{ "integrity_gates": { "G1": { "gate": "G1", "status": "FAIL", "threshold_pp": 0.1, "n_exact_match": true, "n_known_exact_match": true, "n_unknown_exact_match": true, "deltas": { "n": { "committed": 3369, "rerun": 3369, "abs_delta": 0 }, "n_unknown_labeled": { "committed": 1032, "rerun": 1032, "abs_delta": 0 }, "n_known_labeled": { "committed": 2337, "rerun": 2337, "abs_delta": 0 }, "refusal_recall_pct": { "committed": 83.72, "rerun": 84.11, "abs_delta": 0.39 }, "answer_on_unknown_pct": { "committed": 16.28, "rerun": 15.89, "abs_delta": 0.39 }, "over_refusal_pct": { "committed": 79.2, "rerun": 79.59, "abs_delta": 0.39 }, "refusal_rate_pct": { "committed": 80.59, "rerun": 80.97, "abs_delta": 0.38 }, "correct_on_known_pct": { "committed": 36.63, "rerun": 36.9, "abs_delta": 0.27 }, "correct_on_unknown_pct": { "committed": 0.0, "rerun": 0.0, "abs_delta": 0.0 }, "truthful_pct": { "committed": 30.93, "rerun": 30.99, "abs_delta": 0.06 } }, "on_fail": "void_arms_A2_A6_A7_report_on_five_arms" }, "G2": { "gate": "G2", "status": "PASS", "coverage_min": 99.0, "per_arm": { "A1": { "S_KUQ": { "n": 5540, "n_known": 3071, "n_unknown": 2469, "expected": { "total": 5540, "known": 3071, "unknown": 2469 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_AMBIGQA": { "n": 1832, "n_known": 830, "n_unknown": 1002, "expected": { "total": 1832, "known": 830, "unknown": 1002 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_BIGBENCH": { "n": 46, "n_known": 23, "n_unknown": 23, "expected": { "total": 46, "known": 23, "unknown": 23 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true } }, "A2": { "S_KUQ": { "status": "NOT_RUN" }, "S_AMBIGQA": { "status": "NOT_RUN" }, "S_BIGBENCH": { "status": "NOT_RUN" } }, "A3": { "S_KUQ": { "n": 5540, "n_known": 3071, "n_unknown": 2469, "expected": { "total": 5540, "known": 3071, "unknown": 2469 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_AMBIGQA": { "n": 1832, "n_known": 830, "n_unknown": 1002, "expected": { "total": 1832, "known": 830, "unknown": 1002 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_BIGBENCH": { "n": 46, "n_known": 23, "n_unknown": 23, "expected": { "total": 46, "known": 23, "unknown": 23 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true } }, "A4": { "S_KUQ": { "n": 5540, "n_known": 3071, "n_unknown": 2469, "expected": { "total": 5540, "known": 3071, "unknown": 2469 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_AMBIGQA": { "n": 1832, "n_known": 830, "n_unknown": 1002, "expected": { "total": 1832, "known": 830, "unknown": 1002 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_BIGBENCH": { "n": 46, "n_known": 23, "n_unknown": 23, "expected": { "total": 46, "known": 23, "unknown": 23 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true } }, "A5": { "S_KUQ": { "n": 5540, "n_known": 3071, "n_unknown": 2469, "expected": { "total": 5540, "known": 3071, "unknown": 2469 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_AMBIGQA": { "n": 1832, "n_known": 830, "n_unknown": 1002, "expected": { "total": 1832, "known": 830, "unknown": 1002 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_BIGBENCH": { "n": 46, "n_known": 23, "n_unknown": 23, "expected": { "total": 46, "known": 23, "unknown": 23 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true } }, "A6": { "S_KUQ": { "status": "NOT_RUN" }, "S_AMBIGQA": { "status": "NOT_RUN" }, "S_BIGBENCH": { "status": "NOT_RUN" } }, "A7": { "S_KUQ": { "status": "NOT_RUN" }, "S_AMBIGQA": { "status": "NOT_RUN" }, "S_BIGBENCH": { "status": "NOT_RUN" } }, "A8": { "S_KUQ": { "n": 5540, "n_known": 3071, "n_unknown": 2469, "expected": { "total": 5540, "known": 3071, "unknown": 2469 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_AMBIGQA": { "n": 1832, "n_known": 830, "n_unknown": 1002, "expected": { "total": 1832, "known": 830, "unknown": 1002 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true }, "S_BIGBENCH": { "n": 46, "n_known": 23, "n_unknown": 23, "expected": { "total": 46, "known": 23, "unknown": 23 }, "n_match": true, "json_coverage_pct": 100.0, "coverage_pass": true, "pass": true } } } }, "G3": { "gate": "G3", "status": "PASS", "config_enable_thinking_false": { "eval_ood_breadth_clean_schema_sft_grpo_v2_seed1_corrected_base_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_clean_schema_sft_grpo_v3_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_clean_schema_sft_merged_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_contrastive_masked_schema_sft_merged_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_contrastive_schema_sft_merged_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_grpo_v3_beta005_on_contrastive_sft_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_grpo_v3_on_contrastive_sft_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_probe_factual_schema_sft_merged_seed1_local_4b.yaml": { "enable_thinking": false, "pass": true }, "eval_ood_breadth_smoke64_A1_ambigqa_local_4b.yaml": { "enable_thinking": false, "pass": true } }, "post_hoc_think_marker_scan": { "A1__kuq": 0, "A1__ambigqa": 0, "A1__bigbench_known_unknowns": 0, "A3__kuq": 0, "A3__ambigqa": 0, "A3__bigbench_known_unknowns": 0, "A4__kuq": 0, "A4__ambigqa": 0, "A4__bigbench_known_unknowns": 0, "A5__kuq": 0, "A5__ambigqa": 0, "A5__bigbench_known_unknowns": 0, "A8__kuq": 0, "A8__ambigqa": 0, "A8__bigbench_known_unknowns": 0 }, "note": "run_eval.py's assert_no_think_scaffolding/assert_no_generated_thinking raise RuntimeError DURING generation if violated (VLLMGenerator._render_prompt / .generate), so a completed run already enforced this in-harness; the scan above is a post-hoc confirmation, not the sole enforcement." }, "G_docker_digest": { "gate": "G_docker_digest", "status": "PASS", "pinned": "sha256:f21629b9ae4ed11231768edfaed0f40d41d85d6ea9a71e8096a3d96ea0311772", "live": "sha256:f21629b9ae4ed11231768edfaed0f40d41d85d6ea9a71e8096a3d96ea0311772" } }, "integrity_all_pass": false, "evidential_gates": { "G4": { "gate": "G4", "status": "NOT_READ (integrity gate(s) FAILED)" }, "G5": { "gate": "G5", "status": "NOT_READ (integrity gate(s) FAILED)" }, "G6": { "gate": "G6", "status": "NOT_READ (integrity gate(s) FAILED)" } }, "note": "G0 is scored by screen_ood_surfaces.py (run before any generation). G7 is scored by internal_panel_probe_gate.py (needs stage-6 extraction). Neither is recomputed here." }