Report 66 · Five-Question Learning Pilot

Transcribed source page 1 of 2

 NewBrain: five-question learning pilot
 Training, separate reload and independent evidence review | 20 October 2026

 Result: 2/5 before training, 0/5 after training, 0/5 after reload.
 This separate five-question diagnostic does not replace Original12. The latest Original12 score remains 2/12. All
 earlier checkpoints and failed records are preserved. Qwen remains untested.

 Complete-response measure                                Before               After                Reload

 Trained state                                            2/5                  0/5                  0/5

 No-teach control                                         2/5                  2/5                  2/5

 New structured association                               Incorrect            Incorrect            Incorrect

 Two absence-response guards                              Both correct         Both wrong           Both wrong




 How the test was made
 Five public cases were frozen before the run: one structured association, two positive factual-retention checks and
 two missing-information guards. Both factual checks were already wrong, so they were ineligible as retained
 successes. The two absence guards supplied the baseline passes. Overall acquisition-and-retention eligibility was
 not satisfied; no success is claimed.


 Training and scoring
 A separate state received 64 updates over two taught rows: 128 example presentations and 384 target tokens
 including end markers, with learning rate 0.01 and gradient clipping at 5. Twenty explicit whole-answer loss
 measurements and twenty free-generation calls were recorded. The no-teach control received no updates; this was
 not an equal-compute comparison. Independent scoring required the entire frozen answer, without repairing or
 selecting outputs.


 What changed in the answers
 The new answer was expected to be cab, word, one. It generated unknown before teaching, then b after teaching
 and again after reload. Both previously correct unknown responses changed to a and stayed a after reload. All five
 post-training and reloaded responses ended normally; none hit the response cap or reported an unknown input
 token. Failure at the first content token excludes a suffix-format-only explanation.


 What the lower loss means
 The teacher-forced whole-answer loss for the new association fell from 10.673886 to 3.869227 even though its free
 answer remained wrong. Teacher-forced measurement supplies the correct preceding tokens; free generation uses
 its own decisions. Loss reduction alone therefore does not demonstrate useful acquisition.




KIRA LABS | RESEARCH AND EDUCATIONAL REFERENCE | 20 OCTOBER 2026                                                       1

Transcribed source page 2 of 2

 Separate restart test: what was verified
 A distinct process restored the exact saved trained state and generated five trained-state responses and five control
 responses, with zero additional training updates or explicit loss measurements. Every complete trained response
 matched the incorrect post-training response; the control stayed unchanged at 2/5. This verifies the evaluated
 save/reload and output persistence path, not improved learning.


 Independent evidence and failed records
 The training audit checked 219 events and 106 inverse-hash cells. The separate reload audit checked 219 events
 and 107 inverse-hash cells; it also verified 253 copied references, 19 once-only closures, two empty process jobs
 and four complete streams. A prior saved audit stopped because its request used a different path spelling. That
 failed audit and its costs were preserved; a fresh saved-only audit accepted the unchanged model run. The model
 experiment was not replayed.


 What remains unresolved
 The cause of the learning failure is not established. Earlier sampled-gradient and one-step update checks passed
 only their stated scopes. Direct retrieval, per-token probabilities, generalization and complete optimizer correctness
 remain unproven. Two before/after RAM samples do not establish continuous resource compliance; the 8 GiB
 startup guard remains unchanged.


 Next controlled investigation
 A smaller conditional-teaching diagnostic is being prepared: a short one-token association and explicit
 missing-information counterexamples. Its result remains unmeasured. The earlier 2/2 Public2 result used an
 already-taught checkpoint and familiar short questions; the five-question pilot used a different checkpoint and a
 longer unfamiliar context. These differences do not yet isolate a cause. More updates or larger hardware alone are
 not established as a fix.
 Evidence locator

 Training record 987, independent training review 989, distinct reload 994, preserved failed saved review 996 and successful saved review 998. All
 execution, cleanup, evidence and finite-cost records closed before parallel development resumed.




KIRA LABS | RESEARCH AND EDUCATIONAL REFERENCE | 20 OCTOBER 2026                                                                                     2