Acceptance Gating for Chained Coding-Agent Work: Measuring Completion Decay, Recovering It with Deterministic Test-Evidence Verification, and Pricing the Mutation Signal
Preprint v0.1. Empirical study of AI coding-agent contributions to open-source repositories: a deterministic CI gate verifies test-evidence quality via bidirectional validity (fail-before/pass-after) and sampled mutation checks. Evaluated on 5 real task chains (46 runs, 468 task instances) with flash and pro model tiers, a 63-task intervention study (Fisher p=4.2e-10), and a mutation-signal ablation. Companion dataset and run logs: 10.5281/zenodo.24363393.
Authors
- Kaige Yu
Institutions
- Dalian University of Technology (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24365817
- Primary Topic
- Software Engineering Research
- Type
- preprint