Reproducibility deposit — Exposure is not manifestation: measurement target and output resolution jointly determine which behavioural-faithfulness evaluator wins

Reproducibility deposit for the single-author manuscript "Exposure is not manifestation: measurement target and output resolution jointly determine which behavioural-faithfulness evaluator wins". Two measurement targets are scored on the identical 720 replies: exposure (was the reply produced under a trait-inducing prompt) and manifestation (did the trait actually surface). Two instruments are compared at three output resolutions; the estimand interaction excludes zero at all three (+0.207, +0.237, +0.169 AUROC), while the ranking reverses only under the deployed verdict interface. Contents: the 720-reply paired evaluation set with both labels; per-item scores for all three instrument configurations (including the auditor per-item out-of-fold read-out); the 90,314-record held-out set behind Table 1; the human-rating raw files, canonical pipeline and vote matrix (Fleiss kappa 0.130 primary / 0.074 sensitivity); geometry comparison; analysis scripts; label and source provenance manifests; environment and requirements files; SHA256SUMS and a role-tagged manifest. Reproduce: python CIHTD_code/reproduce_table2.py (Tables 2-3 incl. exact McNemar p and all cluster bootstrap intervals), python CIHTD_code/make_sr_figures.py (Figures 1-3), python human_protocol/human_kappa.py (reported kappas from raw ratings, ID-joined). The auditor checkpoint and pinned source archive are held in a private model repository; reviewers are granted access on request during peer review. Changelog v4: single author, new title; analysis re-centred on the estimand-by-output-resolution design; added per-item auditor scores, the three-command analysis pipeline, exact McNemar p, pair-cluster bootstrap statistics, provenance manifests (LABEL_PROVENANCE_v4, SOURCE_MANIFEST_v4, CANONICAL_FILES_v4), and quarantined superseded materials under historical/ folders. Earlier versions carry the previous title, a four-author list, and a single-target analysis, and remain available for the record.