Anonymous research manuscript on verifier-gated repair and checker-objective alignment at test time.
Repair When It Helps: Auditing Checker-Objective Alignment at Test Time
- Manuscript: vgr-tl-anonymous.pdf
- Status: Anonymous research submission
Runtime-visible checkers can detect contract violations, but a checker verdict does not necessarily match the final task objective. This work frames the decision to spend one additional model call on repair as a checker-objective alignment problem.
It introduces VGR-TL, a bounded verifier-gated repair policy with a trace ledger that preserves accepted candidates, repairs rejected candidates with permitted visible feedback, and records provenance, candidate transitions, failures, latency, and cumulative token use. The manuscript evaluates the accuracy-cost-damage trade-off across instruction following, function calling, JSON-schema generation, and code tasks.
- vgr-tl-anonymous.pdf: Anonymous manuscript.
- README.md: Publication overview.
This repository is maintained as an anonymous research release. It intentionally contains no identifying author, affiliation, or contact information.
Citation metadata will be provided after the anonymous review process is complete.