Controlled curriculum: negative acquisition result

These are the full text pages extracted from 73_NewBrain_Controlled_Curriculum_Result_and_Method_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.

NewBrain: controlled curriculum result
 Verified research update | 21 October 2026

 Result: the revised teaching schedule failed on a separate three-case exercise. Complete-answer
 accuracy changed from 2/3 before training to 0/3 afterward. The no-training control remained 2/3. The
 preserved Original12 baseline remains 2/12.

    Assessment                                                            Before                    After


    Separate three-case trained arm                                       2/3                       0/3


    Matched no-training control                                           2/3                       2/3


    Original12 baseline                                                   2/12                      Unchanged; not retested here


    Qwen comparison                                                       Untested                  Untested




 What was tested
 The fixed exercise contained one multiword answer and two questions requiring a missing-information
 response. All three rows were rehearsed during training. Passing required all three complete responses,
 including their required ending. This was a training-and-retention diagnostic, not a held-out generalization test.

 The first 32 updates rehearsed the two missing-information rows. The final 32 rehearsed the positive answer.
 Each row appeared 64 times: 64 updates, 192 example presentations, 1,152 prefix tokens and 512 target
 tokens including end markers.

 The overall dose matched the earlier mixed-batch reference, while both teaching timing and batch composition changed. The result
 therefore does not isolate a pure order-only effect.


 What the measurements show
 The positive answer remained incorrect, and both previously correct missing-information responses became
 incorrect. All generated responses reached normal end markers. The recorded failures were not
 response-length-cap stops or unknown-prefix-token errors.

    Teacher-forced whole-answer loss                                      Before                    After


    Positive answer                                                       9.6923                    2.7964


    Missing-information row 1                                             0.2694                    1.8511


    Missing-information row 2                                             0.1274                    1.1721


 Lower positive-answer training loss did not produce a correct generated answer. The result supports further investigation of
 sequence fitting and interference; it does not identify a definite implementation defect.




Kira Labs | Research update | 21 October 2026                                                                                       1

How the test was run and checked
 1. Preserve the starting point
 A fresh clone of the fixed baseline checkpoint was used. Earlier failed trained states and the original baseline
 were preserved. The separate no-training arm provided a comparison for changes caused by the updates.

 2. Freeze the protocol before execution
 The training schedule, cases, scoring, update count and success criterion were fixed before the run. All
 development tracks were paused with fresh receipts. The startup requirement of at least 8 GiB available RAM
 and the 3 GiB reserve requirement were retained.

 3. Run once and retain the originals
 The launch-reference correction allowed this fresh run to reach the learner. Metadata preparation, the actual
 execution and the final memory/stop observation all returned successfully. The run made 64 training updates,
 12 objective calls and 12 generation calls, plus the 64 objective evaluations internal to the updates.

 This run completed; it was not stopped by the startup RAM guard. The result does not demonstrate that adding RAM would correct
 this learning failure.


 4. Independently audit the complete evidence
 A separate audit process checked the saved original records and recomputed the scientific result. It verified 218
 recorded Native events, 103 raw argument cells, 252 copied references, 19 canonical once-only handle
 closures, two empty process jobs and four complete owner streams with end-of-file evidence.

 These are scoped execution checks. Continuous resource coverage, active numerical-thread count and final
 operating-system-wide I/O remain unqualified or unknown. The audit did not perform new training or new model calls.


 5. Apply the stopping rule
 The required 3/3 training criterion was false. The conditional fresh-process follow-up was therefore blocked
 and remains unrun. The failed state and every original result were retained. A misleading raw status label
 referring to a pending cold audit was preserved, but it was not used to authorize another run.

 Next research and collaboration
 The immediate work is to examine matching training and generation paths, determine why the rehearsed
 sequence is not yet fitted, and define a small controlled fitting-and-retention experiment. A new result must pass
 full-answer scoring and, when eligible, a separately admitted restart test before any claim of persistent
 improvement.

 Research assistance in optimization, sequence modeling and experimental design would directly support this
 work. Additional computing resources can expand reproducible experiments; improved learning still requires
 measurement.

 Evidence identifiers: experiment 1063; independent audit 1064. Public status: NewBrain research status and method notes.
 Earlier dated reports remain historical records.




Kira Labs | Research update | 21 October 2026                                                                                     2