Source-report reading edition · 20 October 2026
Report 68 · Learning Retained After Restart
A fixed, explicitly rehearsed three-question diagnostic improved 2/3 to 3/3 and retained 3/3 after a distinct restart; Original12 remains 2/12.
Original document: 68_NewBrain_Verified_Learning_and_Restart_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 3
KIRA LABS | RESEARCH RESULT | 20 October 2026
NewBrain: Learning Retained After Restart
2/3 → 3/3 → 3/3
Before training · After training · After a fresh-process restart
A separate, fixed three-question diagnostic produced a verified learning-and-persistence result. A previously
missed synthetic mapping became correct; two explicitly rehearsed missing-information responses stayed
correct. A distinct process reloaded the saved state and reproduced all three complete correct answers. The
no-training control stayed at 2/3 throughout.
Measured comparison Before Trained Restart
Trained model: complete answers 2/3 3/3 3/3
No-training control: complete answers 2/3 2/3 2/3
Constant “cab” control 1/3 1/3 1/3
Constant “unknown” control 2/3 2/3 2/3
What this establishes
For these three frozen, explicitly taught rows, the update procedure changed parameters and improved
complete generated answers. All preregistered training eligibility conditions were met: the target mapping was
incorrect before training, both missing-information checks were correct beforehand, and the same fixed rows
were scored afterward. No cases were resampled.
Restart persistence: verified
Distinct restart: 3/3. The separate reload run made zero updates, zero loss evaluations and six free generations
across the trained and no-training arms. The complete outputs matched their post-training counterparts. State,
parameter and output-bias hashes and training counters were preserved. The preregistered joint
training-and-restart criterion passed.
Main Original12 benchmark: still 2/12. Qwen comparison: untested. This three-row result is a separate
diagnostic. Every row, including both missing-information responses, was rehearsed. Unrehearsed retention,
held-out generalization and broad prior novelty were not established.
The earlier 2/5 → 0/5 training result and its 0/5 cold result remain part of the research record. This new controlled result does not
replace those observations.
Kira Labs | Verified learning and fresh-process restart | 20 October 2026 1Source page 2 of 3
How the test was made and run
1. Freeze a small, public diagnostic
Three synthetic token sequences were fixed before outcomes. Their purpose was to separate learning one
conditional mapping from producing the appropriate missing-information response. These are research tokens,
not a test of real-world historical knowledge.
Memory cue / query tokens Expected complete Before → trained
answer → restart
<MEMORY> stated 3 <QUERY> name item cab + end marker Incorrect →
correct → correct
<MEMORY> stated c <QUERY> name item unknown + end marker Correct → correct
→ correct
<MEMORY> stated c <QUERY> give word unknown + end marker Correct → correct
→ correct
2. Train only the experimental arm
The trained arm received 64 gradient updates across all three rows: 192 example presentations, 1,152 prefix
tokens and 384 target tokens including end markers. The batch loss used a normalization denominator of 6.
The separate no-training arm provided a stable comparison.
The same fixed procedure recorded 12 explicit whole-answer loss evaluations and 12 free generations across
arms and assessment points. Training also performed its 64 internal update-objective evaluations. A lower
teacher-forced loss was kept separate from answer accuracy: the score required the model to generate the
complete answer itself.
3. Score complete outputs
A response received credit only if it matched the complete expected answer and ended correctly. The checks
rejected extra tokens, invalid or unknown-input conditions, and response-cap failures. A constant-answer
system could not satisfy all three cases: always answering “cab” passed 1/3, and always answering “unknown”
passed 2/3.
Observed objective changes
Teacher-forced whole-answer loss Before After
Mean over the three trained rows 2.239601 0.283517
Previously missed mapping row 6.321978 0.796056
Loss values describe the optimization objective. The complete-answer result above is the behavioral measurement.
Kira Labs | Verified learning and fresh-process restart | 20 October 2026 2Source page 3 of 3
Restart verification and next milestone Saved-evidence review Independent saved-data reviews checked the retained original training and distinct reload runs. The reviews made no new model, training, provider or native experiment calls. They verified the recorded process identities, ownership, job cleanup, complete output streams, evidence references, provider joins and available exported file-I/O accounting. The training audit covered 107 native cells, 253 copied references and 219 events; the reload audit covered 106 cells, 252 references and 218 events. Each verified 19 canonical once-close operations, four complete output streams and two empty jobs. These are scoped execution checks, not continuous physical RAM/CPU measurement or complete loaded-image equivalence. Resource controls The study retained the 8 GiB available-RAM startup guard, 3 GiB reserve guard, two-CPU limit and existing bounded process lifetimes. This bounded study completed on the current machine. More memory can support larger and concurrent experiments; whether those experiments improve learning must be measured separately. Next learning milestone Test whether the same controlled teaching approach can produce a longer answer with its required source and version tokens and a correct ending. Preserve the missing-information controls, fixed scoring and distinct-process verification. This is preparation for the unchanged twelve-question evaluation; it is not a new measured result. How to describe this result “A fixed, synthetic three-question diagnostic improved from two correct complete answers to three and retained all three after a fresh-process restart. Its untrained control stayed at two. All three examples were taught. The larger twelve-question benchmark remains at two correct answers, and the Qwen comparison remains untested.” Research boundaries This experiment jointly taught all three cases. It does not isolate which change from the earlier five-question procedure caused the improvement. It does not demonstrate untrained-question transfer, an unseen real-world fact, general intelligence, or readiness of a full conversational person. It provides a small positive learning-and-restart result to build on. https://github.com/rmcmurrer81/newbrain https://github.com/rmcmurrer81/KiraWorld https://github.com/rmcmurrer81/Patient-Horizon Kira Labs | Verified learning and fresh-process restart | 20 October 2026 3