Kira Labs / Source-report reading edition
Report 61 · Sampled Gradient Test
Document: 61_NewBrain_Sampled_Gradient_Pass_and_Test_Method_2026-10-19.pdf. This reading edition reproduces text extracted from the original report supplied for publication, separated by source page. It is not a newly verified experiment, and PDF layout, line wrapping and any diagrams may differ. For a test's current status, consult the linked Kira Labs evidence page and dated research updates.
Transcribed source page 1 of 4
KIRA LABS / RESEARCH EVIDENCE UPDATE
A verified pass for
sampled learning gradients
A new diagnostic milestone, with the learning score kept separate.
80 4 2
Coordinate comparisons Directional comparisons Deliberately wrong-gradient
passed passed comparisons rejected
What was measured
The fixed numerical check passed across two declared probe states. No sampled derivative mismatch or
instability between the two perturbation sizes was found. All ten parameter groups met the unchanged
minimum-signal requirement. The largest absolute comparison error was 2.08e-9.
A separate checking role recomputed the comparisons from the complete saved objective measurements. The
saved execution and file-evidence audit also passed. The original checkpoint parameters and all four recorded
training counters were preserved.
Current learning results
Separate experiment Before After Interpretation
Earlier 64-update study 2/12 0/12 Negative trained checkpoint preserved.
Latest 256-update study 2/12 2/12 Two new passes and two lost missing-information
passes; no net gain.
The numerical check performed no training updates, generation or scoring. It does not raise the 12-case
learning score. Qwen has not yet been tested on the comparison. The original baseline remains preserved.
Earlier dated papers remain research snapshots. Read this note with the current learning report (PDF58) and the separate
four-response retention pilot (PDF59).
Kira Labs | NewBrain research evidence | 19 October 2026 1Transcribed source page 2 of 4
HOW THE AUTOMATED TEST WORKS From a frozen plan to independently checked numbers 1. Freeze the inputs and acceptance rules Before execution, the plan fixed three public synthetic examples, two probe states, twenty coordinates across ten parameter groups, two perturbation sizes, one deterministic direction and the rejection control. One probe used the inherited checkpoint; the other used an explicitly declared synthetic parameter recipe. These are diagnostic probes, not a new generalization benchmark. 2. Check resources and run under controlled ownership Parallel work and the reminder were paused. Fresh launch checks used the existing 8 GiB available-memory startup guard and 3 GiB reserve. The startup sample showed about 13.4 GiB available. Controlled processes recorded their results, bounded outputs and cleanup evidence. The guard values were not lowered. 3. Compare analytical and numerical derivatives For a parameter coordinate, the numerical estimate is: (loss(theta + h) - loss(theta - h)) / (2 x h) The two fixed h values were 1e-4 and 1e-5. The acceptance bound was absolute error <= max(1e-8, 0.001 x the larger derivative magnitude). Each coordinate was also checked for agreement across the two h values. A deterministic direction tested a combined change across the parameter vector. 4. Require signal and deliberately test rejection Every parameter group had to meet the predeclared 1e-6 minimum signal. The tested Q, K and V maxima were about 5.04e-6, 9.68e-6 and 5.15e-6. An intentionally incorrect analytical gradient was then checked against both fixed perturbation sizes; both comparisons were rejected. This control reused saved measurements and added no model calls. 5. Preserve originals and check the evidence separately The run recorded 170 objective evaluations. The independent saved-data audit checked 219 execution events, 106 full raw data cells, 253 copied references, 19 one-time resource closes, two empty jobs and four complete output streams. The checking stage made no new model or execution-interface calls. Kira Labs | NewBrain research evidence | 19 October 2026 2
Transcribed source page 3 of 4
INTERPRETATION / NEXT EXPERIMENT
What this helps us
investigate next
The supported conclusion
The sampled raw derivatives of the specified loss are numerically consistent under this fixed test. This
makes an error in those sampled derivative calculations less likely as the explanation for the learning failures. It
does not establish complete gradient coverage, optimizer correctness, independent reconstruction of the model
state, useful new learning or unfamiliar-example generalization.
Why the earlier record remains important
The earlier numerical study remains inconclusive: its Q/K/V samples did not meet the fixed signal floor, and
final evidence publication failed. The completed follow-up used a stronger, predeclared synthetic route gain to
make the check informative. A separate launch attempt stopped on a metadata schema mismatch before model
execution; that mismatch was corrected. The failed originals and their costs remain recorded.
The next learning experiment
Question Planned controlled check
Can it acquire a new association? Compare a taught learner with a separate zero-update, no-teach control from
the same preserved starting state.
Can it retain earlier behavior? Check the new association alongside two missing-information responses,
using fixed questions and complete output sequences.
Does it survive restart? Save the resulting state, start a separate process, and assess the same fixed
outputs again.
Can a larger claim be made? Treat this small test as a mechanism diagnostic. Reassess the larger
benchmark only through a separately defined, preserved experiment.
This acquisition-and-retention integration is in development and remains unrun. Evelyn's separate three-visitor memory test also
remains unrun; saved conversation memory is distinct from changes to neural weights.
Kira Labs | NewBrain research evidence | 19 October 2026 3Transcribed source page 4 of 4
COLLABORATION / NEXT RESEARCH
Where a university can make a
difference
A concrete university collaboration
The next useful partnership combines expertise in optimization, sequence generation, memory interference and
experimental design. Lab computing and research assistants would support controlled repetitions, independent
replication and careful analysis. More memory or a larger GPU would expand the experiments that can be
attempted; equipment alone does not guarantee better learning.
Research area Useful contribution
Learning and optimization Help design acquisition, retention, sequence-generation and forgetting
experiments with independent evaluation.
Systems and computing Run controlled repetitions on university hardware, review resource use and
support reproducible research workflows.
Human-computer interaction Evaluate Evelyn's conversation continuity, names/details across restarts,
visitor separation, voice and usability.
Avatar and biomedical research Study measured deformation and anatomy. Validated organ physiology and
biological nerve-to-brain behavior remain future work.
Evelyn and the avatar remain separate evidence tracks
Evelyn's frozen 14-turn test involves three synthetic visitors, delayed recall, application restart, separation of
visitor details and a corrected preference. It is unrun. Her head, voice and installer still require runtime testing
and private user installation before public release. The existing body result is limited to 512 weighted skin
vertices moving through five commanded poses and returning to their starting positions.
Evidence and review scope
The result was checked by separate internal roles and automated audits. External peer review and independent university
replication remain pending. Current public status: NewBrain research record and KiraWorld research record. Detailed restricted
evidence is retained locally for authorized technical review.
Kira Labs | NewBrain research evidence | 19 October 2026 4