Measured evidence · reviewed October 18, 2026
The values below are from different tests, not a single benchmark or a combined model score. Negative outcomes remain part of the research record.
NewBrain / trained-on Public2
0/2 1/2After 64 updates and 128 repeated presentations, the saved model produced ball correctly after fresh-process reload. The three-token answer ended immediately without a word.
Not evidence of held-out generalization, delayed retention or a useful conversational model.
NewBrain / separate Original12
2/12 0/12Two baseline passes were missing-information responses rather than acquired facts. After a separate 64-update, 384-presentation training run, no complete answers passed.
The preserved 2/12 baseline and failed 0/12 trained snapshot are distinct states. The cause remains unproven.
Fixed-checkpoint gradient sample
30/30Thirty sampled primary coordinates passed at each of three step sizes: 90 central differences, from 181 original objective calls, across ten arrays holding 11,914 total scalars.
Checks a bounded local sample, not every parameter gradient, a higher accuracy score or external university replication.
Avatar / geometry control
512Recorded skin vertices moved through five commanded poses and returned exactly to saved positions under the evaluated geometry-control setup.
Not a verified full-body simulator, passive skin mechanics, touch response, or biological anatomy.
Avatar / distinct synthetic input component
8/8Separate synthetic event-control checks covered input differences, duplicate suppression, independent avatar ownership, epoch/session validation and explicit reset.
Invented numerical inputs; no real nerves or active skin-to-learner connection.
Objective-versus-generation diagnostic
4/4All four of the current diagnostic’s scoped execution controls—normal completion, parent exit, cold-library startup and fixed public worker-limit workload—passed internal independent saved-data review.
The subsequent read-only scientific diagnostic also completed; the controls do not demonstrate better learned answers.
October 18 · completed diagnostic
One fixed Public2 checkpoint and batch. Read-only examination, not retraining.
Two-token loss · including EOS
The saved short answer generated ball and then terminated.
Four-token loss · including EOS
The longer target cup word one still generated an empty answer by immediately choosing EOS.
Combined target-weighted loss
Matches the weighted losses within 1e-12. This rules out a simple batch-reduction mismatch for this fixed checkpoint, but not the cause of early stopping.
Limits: no new SGD updates, scored questions or changed parameters. Individual token logits, loss by position and EOS competition were not measured. No Original12 improvement, delayed retention, Qwen comparison or external replication is implied. Dated research record (access may be required) ↗
After the completed read-only objective check, investigate per-position token logits, end-of-sequence competition, answer-length behavior and a newly predeclared learning/retention experiment. Qualify any body-event input path separately. Any improved result must come from fresh, prespecified scoring rather than replaying the consumed Original12 run.
This is Kira Labs' dated internal research summary. Published code documentation and project-specific source evidence may require private repository access; research and educational demonstrations are not equivalent to scientific replication.
