{ "asset": "B5 measured on vendor-model output: the k-sample judge vs the certificate witness on the 24 committed GLM-5.2 Heartbleed guards", "source": "artifacts/crs/gameability_census_summary.json :: candidates (GLM-5.2, z.ai; all 24 pass the finite battery)", "domain_size": 281474976710656, "n_candidates": 24, "n_unsound": 10, "n_sound": 14, "judge": "uniform k-sample over (p in [0,2^16), r in [0,2^32)); accepts unless a sample is guard-accepted and out of bounds; seed 20260902+k", "witness": "formal.decision_procedure._implies(domain_full + guard, in_bounds) -> Farkas certificate; accept iff certified", "judge_vs_witness": [ { "judge_sample_budget_k": 8, "n_unsound_guards": 10, "trials_per_guard": 400, "false_accept_rate_measured": 1.0, "false_accept_rate_closed_form": 1.0, "witness_checker_false_accept": 0.0 }, { "judge_sample_budget_k": 32, "n_unsound_guards": 10, "trials_per_guard": 400, "false_accept_rate_measured": 1.0, "false_accept_rate_closed_form": 1.0, "witness_checker_false_accept": 0.0 }, { "judge_sample_budget_k": 128, "n_unsound_guards": 10, "trials_per_guard": 400, "false_accept_rate_measured": 1.0, "false_accept_rate_closed_form": 1.0, "witness_checker_false_accept": 0.0 }, { "judge_sample_budget_k": 512, "n_unsound_guards": 10, "trials_per_guard": 400, "false_accept_rate_measured": 1.0, "false_accept_rate_closed_form": 1.0, "witness_checker_false_accept": 0.0 } ], "witness_false_accept": 0, "witness_false_reject": 0, "control": { "canonical_guard_certified": true, "honest_certificate_recheck": true, "corrupted_certificate_recheck": false, "control_passed": true }, "rows": [ { "overhead": 0, "slack": 0, "op": "le", "over_acceptance_exact": 196608, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 0, "slack": 0, "op": "lt", "over_acceptance_exact": 131072, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 1, "slack": 0, "op": "le", "over_acceptance_exact": 131072, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 1, "slack": 1, "op": "le", "over_acceptance_exact": 196608, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 2, "slack": 0, "op": "le", "over_acceptance_exact": 65536, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 2, "slack": 0, "op": "lt", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 2, "slack": 1, "op": "le", "over_acceptance_exact": 131072, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 2, "slack": 1, "op": "lt", "over_acceptance_exact": 65536, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 3, "slack": 0, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 3, "slack": 0, "op": "lt", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 3, "slack": 1, "op": "le", "over_acceptance_exact": 65536, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 3, "slack": 1, "op": "lt", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 3, "slack": 2, "op": "le", "over_acceptance_exact": 131072, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 4, "slack": 0, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 4, "slack": 1, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 4, "slack": 1, "op": "lt", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 4, "slack": 2, "op": "le", "over_acceptance_exact": 65536, "unsound": true, "witness_accepts": false, "n_farkas_certs": 0 }, { "overhead": 4, "slack": 2, "op": "lt", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 5, "slack": 0, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 5, "slack": 1, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 5, "slack": 2, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 16, "slack": 0, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 19, "slack": 0, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 }, { "overhead": 19, "slack": 1, "op": "le", "over_acceptance_exact": 0, "unsound": false, "witness_accepts": true, "n_farkas_certs": 1 } ], "honest_scope": "One class (Heartbleed) and one vendor model's 24 candidates; the judge is a uniform sampler, the shape of tests+fuzzing acceptance, not a specific product. Over-acceptance regions here are 65,536-196,608 inputs of 2^48, so any feasible k leaves the judge blind; the closed form (1-m/N)^k is what the measurement must match. The witness is the repo's own Farkas certificate re-checked by the same arithmetic the 228-line anchor uses.", "seal": "artifacts/seals/glm_inference_floor_2026-09.seal.json", "decider_integer_tightening": true, "class_verdicts_under_this_decider": { "openssl_heartbeat": { "sound": true, "complete": true, "tight": false }, "libxml2_len_overflow": { "sound": true, "complete": true, "tight": true }, "zlib_inflate_extra": { "sound": true, "complete": true, "tight": true }, "libpng_rfc1123": { "sound": true, "complete": true, "tight": true }, "sudo_set_cmnd": { "sound": true, "complete": true, "tight": true }, "curl_ntlm_type3": { "sound": true, "complete": true, "tight": true } } }