Replication package: What AI Coding Agents Claim and What We Can Measure (IntentValidator, NFR claim catalog, two-rater manual validation)

Replication package for the manuscript What AI Coding Agents Claim and What We Can Measure: Non-Functional Requirement Claims, Extraction Validity, and Intent–Outcome Consistency in Agent-Authored Pull Requests, prepared for submission to Empirical Software Engineering (special issue: Agentic Software Engineering). Contents: the IntentValidator pipeline source snapshot and run configurations; the per-claim datasets of the canonical agent run, of the all-repository human comparison cohort (with its protocol, draw listings and order of events) and of the superseded first human cohort, plus the re-runs that document the effect of the pipeline's timeouts; the 186-pattern NFR claim catalog, boilerplate-stripping rules and the full match table; the extraction-validity ladder and regression scripts; the manual-validation kit (stratified samples, rubric, blind labelling tool, both human raters' sheets with untouched exports and SHA-256 sums, three disclosed LLM annotation passes with per-item rationales and provenance, the LLM-judge sensitivity, the adjudication queue, a recall probe of unflagged pull requests with blind sheets and a disclosed LLM screen, and a kit for a rater from outside the author team); the consistency analysis on the human-validated claims, the Maintainability-Index floor sensitivity and the all-language cross-check of the index against independent Halstead implementations (per-claim results); every analysis script that regenerates the manuscript's numbers and figures; and a checksum manifest of every file with named reproduction checks. The archive carries the replication tree only; the manuscript itself is not included. The AIDev dataset itself is not redistributed; obtain it from its own Zenodo record (10.5281/zenodo.16919272). The 25 September 2026 audit revision adds explicit sampling and interpretation limitations, corrected stratification of diagnostic agreement intervals, clarified nominal instrument coverage, and reproducibility notes. Original human labels and canonical claim verdicts are preserved; catalogue label/rubric conflicts remain for human reassessment. See replication/09-paper/emse-2026/replication/REVIEW-NOTES-2026-09-25.md.