Rebuild Benchmark Audit

Audit and construct verifier-graded "rebuild the program" benchmarks and RLVR reward environments — hunting duplicate reward, shortcut-passable tests, ambient implementations already installed in the task image, proxy/short-circuit assertions, wall-clock time bombs, unreachable tests with missing fixtures, and grader-only dependencies. Use this whenever the user mentions ProgramBench, SWE-style or agentic benchmark construction, cleanroom task images, gold/dummy test validation, behavioral test suites as reward, verifier design, reward hacking, test-suite deduplication, or asks why a model scores suspiciously high or low on a coding benchmark — even if they don't use the word "benchmark". Also use it when designing the same recipe for APIs, libraries, protocol daemons, or data pipelines.

daedalus Updated 1 repo stars

File contents

daedalus/skills commit 479b9e711e

Frequently asked questions

npx get-skill add daedalus/rebuild-benchmark-audit