Kira Labs / Source-report reading edition
Report 64 · One-Step Training Update
Document: 64_NewBrain_One_Step_Training_Update_Verification_2026-10-20.pdf. This reading edition reproduces text extracted from the original report supplied for publication, separated by source page. It is not a newly verified experiment, and PDF layout, line wrapping and any diagrams may differ. For a test's current status, consult the linked Kira Labs evidence page and dated research updates.
Transcribed source page 1 of 2
NewBrain: verified training-update mechanics
Kira Labs | 20 October 2026
A controlled one-update engineering diagnostic, independently checked from saved evidence.
11,914 0 3 of 3
parameter values checked bit mismatches wrong arithmetic controls rejected
What passed
A single native training update was applied to a disposable clone of a preserved checkpoint. A separate process recomputed
the expected clipped update from the saved before-state and gradient, then compared every saved parameter value. 8,917
values changed; all 11,914 matched the expected result exactly. The maximum absolute parameter error was zero. All
four exposure counters matched their expected increments.
The reported gradient norm was 10.80523171276576; the independent calculation differed by approximately 1.78 x 10^-15.
The clip multiplier, 0.4627388040270147, matched exactly. Teacher-forced loss decreased from 2.355075 to 2.327038.
This is an optimization measurement, not a correct-answer learning result.
How the independent check worked
The checker read the immutable saved evidence from the one completed execution. It did not rerun the model, train another
checkpoint, or generate answers. It reconstructed the expected arithmetic and checked before/after values, counters, file
identities, complete streams and scoped cleanup records.
Deliberately wrong saved-data calculation Detected mismatches Outcome
Reverse the update sign 8,917 Rejected
Double the update scale 8,917 Rejected
Shift parameter indices 9,058 Rejected
These three controls demonstrate that the checker detects these particular arithmetic errors. They were not additional training runs and do
not prove it detects every possible defect.
What this result does not establish
This verifies one-step update mechanics for one frozen state. It does not establish complete gradient or optimizer
correctness, successful acquisition of an unfamiliar fact, generalization, a higher Original12 score, or equivalence of every
loaded binary. The diagnostic used one training update, zero generated answers and no reusable checkpoint product.
Numeric before-state, gradient and after-state observations were saved for arithmetic checking.
Kira Labs | Verified engineering evidence and explicit research limits 1Transcribed source page 2 of 2
Methods, limits and next evidence
A reusable technical explanation for university and research discussions.
Execution and evidence boundaries
The run retained the 8 GiB startup guard, 3 GiB available-memory reserve, two-CPU affinity and approximately 0.8 GiB
NewBrain family limit. Startup and post-run available-memory samples exceeded the startup guard. Two samples do not
demonstrate continuous memory compliance or measured CPU attribution. Stop checks and finite deadlines remained
part of the execution controls.
The scoped saved-evidence review covered 218 events, 108 raw cells, 252 copied references, 19 once-close records, two
empty job records and four EOF-terminated streams. File lengths, hashes, original execution returns and counters were
joined before acceptance. Evidence and resource records were closed before parallel work resumed.
An earlier audit stopped because its manifest omitted two required original records. That failed audit and its costs were
preserved. A distinct corrected audit checked the same unchanged training execution successfully. The training update was
not replayed. An earlier metadata-only startup failure is also preserved separately.
Standing learning results remain separate
Assessment Accepted result and limit
Original12 main assessment 2/12 before and 2/12 after 256 updates and fresh-process reload. Two new correct responses and
two lost baseline missing-information passes; no net gain. The older 64-update 0/12 state remains
preserved.
Public2 taught-answer diagnostic 1/2 to 2/2 after a separate reload. Narrow taught-answer persistence result; not Original12
improvement or proof of unfamiliar-fact acquisition.
Separate isolated three-row study 2/3 before and 2/3 after 64 updates. The new association remained incorrect; two absence
responses remained correct. Cold reload is unrun.
Numerical derivative sample 80 coordinates and four directions across ten parameter groups passed; two wrong-gradient
controls were rejected. This is a separate sampled check, not complete gradient correctness.
Qwen comparison Untested. No comparison result is claimed.
Next learning milestone
Teach a fixed unfamiliar association, test a complete freely generated answer and its ending, retain fixed earlier and
missing-information responses, and repeat recall after a fresh-process restart. Use matched no-teaching and deliberately
incorrect-label controls. Freeze the cases and criteria before execution; never substitute easier cases after seeing results.
Evelyn and Patient Horizon
Evelyn's reviewed private package is prepared, but installation, provider startup, generated conversation, voice,
visitor-memory persistence and the animated head still need runtime evidence. Private user installation and testing precede
public release. Physiological lung function and lung-brain coupling remain research objectives, not demonstrated clinical
capabilities.
Equipment and university expertise would support reproducible diagnosis, acquisition-and-retention experiments, anatomy and simulation
validation. More RAM or a larger GPU alone does not guarantee successful learning.
Kira Labs | Verified engineering evidence and explicit research limits 2