Verifier card — aime2025 · grade A
Clean — no reward-hack signature found; the reward held against the full battery below.
- source:
prime-hub· grader:deterministic*· verifiers:0.1.15.dev187· pass_threshold:0.5· rows: 2 - seal:
record_sha256:5b465b718926fc45— re-running this audit reproduces it byte-for-byte
What we tested
- Degenerate-probe — 16 content-free completions: all below pass_threshold ✓
- IPT invariance — n/a (no free-text reference completion to perturb)
- Verifier-completeness — abstained (no usable scalar ground-truth answer)
- Exploit-search — searched the reward space; no accepted non-attempt ✓
What we found
- No reward-hack signature. Every degenerate probe scored below pass_threshold (0.5); the reward was invariant under cosmetic re-rendering; and the exploit-search (where run) found no accepted non-attempt. A defended reward.
Reproduce
stardata audit aime2025 --row-seed 0
Coverage & limitations
- grader class audited:
deterministic(inferred from the replayable reward) - instruments run: degenerate-probe, IPT invariance, verifier-completeness, exploit-search
- scope: an environment-side audit of reward gameability — it does not guarantee your policy won't find other optimizations, nor does it assess task difficulty (a separate difficulty certificate).