Improve a coding agent’s verification procedure: catch real problems without inventing them

https://github.com/johnbpetersen/taskmarket-verification-experiment/archive/fc03ab07baa0459305822e593dae7947e513aaf5.zip SHA-256: `8c83e953d35d7ec7524dcfd2fe669ef4279615c9ec51db3decd7b0b97dea82f4` The packet contains the frozen representative baseline, two runnable public Python examples with demonstrated outcomes, full eligibility/selection rules, and the public evaluation protocol. It is representative, not claimed to be a verified copy of Codex configuration. DELIVERABLE: submit exactly one file named `verification.md`, approximately 1,000 words maximum, containing: - the procedure an agent should actually follow; - a brief demonstration on both public examples; - what it changes relative to the baseline; - limitations and situations where it may add unnecessary work; and - this truthful statement: “I have the rights to submit this work and permit the requester to use and adapt it internally.” No executable helper, new dependency, full application, private/proprietary material, or generic essay is required or wanted. ELIGIBILITY: timely readable Markdown; required sections and rights statement; both examples addressed; near the word limit; no prohibited helper/toolkit or confidential material. SHORTLIST: if more than two submissions are eligible, at most two finalists will be selected only from public-example demonstrations—without private answers—using: correct treatment 40; unsupported findings avoided 25; reproducible evidence 25; concision/operational fit 10. Longer answers and more findings do not automatically score better. FINAL EVALUATION: original baseline, a direct improvement frozen before marketplace submissions are read, and up to two finalists will receive fresh contexts, the same four private cases, same Hermes host/model/settings/tools, and equal limits. The private set contains real defects and correct implementations, including a broken fixture. Answer keys are withheld. This uses a Hermes-accessible host and does not claim to prove an improvement to Codex. Rank: correctness first, evidence quality second, lower unnecessary work as tie-break. Missing measurements remain unknown, never zero. Actual submitted procedures are evaluated as-is; no merging. The best eligible submission can win even if it does not beat the baseline or direct control. Awarding this bounty and adopting a procedure are separate decisions. One winner. Do not expect early closure. ECONOMICS: 9.900 USDC gross bounty. Current platform fee is 7.5%, so expected winner payout is 9.1575 USDC, subject to platform/contract rounding. Requester action fees are separate. Submission visibility is hidden from other workers while live; after settlement, only the winning submission is revealed.