October 20, 2026 · Source research report 72
Where the longer answer fails
A read-only diagnostic analyzed the saved failed longer-answer state and identified two incorrect sequence decisions. No teaching updates or learning gain occurred.
These are the full text pages extracted from 72_NewBrain_Sequence_Learning_Diagnostic_and_Method_2026-10-20.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.
Source page 1 of 2
KIRA LABS / RESEARCH EVIDENCE
NewBrain sequence-learning diagnostic October 20, 2026
Where the longer answer fails
A completed read-only diagnostic localized two weak decisions in a separate three-case teaching exercise. It did not
train the model, change a checkpoint, or rerun the larger assessment.
Original12 New training updates Qwen comparison
2/12, unchanged 0 in this diagnostic Still untested
What was measured
Decision inspected Required symbol Model's choice
Immediately after the correct first answer word Next word End marker
4.38% probability 4.77% probability
At the next position, with the correct preceding word Following word A different word
supplied 3.62% probability 16.97% probability
Supplying correct preceding words is an inspection technique, often called teacher forcing. It reveals a decision at that history; it does not
establish that the model generated the whole history unaided.
What this tells us
The first suffix decision narrowly favors ending the answer. A later word decision is also incorrect even with the
correct preceding word supplied. Simply preventing an early end would therefore leave another demonstrated
sequence error.
The two missing-information controls still strongly favor their expected tokens at the fixed diagnostic positions. These
observations narrow the next learning experiment without proving a particular implementation defect.
What remains unresolved
This is a three-case diagnostic, separate from Original12. It does not by itself explain every error in the larger
12-case study, establish generalization, or demonstrate a new learning gain. The saved negative experiments and
the original baseline remain preserved.
Next step
Prepare one separately defined teaching-order comparison at the same dose, with a preserved baseline, an
untreated control, both missing-information guards, and a fresh-process reload after a joint pass. The stopped failed
experiment will not receive further updates. This next experiment is preparation, not a successful result.
Synthetic research exercise | Preserved original evidence | No clinical claim Page 1Source page 2 of 2
KIRA LABS / RESEARCH EVIDENCE
NewBrain sequence-learning diagnostic October 20, 2026
How the test was built and checked
Scope: one preserved failed state; three fixed synthetic examples; eight fixed answer positions. One example has a
longer structured answer and two are missing-information controls.
1. Freeze the inputs and plan
The checkpoint, protocol metadata, source versions, fixed positions and limits were pinned before execution. The
diagnostic used the preserved failed state as read-only input. Training, free generation and new model-state products
were prohibited.
2. Run once with current gates
Parallel tracks were held. Current thread, paused reminder, source closure and fresh test receipts were checked.
Available RAM was 16.8 GiB before the run, above the unchanged 8 GiB startup guard; the 3 GiB reserve guard and
existing process limits remained in force.
3. Inspect the decisions without further teaching
One native objective batch produced eight fixed gold-history readouts. The complete loss and output probabilities,
logits and output-bias gradients were retained for independent reconstruction. No result-based retry was used.
4. Recompute the arithmetic independently
Check Verified result
Normalized loss 1.0836942996357317; reconstruction difference 0
Output-bias gradient All 42 coordinates checked; maximum absolute difference 1.39e-17
First suffix logit margin Required word minus end marker: -0.0854257
The first suffix margin equals a shared output-bias difference of -0.6886893 plus a residual of +0.6032636 from all other logit terms. This
is descriptive algebra, not proof that output bias caused the failure. The check does not newly validate all recurrent-gradient coordinates.
5. Audit originals, cleanup and costs
The independent saved-evidence review checked 219 native events, 106 raw argument cells by inverse hash, 19
canonical once-closes, four complete owner streams and two empty process Jobs. Available original file-I/O
evidence was closed. Full selected allowances, including unused portions, were permanently closed.
The audit used saved evidence only: no new model, provider or native-execution calls. Source-exported I/O counters are not complete
operating-system-wide totals. Continuous RAM/CPU behavior and a causal defect remain unproven.
Evidence: consumed execution 1050 and independent saved-evidence audit 1051. Public methods and current status. Complete private
originals and failed checkpoints are retained.
Synthetic research exercise | Preserved original evidence | No clinical claim Page 2