{
 "schema": "mimese.audit.index/1",
 "updated_at": "2026-10-13T09:01:47+00:00",
 "reports": [
  {
   "id": "selftest",
   "file": "selftest.json",
   "title": "Audit self-test: 14 planted flaws, 14 clean verifiers",
   "kind": "self-test",
   "created_at": "2026-10-13T09:01:47+00:00",
   "summary": "Self-test: 14 of 14 planted flaws caught by the check designed for them; 0 false positives on 14 clean verifiers. Acceptance (>= 12/14 caught, 0 false positives): passed.",
   "targets": 28,
   "worst": "S3"
  },
  {
   "id": "math-verify-0.9.0-symbolic",
   "file": "math-verify-0.9.0-symbolic.json",
   "title": "Math-Verify 0.9.0 on symbolic answers",
   "kind": "public",
   "created_at": "2026-10-13T05:43:41+00:00",
   "summary": "Math-Verify 0.9.0 with the gold written as $answer$, on 64 symbolic tasks (16 each: fractions, radicals, intervals, polynomial expressions): no findings in 37 probes. All 209 equivalent forms of the correct answers earned full reward (slash fractions, terminating decimals, \\dfrac, 'x = …', \\left( \\right], 'x \\in …', reordered unions and terms, partly factored polynomials), and every near miss earned nothing (2-decimal roundings, flipped brackets, shifted endpoints, x^2 → x^3, dropped roots). In our run the 101-box spray hit Math-Verify's 5-second comparison timeout on 6 of 64 tasks; it was rejected there too.",
   "targets": 1,
   "worst": "S0"
  },
  {
   "id": "math-verify-0.9.0",
   "file": "math-verify-0.9.0.json",
   "title": "Math-Verify 0.9.0: answer-agnostic reward-hacking probes",
   "kind": "public",
   "created_at": "2026-10-11T05:33:13+00:00",
   "summary": "Math-Verify 0.9.0 earned no reward from any of 30 probes in either configuration, on 60 integer-answer arithmetic tasks: 7 null inputs, 6 kinds of wrong-answer mutation, 12 adversarial inputs (answer sprays, hedges, infinity and NaN, grader and JSON injections, padding, malformed LaTeX) and 4 exploits proposed by an LLM attacker, which are saved so the audit can be replayed. The reference solution earned full reward every time. Scope: integer answers only; symbolic, multi-part and free-form answers are not covered by this report.",
   "targets": 2,
   "worst": "S0"
  }
 ]
}