Naive Defect-Recidivism Mining Is Inflated by Agent Workflow Artefacts: A Construct-Validity Study at Corpus Scale

<title>Abstract</title> <p> <bold>Context</bold> . Mining version-control history for recidivism signals is a tempting shortcut for comparing AI-agent and human fix durability. <bold>Objective</bold> . We test whether such signals validly measure genuine defect recurrence. <bold>Method</bold> . We define Defect Resolution Debt (DRD) from timestamped Git/GitHub events and apply it to a 3,040-defect mined corpus (66 repositories), with confirmatory statistics on the 3,003-defect Copilot-versus-human inference subset (1,001 Copilot PR-agent, 2,002 human). We manually validate 200 flagged defects in two batches (κ = 0.654) and run a deduplication algorithm against the GitHub Issues Timeline API on 151 multi-attempt defects. <bold>Results</bold> . The naive signal is striking: Copilot fixes are flagged recidivist 1.64× more often than human ones (13.5% vs 8.2%; OR = 1.74, p = 0.002). Manual validation reverses it (precision 9.5% agent vs 41.3% human, OR = 6.69, p &lt; 0.0001; a second rater confirms). A programmatic corpus-scale correction corroborates this: collapsing 154 draft-to-final pull-request pairs reduces agent recidivism to 6.5% versus 7.4% for humans (RR = 0.88, p = 0.41). <bold>Conclusion</bold> . The naive OR survives no correction we apply; corrected estimates are indistinguishable from no difference (adversarial bound OR = 1.42, p = 0.34). A matched-pairs analysis (179 pairs) finds a significant reversal (RR = 0.48, McNemar p = 0.035), treated as hypothesis-generating: not pre-registered and contingent on an equal-recall assumption we have not validated. We contribute a corrected measurement protocol, a validated artefact taxonomy, and a deduplication algorithm. DRD's corrected precision is 32.4% (up from 24.5%). </p>