Report 63 · Isolated Learning Result

Original document: 63_NewBrain_Isolated_Learning_Result_and_Verification_Method_2026-10-20.pdf

This is a complete text reading edition extracted page by page from the supplied PDF, not the original binary PDF. Text, tables and spacing may differ from the PDF; screen captures from Report 70 are linked separately. The original source is retained, and the latest status is on the evidence page.

 NEWBRAIN | RESEARCH ADDENDUM


 Isolated learning study:
 result and verification method
 Kira Labs | 20 October 2026

 The latest small training study did not produce a newly correct answer. Its saved execution evidence
 has now passed the scoped internal audit. The separate process-restart stage remains unrun.

    Fixed public test                                            Before 64 updates                  After 64 updates


    One new association                                          Incorrect                          Incorrect


    Missing-information question 1                               Correct                            Correct


    Missing-information question 2                               Correct                            Correct


    Strict complete-response total                               2/3                                2/3


 The two correct responses were missing-information controls. They were not two newly learned facts. The
 no-teaching control stayed unchanged. All 64 update records showed a changed parameter hash; changed
 weights alone do not establish useful learning.

 How this relates to the main result
 Original12 remains 2/12 before and 2/12 after the latest 256-update main study and a fresh-process
 reload: two newly correct complete responses and two lost missing-information passes. The original baseline
 and the earlier separate 64-update state that scored 0/12 remain preserved.

 A different, earlier Public2 teaching experiment improved from 1/2 to 2/2 and retained that result after reload.
 This narrow success does not establish improvement on Original12 or on the new three-row study. The Qwen
 comparison remains untested.

 Current conclusion: no net improvement on the main assessment, and no new correct association response in this isolated training
 observation. No generalization or acquisition-and-retention success is claimed.




Kira Labs | NewBrain | Verified status: 20 October 2026                                                                       1/3

 How the automated test was made and run
 1. Freeze the question, controls and scoring
 The study used one fixed public synthetic association and two questions requiring an answer that information
 was unavailable. Before training, the association had to be incorrect and both controls correct. Those
 eligibility checks passed. The expected responses and evaluation rules were fixed before the run.

 Generation was scored as a complete response, including its required answer, source/version fields and
 ending. The evaluator did not repair answers, choose a more favorable checkpoint or replace failed cases.
 This was a test of taught examples, not a held-out generalization benchmark.

 2. Run one declared training schedule
 One training process used the native all-parameter update path for 64 updates. The record contains 12
 explicit objective evaluations, 64 objective evaluations inside the update routine and 12 generated responses
 across the training and zero-update control measurements. The zero-update control did not receive equal
 training compute.

 All parallel development paused for the test, cleanup and evidence closure. The 8 GiB available-memory
 startup guard and 3 GiB reserve remained unchanged. The original checkpoint and failed originals were
 preserved; this new state was not adopted as the default model.

 3. Verify saved originals without retraining
 The first saved-evidence audit stopped before scoring because matching large integer file-identity values
 appeared as integers in one receipt and decimal strings in another. That failed attempt remains consumed
 and preserved.

 A narrow checker correction received separate source review. Thirty-five pure metadata controls then
 passed, including exact values above the usual floating-point integer boundary, malformed-value rejection
 and rejection of genuinely changed identities. These controls validate metadata handling, not learning.

 A fresh saved-only audit read the existing run's originals and accepted its scoped process, file, counter and
 response evidence. It did not repeat training. The audit checked complete byte lengths and hashes, original
 reference joins, stream completion, closures and the declared state counters.

 Verification limits: internal role separation is not external peer review. The audit did not independently reconstruct the model, prove
 the complete optimizer correct, establish continuous RAM/CPU behavior or qualify every opaque API-argument ownership detail.




Kira Labs | NewBrain | Verified status: 20 October 2026                                                                                2/3

 What is established, and what comes next
 Verified progress
 The sampled numerical-gradient study passed 80 coordinate and four directional comparisons across ten
 parameter groups, while rejecting two deliberately wrong-gradient controls. The latest evidence-checker
 controls passed 35 comparisons. The native training run and its saved evidence can now be examined as a
 documented negative learning observation.

 These are useful engineering and diagnostic results. They do not substitute for a newly correct answer, a
 higher Original12 score or a functioning clinical simulation.

 What the result does not explain yet
 The new association remained incorrect after teaching, even though the recorded parameter hashes
 changed. The present result does not establish whether the limiting factor is the actual update
 implementation, the distribution of training objectives, representation or selection of the answer, or response
 generation.

 A gradient check concerns a sampled derivative of a specified loss. It does not by itself establish that the live
 optimizer applies the intended update to every array, or that reducing that loss improves a strict
 complete-answer score. More RAM or a larger GPU would support experiments but would not guarantee a
 learning fix.

 Next bounded measurements
 The primary next diagnostic is being prepared to compare one actual optimizer step with independently
 calculated expected parameter changes and counter changes, including deliberately incorrect controls.
 Objective weighting and first-token versus later-token errors also require separate examination. These
 follow-ups are prepared work, not completed tests.

 A distinct-process cold restore of the accepted experimental state remains pending. It must use a fresh test
 admission and preserve the negative training observation. The main Original12 assessment and the Qwen
 comparison also require separately defined, verified runs.

 How a university could help
 The most useful immediate support is machine-learning expertise in optimization, sequence objectives and
 experimental design; independent replication; research assistants to run fixed tests; and stable laboratory
 computing access. Additional memory, GPU capacity and remote server access would improve access to
 reproducible comparisons and larger experiments.

 Related material: PDF61 documents the sampled gradient result and method. PDF58 records the current main learning result.
 PDF59 covers the earlier four-response pilot. PDF62 describes the proposed Patient Horizon lung demonstration. The verified avatar
 result remains movement and restoration of 512 sampled weighted skin vertices; realistic organs and biological nerve-to-brain
 function remain unvalidated.

 This addendum is reusable across research meetings. It separates measured results, unresolved questions and future work. No new
 public-release readiness claim is made for Evelyn's installer, voice or animated head.




Kira Labs | NewBrain | Verified status: 20 October 2026                                                                         3/3