Source-report reading edition · 20 October 2026
Report 69 · Longer-Answer Diagnostic
A separate longer-answer test stayed 2/3 before and after training: the first content token improved, but the required source and version tokens were omitted.
Original document: 69_NewBrain_Full_Response_Diagnostic_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.
Source page 1 of 2
KIRA LABS | VERIFIED RESEARCH UPDATE | 20 October 2026 Longer-answer learning diagnostic 2/3 → 2/3 Complete answers before and after training · Separate three-question test The latest controlled experiment improved the first answer token on one synthetic example, but did not produce the complete required answer. The two explicitly rehearsed missing-information responses stayed correct. The no-training control stayed at 2/3. The predefined full-response success criterion failed. Positive example Generated answer Before training b + end marker After training cab + end marker Required complete response cab word one + end marker In this diagnostic, “cab” is the answer token, “word” represents its source and “one” its version. These are synthetic research labels. Generating only “cab” earns no complete-answer credit. What improved, and what remains unresolved The model changed the wrong first token to the correct one. Its teacher-forced whole-answer loss on that example fell from 9.692280 to 2.136134. However, it generated an end marker before the source and version. Lower training loss and a correct first token are useful observations, but neither establishes a complete learned response. How this relates to the positive restart result The earlier, shorter-answer diagnostic remains a verified positive result: 2/3 before training, 3/3 afterward and 3/3 after a distinct fresh-process restart. Its no-training control stayed at 2/3. All three rows were explicitly rehearsed. Report 68 explains that study. The longer-response experiment here is separate and does not replace its result. Main Original12 benchmark: still 2/12. Qwen comparison: untested. The latest experiment did not rerun the main benchmark, establish generalization, or demonstrate an unseen real-world fact. Kira Labs | Full-response learning diagnostic | 20 October 2026 1
Source page 2 of 2
Method, evidence and next work How this test was built and performed Three public synthetic queries and their complete answer sequences were fixed before the outcomes. Eligibility required the positive example to be incorrect initially and both missing-information checks to be correct. All eligibility checks passed, and no examples were resampled. The trained arm received 64 gradient updates, 192 example presentations, 1,152 prefix tokens and 512 target tokens including end markers. The batch normalization denominator was 8. Both missing-information examples were explicitly taught. The separate no-training arm received no updates. The procedure recorded 12 explicit whole-answer loss evaluations and 12 free generations across arms and assessment points. Training also made 64 internal update-objective evaluations. Complete-answer scoring required all expected tokens and the correct ending; a constant answer could not satisfy the joint criterion. Independent verification A separate saved-data audit checked the retained training outputs without making a new model or training call. It verified the original evidence references, exported file-I/O accounting, complete output streams, recorded ownership and cleanup. The experiment and all selected resource allowances were closed, preserving used and unused costs and all failed originals. The audit verified 107 native evidence cells, 253 copied references, 219 events, 19 canonical once-close operations, four complete output streams and two empty jobs. These checks establish scoped execution evidence; they do not establish continuous RAM/CPU measurement, complete optimizer correctness or a diagnosis of the learning failure. Decision and next investigation The full-response criterion failed, so the proposed cold confirmation remains unrun and unfunded. The failed bridge does not authorize another exploratory bridge or automatic progression to the twelve-question experiment. The next work is analysis of the existing answer, source, version and ending evidence, together with static review of target construction and decoding. Early stopping is the observed failure. A specific implementation defect, memory-readout problem or forgetting mechanism has not been established. Computing resources The run completed on the current machine with the existing 8 GiB available-RAM startup guard, 3 GiB reserve guard and two-CPU limit. More RAM, GPU capacity and laboratory support can enable larger reproducible studies and concurrent work; they do not guarantee a learning improvement. https://github.com/rmcmurrer81/newbrain/blob/main/docs/research/2026-10-20-current-status-and-patient-horizon-plan.md https://github.com/rmcmurrer81/KiraWorld/blob/main/docs/research/2026-10-20-current-status-and-patient-horizon-plan.md Kira Labs | Full-response learning diagnostic | 20 October 2026 2