Verification strength does not buy discovery quality: a pre-registered, two-domain, hundred-pair test of verifier-bound automated research — Artifact Record

Abstract

Artifact record for the manuscript "Verification strength does not buy discovery quality: a pre-registered, two-domain, hundred-pair test of verifier-bound automated research". Contents: frozen pre-registration protocol (v1.0-v1.2 with changelogs), per-pair JSONL ledgers for two domains (job-shop scheduling, n=100 pairs; vehicular edge offloading on Beijing T-Drive, n=100 pairs), mutation-battery populations and scores, analysis scripts generating every number in the paper, exploratory pre-work probes, and the submitted manuscript. All quantitative results are generated by the released analysis scripts; none are hand-computed. Seed-to-figure reproduction: rerunning the scripts on the released ledgers reproduces every number and figure. Provenance: holdout instances evaluated exactly once per promoted rule (HoldoutGate, one-shot semantics). Three domain-D pairs (46, 71, 98) lack an S+ or S- promotion; the paired group contrast there is computed on the 97 valid pairs. Union-pool and tension-gradient analyses are exploratory (post-hoc) and labeled as such in the manuscript. License: code MIT; data and protocol CC-BY-4.0.