Complete synthetic response learned and retained after restart

These are the full text pages extracted from 74_NewBrain_Verified_Full_Response_Learning_and_Restart_2026-10-21.pdf, supplied to Kira Labs on October 21. This is a readable web edition, not the original downloadable PDF. Line breaks, tables and graphics may differ from the PDF; numeric claims and experiment boundaries are preserved. The private underlying training artifacts are not published. Original12 remains 2/12; Qwen is untested.

KIRA LABS | INDEPENDENT RESEARCH




 NewBrain learned a complete response
 and preserved it after restart
 Kira Labs | Verified research update | October 21, 2026

 A small, controlled experiment improved from 2/3 to 3/3 complete responses. A separate process then reloaded
 the saved trained state and reproduced 3/3, with no additional training. The no-training control remained at 2/3
 throughout.


   Condition                                     Before              After teaching                Fresh reload



   Trained learner
                                                       2/3                    3/3                           3/3
   No-training control                           2/3                 2/3                           2/3


 What was actually learned
 One fixed synthetic query initially produced "b". After teaching it produced the required complete sequence "cab
 word one", followed by the normal end marker. It produced the same sequence after reload. Two
 missing-information responses remained exactly "unknown", with normal endings and no added tokens.

 The three positions in the synthetic positive response exercise answer, source-label and version-label formatting.
 These toy labels are not real bibliographic evidence.

 Why this is useful
 The result extends earlier narrow short-answer work to a complete multi-token response while maintaining two
 explicitly rehearsed missing-information responses. It provides a working training configuration to investigate on a
 larger, independently evaluated task.

 Scope and limits
 All three examples were explicitly rehearsed. This establishes a specific acquisition-and-restart result. It does not
 establish held-out generalization, retention of unrehearsed knowledge, broad intelligence, natural conversation or
 readiness to replace an existing language model.

 The separate Original12 assessment remains 2/12. Qwen has not yet been tested in the comparison. The new
 3/3 result must not be presented as 3/12 or as a new Original12 score.

 Evidence: training 1073 / independent audit 1074; fresh-process reload 1075 / independent audit 1076. Tests, cleanup, original
 records and all selected resource allowances are closed.




Synthetic research demonstration | October 21, 2026                                                                               1

KIRA LABS | INDEPENDENT RESEARCH




 How the automated test worked
 Methods and interpretation | Same frozen plan for training and restart

 1. Freeze the task and starting state
 Three fixed public synthetic examples were selected before execution: one positive mapping and two
 missing-information responses. Eligibility required the positive response to be wrong before teaching and both
 missing-information responses to be correct. There was no resampling. A preserved starting checkpoint was cloned;
 earlier failed trained states were not continued.

 2. Run one fixed, balanced teaching dose
 The learner received 256 native stochastic-gradient updates. Every update used all three examples, with learning
 rate 0.01 and gradient clipping at 5. This totaled 768 example presentations, 4,608 prefix tokens and 2,048 target
 tokens including end markers. Each example appeared 256 times.

 Measurements were fixed at updates 0, 64, 128, 192 and 256. The complete-response counts were 2/3, 2/3, 2/3, 3/3
 and 3/3. Only the final 256-update checkpoint governed acceptance. Decreasing training loss alone did not count as
 successful learning.

 3. Save, then reload in a separate process
 A distinct worker loaded the saved final state. It performed zero training updates and zero loss evaluations, making
 six generation calls: three taught cases and three no-training-control cases. State identity, parameters, output bias
 and four training counters matched the accepted training evidence. The taught outputs persisted exactly.

 4. Score the original outputs independently
 The independent audit read saved originals and made no new model calls. Correctness required the entire expected
 token sequence, a normal end marker, no unwanted extra tokens, no unknown-input-token error and no
 response-length-limit failure. Constant-output controls failed the joint three-case criterion.

 The audit also checked process records, containment, complete output streams, cleanup and scoped file I/O. Both
 launches passed the existing 8 GiB free-RAM startup guard; the 3 GiB reserve and other execution limits remained
 in place. These checks are not continuous RAM/CPU measurement or proof of every loaded binary's behavior.

 What changed, and what comes next
 A previous separate 64-update curriculum scored 2/3 before teaching and 0/3 afterward. That negative result and its
 state remain preserved. The present balanced 256-update study changes dose and training composition, so it does
 not isolate a single causal defect or prove which change was responsible.

 The next research step is to transfer the successful method to the original authorized teaching material while
 preserving the frozen 12-question assessment. Learning objectives, missing-information behavior, independent
 evaluation and the Qwen comparison still require investigation. Additional hardware can support reproducible work; it
 cannot guarantee a higher learning score.




Synthetic research demonstration | October 21, 2026                                                                      2