loss accounting.

Conclusion: no benchmark gain. Both prior correct missing-information answers were lost, even though two formerly incorrect substantive answers became correct. The original checkpoint was preserved and the trained candidate was not adopted. This intervention combined extra development teaching with weighting, so the effect of weighting alone was not isolated against matched standard-training exposure.

October 27 · NB21 symbolic acquisition in the original neural core

Fresh taskFrozenTrained
Direct known-token symbolic questions0/1211/12
Reordered questions0/124/12

NB21 used isolated copies of the historical original core and unchanged standard cross-entropy/gradient code. Across three fresh worlds it taught four known-token symbolic associations per world, completing 768 updates and 3,072 teaching presentations. All three worlds gained new direct correct complete answers (4, 3 and 4 out of 4). A distinct process reproduced all 48 frozen/trained answer records, including exact emitted tokens, EOS, state/counter information and numerical scores, with no further updates.

This separate task did not measure delayed retention, new vocabulary, broad conversational understanding or the accepted Original12 rubric. Reordered-question performance remained only 4/12.

What should be tested next?

Preserve both completed tests and preregister fresh independent acquisition-plus-abstention-retention cases with equal teaching budgets, complete free answer scoring and distinct-process restart. Include a matched standard-weighting control to isolate the effect of token weighting. Do not tune against protected Original12 answers or reuse consumed cases.

Private state and protected benchmark content are not published. Mirrored reports in Kira World and Patient Horizon describe these same results, not separate confirmations. The work is experimental research, not external peer review, general intelligence, Qwen parity, medical validation or an autonomous synthetic person's abilities.