Traceability and estimate coverage as corpus-eligibility constraints: a probe of 38 Apache projects

Replication package and preprint for a probe of traceability and estimate coverage as corpus-eligibility constraints across 38 Apache projects. The study clones 38 Maven-built, Jira-tracked Apache repositories at pinned commit shas and measures two issue-commit traceability rates per project: the commit-side rate that the literature publishes, and the ticket realisation rate that a ticket-anchored study actually depends on. Twelve projects clear a commit-side threshold of 0.80 that was fixed in predictions/PREDICTIONS.md at commit ca076a9 before any project was cloned, and never moved. The measure is validated against an independently constructed dataset: four of five projects overlapping SEOSS 33 agree within 3.3pp on corpora seven years apart, and the fifth agrees to +0.0041pp once commit scope is matched. A 200-commit seeded manual sample puts matcher precision at 195 of 200. Contents. Per-project measurements and the full 38-row eligibility table; the three-channel visibility table; every script that produced a number in the paper; the pinned repository shas; 2,491 frozen Jira cache files with a per-file SHA-256 manifest; the pre-registration file and its commit; the held-out and replication records; dated working logs left unrewritten; and an adversarial audit report, included unedited. The paper is included as a PDF. On one deliberate disagreement in the results. The frozen and live estimators of the commit-side/ticket-side association differ in sign: Spearman rho = -0.010 across 33 projects under the frozen snapshot, +0.518 across the twelve measured exactly and live. This is reported rather than resolved, and paper/DIRECTION_TENSION.md sets out both measurements, the populations they are taken over, and the decomposition that explains the difference. It is a documented tension, not an inconsistency. Six operationalisations of the original research question were tested and retracted with named causes; those retractions are recorded in the package rather than removed from it. Personal data. Developer email addresses appearing in Apache commit metadata have been replaced with salted pseudonyms, following the practice of SEOSS 33. Commit hashes are unmodified, so every determination remains checkable against the public repositories. Use of AI tools. Large language model assistants (Anthropic's Claude) were used in this work: to draft and run the analysis scripts, all of which are committed and re-runnable; to draft manuscript prose from the author's outlines, findings and decisions, revised by the author; to check citations against source texts; and to conduct an adversarial audit that re-derived the load-bearing results from primary artifacts with separately written code, and reported errors that were then corrected. The research questions, the study design, the pre-registered threshold and its commitment before any data was seen, the decision to hold that threshold when moving it would have helped, the six retractions, and the interpretation of the results are the author's. The author takes full responsibility for all content, including any error a tool introduced and the author did not catch. Section VIII-C of the paper states this in full, and paper/DISCLOSURE_VERIFICATION.md separates what is observable in the repository's record from what is inferred about sessions predating it.