[devin-ai-integration]

Summary

The Weight-Space Programming paper improves Qwen3.5-0.8B mainly by suppressing generation loops. Kev is a prefill-only decision model built on Qwen-Base. This PR tests both the paper's published donor edit and a second, Kev-specific edit of its decision head. Neither clears its preregistered quality gates; keep the released checkpoint unchanged.

Two hypotheses, not a direction search

Round 30: transfer the verified donor delta. Reconstruct the sparse artifact, verify original/edited tensor hashes, and add dose * (edited_donor - original_donor) to Kev's two base down_proj matrices. Copying replacement weights would incorrectly overwrite Kev-Base. Full/half doses leave the adapter and head untouched. Full-dose transfer accuracy changes −0.152 pp, paired 95% CI [−0.457, 0.000]; Brier gets slightly worse. R30 report.

Round 31: reweight Kev's own pointer channels. The pointer head scores each answer by summing 256 products of projected decision/option features. Unequal channel gains can change rankings, unlike scalar temperature. Fit 255 effective gains in one damped analytic Newton solve on 1,536 training records / 2,177 questions, then fold them into head.q.weight and its bias. Gains have mean one and deviate at most ±25%; backbone, LoRA and key projection remain frozen. This is supervised analytic fitting, not a data-free or literally derivative-free method; no iterative backpropagation is used. A seed-0 coordinate permutation preserves the gains' distribution as an attribution control. Registration precedes inference; evaluation records/groups/exact states are excluded from fitting. R31 report.

Kev-specific edit results

Kev bf75a6a8848ea6960ff2ed108d9ed44c2941174f, H100, development partitions only, unchanged shipped T=2.3510958125672174. Clean knowable questions:

Panel	n	Baseline accuracy	Learned	Half	Shuffled
transfer-v4	656	64.787%	64.787%	64.787%	64.787%
hard-v1	1,083	59.372%	59.464%	59.464%	59.464%
documents-v1	920	84.239%	84.239%	84.239%	84.239%

Hard accuracy improves only +0.092 pp [0.000, +0.279]; learned-minus-shuffled accuracy is 0 [0, 0]. The accuracy and attribution gates fail. Hard Brier improves 0.530870 → 0.529890 and NLL 0.973454 → 0.971386; documents also improves slightly. These probability gains are measurable, but the control improves them too. ECE improvements are not separated from zero. We do not replace the primary with a favorable secondary after seeing results.

Raw rows, reports, fit selection and small gain/statistic tensors are committed; no model checkpoint is published. Both canonical diagnostic readouts select no candidate. Combined metered app compute $2.24694 ($1.21240 + $1.03453), below the $10 cap; storage/egress separate. No locked-test reads, Hub writes, serving changes, release selection or post-result search. These negative pilots do not rule out every possible Kev-specific weight edit.

Round 32: registered, not yet executed

The user approved a different $5 pilot: a short LoRA continuation on teacher-verified training errors, compared with unchanged Kev and an ordinary-training continuation. error_repair_pilot.py reserves disjoint mining/control pools before reading predictions, asks open-weight DeepSeek-V4-Pro to solve the errors without seeing their labels, and retains only answers that agree with programmatic gold. Both continuations use 128 hard examples matched by source, identical 128-record replay, and 32 optimizer steps. This tests the error-selection plus teacher-verification package, not teacher-generated labels against gold; accepted labels are identical to solver gold.

The preregistration requires targeted hard accuracy to beat both baseline and ordinary continuation, with frozen-temperature calibration and transfer guards. API calls reserve cost before dispatch, never retry ambiguous charges, and stop at $1; two bounded GPU calls leave the total below $5. No new paid calls have run: teacher API access is pending. This section is preparation, not a quality result or a release candidate.

Test plan

• R32 local suite: 644 passed, 26 archive-only skips; Ruff/Pyright and structural registration preflight pass. Cost reservations, blind prompts, teacher agreement/failure stops, source matching and train/eval disjointness tested without live API calls.
• R32 live teacher calls, GPU training/evaluation, baseline parity and full artifact validation: blocked on API access. No GPU quality claim yet.
• Repository fast suites: 623 passed, 26 archive-only skips before the final saved-evidence test; final focused suites: 50 passed.
• Analytic gradient/Hessian agrees with autograd on synthetic data; folded weights match gain-adjusted logits; bounded mutation and exact restoration tested.
• All four fresh R31 baseline row files are byte-identical to R30. Complete coverage, no truncation/rejection, all packing/sibling checks pass.
• R31 changes only query weight/bias; R30 only the intended two matrices. All other parameter hashes and exact restoration checked on H100.
• Saved fit reproduces within 1e-12 on another CPU; all 36 R31 baseline bootstraps plus control comparison, and all 24 R30 comparisons reproduce from committed rows.
• Ruff, Pyright, round validation and diff checks pass. Metric reproduction tolerates only 1e-12 CPU reduction variance, not changed decisions.
• Thermonuclear self-review: canonical loader/evaluator/metrics/round helpers reused; no shared inference/training code changed; no new file over 1,000 lines.

Link to Devin session: https://app.devin.ai/sessions/f7a092ee39c7463f8469ab4f3229f5bd
Open in Devin Desktop: https://app.devin.ai/desktop/session/f7a092ee39c7463f8469ab4f3229f5bd?variant=devin
Requested by: @jaredpalmer

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring