Receipt-Bound Delayed-Contradiction Benchmark: Reviewer Artifact V6

This reviewer artifact accompanies the manuscript "When Does a Correction Change the Action? A Receipt-Bound Matched Benchmark of Delayed Contradiction in Language Agents." It releases 20 selected synthetic delayed-contradiction tasks, exact answer labels, terminal verifier specifications, 480 execution-unit bindings, redacted completed-stage outputs, 840 redacted attempt receipts, task- and arm-level result tables, frozen validation receipts, and standard-library-only reconstruction scripts. The artifact deterministically reconstructs all 480 terminal scores, all eight configuration means, and 60 primary matched-pair effects. It supports reviewer inspection and exact table reconstruction; it is not a fresh provider rerun, expert validation, human ground truth, deployment evidence, or a universal method-superiority claim. API credentials, provider accounts, provider request identifiers, invoices, raw private request payloads, and unselected holdout tasks are excluded.