CodeTrace-Agent Benchmark Lifecycle Reliability Research Compendium

This research compendium provides the code, frozen protocols, de-identified task-level evidence, machine-readable results, executed notebook, and editable figures supporting the article Auditing the Lifecycle Reliability of a Repository-Level Coding-Agent Benchmark: A Multi-Checkpoint Empirical Study. The package operationalizes a claim-evidence contract across construction and replay, cross-host transfer, verifier adequacy, process evidence, selection provenance, and external applicability. Version 1.0.1 adds the completed, separately reported four-task second-designer behavioural-challenge supplement. All four constructions were executable and all four escaped the registered verifier; the 2/2 versus 2/2 result does not establish concern-control discrimination. Registered-verifier reward is not treated as semantic correctness, and the package does not produce a single global reliability score. Reviewer identities, access tokens, private reviewer worksheets, copied upstream repository snapshots, raw model dialogues, machine-local receipts, and manuscript files are excluded. File-level licensing is documented in DATA-LICENSE.txt: team-created data, annotations, results, metadata, documentation, and figure sources are CC BY 4.0; analysis code is MIT-licensed.