test(agent): compare verification quality and long-command guidance

[minixalpha]

This experiment tests whether stronger headless verification guidance improves evidence quality, and compares it with a separate long-command guidance candidate. It keeps runtime tools and budgets unchanged, freezes all three source revisions, and completes all 24 registered trials without replacement runs.

The shared valid comparison is control 5/6, verification 6/6, operations 6/6. Both candidates gain the first scheduler repeat, whose control checks each output separately but misses the public cross-file shape limit. All arms pass 4/4 cases on the two tasks outside prior tuning. Two pipeline controls fail during verifier dependency setup before tests start; their native zeros remain recorded and are excluded from capability comparisons. This is an exploratory result, and the PR remains a draft rather than a recommendation to promote the prompt.

Task, repeat 1 / repeat 2	Control	Verification	Operations
Pipeline parallelism	Verifier error / Verifier error	Fail / Pass	Fail / Fail
Inference batching scheduler	Fail / Pass	Pass / Pass	Pass / Pass
Build Cython extensions	Pass / Pass	Pass / Pass	Pass / Pass
Cancel async tasks	Pass / Pass	Pass / Pass	Pass / Pass

The verification arm performs five negative controls of varying strength, including bad async implementations and malformed scheduler plans. It still drops a failing tied-weight pipeline case, and uses 9.1% more agent time on common valid cases. Operations incurs three Bash timeouts versus one in control and overstates a failed numerical self-check. Three trials have incomplete cost accounting, retained as known lower bounds. The report recommends requirement-to-evidence coverage, preserving failed cases, final-artifact checks, managed command execution and reliable verifier preflight for the next experiment.

The branch includes the verification prompt candidate, frozen comparison plan, reusable standard-runner orchestration, a collector that verifies recorded artifacts, complete structured results, and a manual unblinded 24-trial audit. Candidate runtime sources remain unchanged after registration. No model or official verifier was rerun during the final audit.

• Final report
• Registered plan and reproduction
• Results and accounting

Validation: 105 Harbor tests passed, including collection tests for native-zero preservation, partial accounting, eligible pairing and changed evidence. The frozen prompt/runtime previously passed 351 core tests, 4 skipped. All 24 ATIF trajectories and installed wheel/source manifests validate; frozen task/config/prompt hashes and four Docker image IDs match. Verifier retries retain native limits and occur only after timeouts, at most three attempts. Agent installation preflight passed, while all four verifier dependency profiles remain explicitly not_configured.

Limitations: two repeats, two tasks outside prior tuning, manual unblinded audit, prior pipeline evaluation exposure, provider/network variation and invalid pipeline controls. No confirmed external benchmark-answer exposure was found in the 24 traces. Research notes in both languages are unchanged.