AuditCoder Makes the Construction Trace a First-Class Output, But Pays 7.5-8.5 Points of pass@1 For It
AuditCoder emits an auditable construction trace alongside the program, using a contract-annotated task graph that binds stable responsibility identities to each commitment, its implementation, provenance, validation evidence, and intervention history; on validation failure a conservative locator maps evidence to a node or dependency branch — or abstains — and bounded repair regenerates only that region while reusing the frozen complement. It reaches 82.5-83.0% pass@1 on APPS and 75.0-82.0% on ClassEval, beating CoT+retry but trailing AgentCoder by 7.5-8.5 points. An audit of 200 APPS records yields 0.9725 task-macro decision-code trace coverage, though the locator only localizes 26 of 60 failures, 17 of which repair successfully.
↳ Follow the thread