{
  "decision": "stop_before_action_only_confirmation_and_menagerie",
  "downstream": {
    "action_only_trained": false,
    "confirmation_run": false,
    "menagerie_authorized": false,
    "menagerie_seeds_consumed": 0
  },
  "evaluation": {
    "checks": {
      "bootstrap_vs_apex": false,
      "bootstrap_vs_sample_more": false,
      "commit_absolute": true,
      "commit_retention": true,
      "invalid_actions": false,
      "locality": false,
      "trained_family_retention": true,
      "transfer_vs_apex": false,
      "transfer_vs_sample_more": false,
      "verification_absolute": true,
      "verification_retention": false
    },
    "contrasts": {
      "compact_vs_apex": {
        "ci95": [
          -0.4444444444444444,
          -0.2222222222222222
        ],
        "delta": -0.3333333333333333,
        "left_only": 0,
        "left_success": 0.3472222222222222,
        "n": 72,
        "right_only": 24,
        "right_success": 0.6805555555555556
      },
      "compact_vs_apex_sample_more": {
        "ci95": [
          -0.2777777777777778,
          -0.08333333333333333
        ],
        "delta": -0.18055555555555555,
        "left_only": 1,
        "left_success": 0.3472222222222222,
        "n": 72,
        "right_only": 14,
        "right_success": 0.5277777777777778
      },
      "trained_compact_vs_apex": {
        "ci95": [
          0.0625,
          0.2708333333333333
        ],
        "delta": 0.16666666666666666,
        "left_only": 8,
        "left_success": 1.0,
        "n": 48,
        "right_only": 0,
        "right_success": 0.8333333333333334
      }
    },
    "locality": {
      "ceiling": 0.15,
      "mean_entropy_delta": -0.009434346032018481,
      "median_non_target_centered_logit_drift": 0.38596151769161224,
      "passed": false
    },
    "trained_dev": {
      "apex_replay": {
        "commit_given_verified": 1.0,
        "invalid_action_rate_per_turn": 0.23706896551724138,
        "mean_sampled_tokens": 1643.2708333333333,
        "mean_turns": 4.833333333333333,
        "n_tasks": 48,
        "per_family": {
          "collision_index": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "label_intervals": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "segmented_rollup": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "specificity_router": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "stable_merge": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "weighted_quota": {
            "n": 8,
            "submit_rate": 0.0,
            "success": 0.0
          }
        },
        "submit_rate": 0.8333333333333334,
        "success": 0.8333333333333334,
        "verified_given_success": 1.0
      },
      "compact": {
        "commit_given_verified": 1.0,
        "invalid_action_rate_per_turn": 0.0,
        "mean_sampled_tokens": 401.8333333333333,
        "mean_turns": 4,
        "n_tasks": 48,
        "per_family": {
          "collision_index": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "label_intervals": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "segmented_rollup": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "specificity_router": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "stable_merge": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "weighted_quota": {
            "n": 8,
            "submit_rate": 1.0,
            "success": 1.0
          }
        },
        "submit_rate": 1.0,
        "success": 1.0,
        "verified_given_success": 1.0
      }
    },
    "transfer_dev": {
      "apex_replay": {
        "commit_given_verified": 0.9387755102040817,
        "invalid_action_rate_per_turn": 0.09278350515463918,
        "mean_sampled_tokens": 1426.3194444444443,
        "mean_turns": 5.388888888888889,
        "n_tasks": 72,
        "per_family": {
          "dependency_order": {
            "n": 18,
            "submit_rate": 0.6111111111111112,
            "success": 0.7222222222222222
          },
          "recursive_overlay": {
            "n": 18,
            "submit_rate": 1.0,
            "success": 0.0
          },
          "retry_schedule": {
            "n": 18,
            "submit_rate": 0.9444444444444444,
            "success": 1.0
          },
          "ttl_cache": {
            "n": 18,
            "submit_rate": 1.0,
            "success": 1.0
          }
        },
        "submit_rate": 0.8888888888888888,
        "success": 0.6805555555555556,
        "verified_given_success": 1.0
      },
      "apex_sample_more": {
        "commit_given_verified": 0.875,
        "invalid_action_rate_per_turn": 0.09615384615384616,
        "mean_sampled_tokens": 2452.4583333333335,
        "mean_turns": 7.944444444444445,
        "n_tasks": 72,
        "per_family": {
          "dependency_order": {
            "n": 18,
            "submit_rate": 0.0,
            "success": 0.2777777777777778
          },
          "recursive_overlay": {
            "n": 18,
            "submit_rate": 0.4444444444444444,
            "success": 0.05555555555555555
          },
          "retry_schedule": {
            "n": 18,
            "submit_rate": 0.5,
            "success": 0.9444444444444444
          },
          "ttl_cache": {
            "n": 18,
            "submit_rate": 0.4444444444444444,
            "success": 0.8333333333333334
          }
        },
        "submit_rate": 0.3472222222222222,
        "success": 0.5277777777777778,
        "verified_given_success": 0.42105263157894735
      },
      "compact": {
        "commit_given_verified": 1.0,
        "invalid_action_rate_per_turn": 0.26,
        "mean_sampled_tokens": 2422.5138888888887,
        "mean_turns": 6.944444444444445,
        "n_tasks": 72,
        "per_family": {
          "dependency_order": {
            "n": 18,
            "submit_rate": 0.0,
            "success": 0.0
          },
          "recursive_overlay": {
            "n": 18,
            "submit_rate": 0.0,
            "success": 0.0
          },
          "retry_schedule": {
            "n": 18,
            "submit_rate": 1.0,
            "success": 1.0
          },
          "ttl_cache": {
            "n": 18,
            "submit_rate": 0.2222222222222222,
            "success": 0.3888888888888889
          }
        },
        "submit_rate": 0.3055555555555556,
        "success": 0.3472222222222222,
        "verified_given_success": 0.88
      }
    }
  },
  "input_sha256": {
    "apex_merge": "ece3c696e1abfb071d33ba523d8007bb1506d22047a1c11aa57a48a280cd1b0c",
    "apex_training": "f82f38d943043c62c7a24d1cabde5b7a4bbe7c2a8c291e28566f5c2db94b0b91",
    "compact_merge": "19ff48133f81760ff2065a3928c284facce012b0d666fa54427e222cfdabc44b",
    "compact_training": "3a0621de184ec211b186df525de7f075ecc96709b6057fb05a5b61e29d4c2f38",
    "evaluation_config": "a1af957d0d95edfdc5b55bcd598243decfb5f0c0f7105da67e71d0da5a563313",
    "evaluation_script": "dc54b20f623b16f884b00a21ac55f7ce4beaf8239e575bd2c835afee23371998",
    "locality": "483cc5b13218c6bb8386b0ef8da4be10c297d7eb39c280f64074c4411f8123c8",
    "merge_script": "8369ad49f3caf70b0802a4326afbaecc896a343912491dd2ad2c14f6e7eec927",
    "primary_gate": "f7cf3c753155d82f3c1e082c7d5255f3a9c481fce8d5fd0c3c50e30190e022bd",
    "sample_more": "d87395ea959678747defdbeebb0a5fd97982fcefecfe3025e98eb25c24ab031f",
    "trained_apex": "c231089d53020e827b6df1f3e3a0b8f426fb4ba99784cc507a6e0412d455fbb8",
    "trained_compact": "7175be5d9def65857a0ca3d22ead9b50b8a5776dd2a6f26dd8af35e0fd65b03d",
    "training_script": "f81f470d3fa3499b2569e38ad61afd353c5ab1ed354304dae840f70eaf219996",
    "transfer_apex": "fb6c180b44886bfd8e091542e883ed2e44ded00e7d948fc126033112cfdbb2fc",
    "transfer_compact": "5c2b765c363426e673baa98090872ac4ea616eb430fe344c176f84467c27e8da",
    "vllm_runner": "3908eb25ebf059a867c1d7581f1a98186470f5a6c597132a12095b6922d33dcf"
  },
  "merge": {
    "apex_replay": {
      "adapter_weights_sha256": "b2ba9aad45bb0b9a84ad906b420c007b1fbf2db5f7df6205f6011fd094873ba0",
      "applied_lora_modules": 128,
      "merged_weights_sha256": "c933168075a0fc011f31ee54f83d722cc91423ce1a0ac19bfe40dfab92f608d5",
      "nonzero_lora_modules": 128
    },
    "compact": {
      "adapter_weights_sha256": "70fb8e3dff503863c041fa3a9290c2cafa50ae9759bc79103e64cc97deeee8a1",
      "applied_lora_modules": 128,
      "merged_weights_sha256": "8024a1944e179e09f9486d6039f7b1a4fd3f731350594862adcf288c7621c80b",
      "nonzero_lora_modules": 128
    }
  },
  "model": "Qwen/Qwen3.5-4B",
  "policy_transition_diagnostics": {
    "trained_compact": {
      "after_failed_patch": {
        "next_operator_counts": {},
        "total": 0
      },
      "after_failed_test": {
        "next_operator_counts": {},
        "next_patch_rate": null,
        "total": 0
      },
      "after_passed_test": {
        "next_commit_rate": 1.0,
        "next_operator_counts": {
          "COMMIT": 48
        },
        "total": 48
      },
      "canonical_inspect_patch_verify_commit": {
        "count": 48,
        "rate": 1.0
      },
      "n_trajectories": 48,
      "tasks_repeating_identical_failed_patch": 0
    },
    "transfer_apex_replay": {
      "after_failed_patch": {
        "next_operator_counts": {
          "INSPECT": 3
        },
        "total": 3
      },
      "after_failed_test": {
        "next_operator_counts": {
          "INVALID": 2,
          "PATCH": 24
        },
        "next_patch_rate": 0.9230769230769231,
        "total": 26
      },
      "after_passed_test": {
        "next_commit_rate": 0.9846153846153847,
        "next_operator_counts": {
          "COMMIT": 64,
          "VERIFY": 1
        },
        "total": 65
      },
      "canonical_inspect_patch_verify_commit": {
        "count": 36,
        "rate": 0.5
      },
      "n_trajectories": 72,
      "tasks_repeating_identical_failed_patch": 0
    },
    "transfer_compact": {
      "after_failed_patch": {
        "next_operator_counts": {
          "INSPECT": 54
        },
        "total": 54
      },
      "after_failed_test": {
        "next_operator_counts": {
          "INSPECT": 17,
          "INVALID": 11,
          "VERIFY": 20
        },
        "next_patch_rate": 0.0,
        "total": 48
      },
      "after_passed_test": {
        "next_commit_rate": 1.0,
        "next_operator_counts": {
          "COMMIT": 22
        },
        "total": 22
      },
      "canonical_inspect_patch_verify_commit": {
        "count": 18,
        "rate": 0.25
      },
      "n_trajectories": 72,
      "tasks_repeating_identical_failed_patch": 18
    }
  },
  "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
  "runtime": {
    "driver": "550.127.08",
    "gpu": "NVIDIA RTX 6000 Ada Generation",
    "gpu_memory_mib": 49140,
    "inference_backend": {
      "engine": {
        "gpu_memory_utilization": 0.85,
        "max_model_len": 8192,
        "max_num_batched_tokens": 16384,
        "max_num_seqs": 32
      },
      "evaluation": {
        "answer_max_tokens": 256,
        "blocks": {
          "trained_dev": {
            "families": "train",
            "seed": 73200,
            "tasks_per_family": 8
          },
          "transfer_confirm": {
            "families": "transfer",
            "seed": 73400,
            "tasks_per_family": 18
          },
          "transfer_dev": {
            "families": "transfer",
            "seed": 73300,
            "tasks_per_family": 18
          }
        },
        "deep_run_seed": 73310,
        "deep_turns": 8,
        "sample_more_trajectories": 2,
        "sample_more_turns_each": 4,
        "sample_run_seed": 73311,
        "sample_temperature": 0.8,
        "sample_top_k": 20,
        "sample_top_p": 0.95,
        "think_budget": 512
      },
      "merged_composite_checkpoints": true,
      "name": "vllm",
      "runner_schema_version": 4,
      "runtime_lora": false,
      "same_backend_all_arms": true,
      "torch": "2.11.0+cu129",
      "transformers": "5.13.0",
      "version": "0.24.0"
    },
    "python": "3.12.3",
    "training_packages": {
      "bitsandbytes": "0.49.2",
      "peft": "0.19.1",
      "torch": "2.11.0+cu129",
      "transformers": "5.13.0"
    }
  },
  "schema_version": 1,
  "status": "PRIMARY_GATE_FAILED",
  "training": {
    "apex_replay": {
      "effective_batch_size": 16,
      "effective_dataset_epochs": 2.0,
      "loss_implementation": "exact_sequence_chunked_checkpointed_cross_entropy",
      "optimizer_steps": 584,
      "peak_cuda_bytes": 48578297344,
      "training_loss": 0.42299116484514654,
      "wall_seconds": 10007.277540124953
    },
    "compact": {
      "effective_batch_size": 16,
      "effective_dataset_epochs": 1.8010794140323825,
      "loss_implementation": "exact_sequence_chunked_checkpointed_cross_entropy",
      "optimizer_steps": 584,
      "peak_cuda_bytes": 48578178560,
      "repository_operator_rows": {
        "COMMIT": 129,
        "INSPECT": 129,
        "PATCH": 129,
        "VERIFY": 129
      },
      "repository_rows": 516,
      "repository_tasks": 129,
      "training_loss": 0.437239117410085,
      "wall_seconds": 9633.08647069335
    }
  }
}
