{
  "schema_version": 1,
  "batch": "2026-10-30-overlap-components",
  "seeds": [
    42,
    43,
    44
  ],
  "evaluation_tier": "l1_mechanism",
  "components": [
    {
      "key": "concur",
      "source": "src/auto_research/post_training/concur.py",
      "device": "cpu",
      "checks": [
        "all_pass_auxiliary_removed",
        "below_ceiling_ties_not_collapsed",
        "binary_primary_has_no_trust_reading",
        "ceiling_tied",
        "cone_energy_identity",
        "mean_preserved",
        "primary_order_preserved",
        "sixty_one_readings_required",
        "whitening_is_whole_batch"
      ],
      "metrics": [
        "all_pass_magnitude",
        "order_energy",
        "projection_mean_error",
        "pythagorean_error",
        "tied_energy",
        "valid_trust_readings"
      ],
      "gates": [
        {
          "metric": "projection_mean_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "pythagorean_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "all_pass_magnitude",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "valid_trust_readings",
          "op": "eq",
          "value": 61
        }
      ]
    },
    {
      "key": "merlin",
      "source": "src/auto_research/foundation_models/merlin.py",
      "device": "cuda",
      "checks": [
        "alignment_mean_independent_reference",
        "alignment_variance_independent_reference",
        "completed_segments_have_exact_boundary",
        "moments_gate_queries_receive_real_gradient",
        "raw_second_is_not_variance",
        "readout_aggregation_reference",
        "segment_future_mask_blocks_readout_changes",
        "two_softmax_mixture_not_logit_mixture"
      ],
      "metrics": [
        "future_readout_leak_error",
        "gradient_norm",
        "mean_reference_error",
        "minimum_gradient_group_norm",
        "model_calls",
        "raw_moment_negative_control_gap",
        "single_softmax_negative_control_gap",
        "variance_reference_error"
      ],
      "gates": [
        {
          "metric": "mean_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "variance_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "future_readout_leak_error",
          "op": "eq",
          "value": 0
        },
        {
          "metric": "minimum_gradient_group_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "crop",
      "source": "src/auto_research/agent_research/crop.py",
      "device": "cpu",
      "checks": [
        "callback_doubles_not_capability",
        "literal_gradient_concatenation",
        "optimizer_reference_field_absent",
        "score_matches_equation4",
        "strict_length_gate",
        "unvalidated_final_not_selected",
        "validated_prompt_selected"
      ],
      "metrics": [
        "brevity_callback_count",
        "evaluated_prompt_count",
        "reference_normalized_length",
        "reference_normalized_score"
      ],
      "gates": [
        {
          "metric": "evaluated_prompt_count",
          "op": "eq",
          "value": 2
        },
        {
          "metric": "brevity_callback_count",
          "op": "eq",
          "value": 2
        },
        {
          "metric": "reference_normalized_score",
          "op": "gt",
          "value": 0.91999999
        },
        {
          "metric": "reference_normalized_score",
          "op": "lt",
          "value": 0.92000001
        },
        {
          "metric": "reference_normalized_length",
          "op": "eq",
          "value": 0.4
        }
      ]
    },
    {
      "key": "pinger",
      "source": "src/auto_research/reproductions/pinger_components.py",
      "device": "cuda",
      "checks": [
        "actual_actor_update",
        "dense_stream_invariant",
        "frozen_reference",
        "no_model_or_serving_calls",
        "ordinary_causal_mask_is_not_equivalent",
        "prefix_penalty_changes_parent_spread",
        "prefix_sum_not_last_level_only",
        "zero_penalty_matches_top_width"
      ],
      "metrics": [
        "actor_gradient_norm",
        "causal_mask_negative_control_gap",
        "dense_compaction_max_error",
        "greedy_selected_parents",
        "last_prefix_only_loss_gap",
        "loss_after",
        "loss_before",
        "penalized_selected_parents",
        "pretrained_model_calls"
      ],
      "gates": [
        {
          "metric": "dense_compaction_max_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "causal_mask_negative_control_gap",
          "op": "gt",
          "value": 1e-06
        },
        {
          "metric": "loss_after",
          "op": "lt",
          "other_metric": "loss_before"
        },
        {
          "metric": "actor_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "penalized_selected_parents",
          "op": "gt",
          "other_metric": "greedy_selected_parents"
        },
        {
          "metric": "pretrained_model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "pinterest-rlbp",
      "source": "src/auto_research/reproductions/pinterest_rlbp.py",
      "device": "cuda",
      "checks": [
        "CPC_reduction_retains_negative_sign",
        "actual_reward_regression_update",
        "exact_top_k_support",
        "fusion_changes_preference",
        "nine_bounded_PID_actions",
        "no_live_bidder_or_model_calls",
        "opposite_backup_directions",
        "ordinary_IPS_not_self_normalized"
      ],
      "metrics": [
        "actor_gradient_norm",
        "derivative_gain_error",
        "ips_reference_error",
        "max_min_backup_error",
        "policy_row_mass_error",
        "pretrained_model_calls",
        "td_loss_after",
        "td_loss_before",
        "wrong_cpc_max_backup_gap"
      ],
      "gates": [
        {
          "metric": "td_loss_after",
          "op": "lt",
          "other_metric": "td_loss_before"
        },
        {
          "metric": "actor_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "max_min_backup_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "wrong_cpc_max_backup_gap",
          "op": "gt",
          "value": 1
        },
        {
          "metric": "policy_row_mass_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "derivative_gain_error",
          "op": "eq",
          "value": 0
        },
        {
          "metric": "ips_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "pretrained_model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "fold-bias",
      "source": "src/auto_research/foundation_models/fold_bias.py",
      "device": "cuda",
      "checks": [
        "context_projection_mean_constraint",
        "dropping_bias_changes_attention",
        "keys_really_rounded_before_value_refit",
        "offline_fit_is_frozen",
        "rank_one_logit_identity_reference",
        "ridge_constrained_kkt_reference",
        "rounded_value_fit_normal_equations",
        "same_value_output_error_bound"
      ],
      "metrics": [
        "cache_rounding_attention_gap",
        "drop_bias_output_error",
        "kkt_reference_error",
        "logit_identity_error",
        "minimum_output_bound_slack",
        "model_calls",
        "projection_mean_error",
        "value_normal_equation_error"
      ],
      "gates": [
        {
          "metric": "logit_identity_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "projection_mean_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "kkt_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "value_normal_equation_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "cache_rounding_attention_gap",
          "op": "gt",
          "value": 1e-08
        },
        {
          "metric": "minimum_output_bound_slack",
          "op": "ge",
          "value": -1e-12
        },
        {
          "metric": "model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "swim",
      "source": "src/auto_research/foundation_models/swim.py",
      "device": "cuda",
      "checks": [
        "current_state_is_not_gated",
        "cyclic_hinge_negative_reference",
        "gates_and_inverse_model_really_train",
        "gaussian_nll_independent_reference",
        "infonce_selected_denominator_reference",
        "latent_geometry_receives_real_gradient",
        "sqrt_gate_distance_reference",
        "stage_two_gate_target_is_detached"
      ],
      "metrics": [
        "behavior_only_gate_gradient_norm",
        "correlation_loss",
        "environment_interactions",
        "gate_gradient_norm",
        "gate_parameter_delta",
        "hinge_reference_error",
        "infonce_reference_error",
        "metric_reference_error",
        "nll_reference_error",
        "predictor_gradient_norm"
      ],
      "gates": [
        {
          "metric": "nll_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "hinge_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "infonce_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "metric_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "behavior_only_gate_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "gate_parameter_delta",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "environment_interactions",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "mpopd",
      "source": "src/auto_research/post_training/mpopd.py",
      "device": "cuda",
      "checks": [
        "conditional_support_normalized",
        "eta_endpoints_match",
        "first_prefix_weight_one",
        "geometric_not_arithmetic",
        "padding_gradient_zero",
        "proposal_update_decreases_fixed_objective",
        "student_update_decreases_fixed_objective",
        "teacher_and_old_proposal_frozen",
        "weights_detached"
      ],
      "metrics": [
        "active_tokens",
        "geometric_arithmetic_distance",
        "proposal_gradient_norm",
        "proposal_loss_after",
        "proposal_loss_before",
        "student_gradient_norm",
        "student_loss_after",
        "student_loss_before"
      ],
      "gates": [
        {
          "metric": "student_loss_after",
          "op": "lt",
          "other_metric": "student_loss_before"
        },
        {
          "metric": "proposal_loss_after",
          "op": "lt",
          "other_metric": "proposal_loss_before"
        },
        {
          "metric": "student_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "proposal_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "geometric_arithmetic_distance",
          "op": "gt",
          "value": 1e-05
        },
        {
          "metric": "active_tokens",
          "op": "eq",
          "value": 7
        }
      ]
    },
    {
      "key": "emma",
      "source": "src/auto_research/agent_research/emma.py",
      "device": "cpu",
      "checks": [
        "old_snapshot_not_mutated",
        "one_skill_entry_changed",
        "overflow_preserves_completed_prefix",
        "overflow_preserves_suffix",
        "six_contract_violations_refused",
        "specialist_inherits_current_generalist",
        "unused_edit_cannot_claim_exercised"
      ],
      "metrics": [
        "original_plan_steps",
        "original_skill_count",
        "rejected_negative_controls",
        "split_plan_steps",
        "updated_skill_count"
      ],
      "gates": [
        {
          "metric": "original_skill_count",
          "op": "eq",
          "value": 0
        },
        {
          "metric": "updated_skill_count",
          "op": "eq",
          "value": 1
        },
        {
          "metric": "rejected_negative_controls",
          "op": "eq",
          "value": 6
        },
        {
          "metric": "original_plan_steps",
          "op": "eq",
          "value": 3
        },
        {
          "metric": "split_plan_steps",
          "op": "eq",
          "value": 4
        }
      ]
    },
    {
      "key": "evalalloc",
      "source": "src/auto_research/agent_research/evalalloc.py",
      "device": "cpu",
      "checks": [
        "MAP_predictions_finite",
        "acquisition_without_replacement",
        "budget_self_fit_and_hidden_cells_refused",
        "cached_cells_not_reexecuted",
        "common_core_preserved",
        "full_observation_correction_zero",
        "item_parameters_frozen_during_MAP",
        "pooled_slope_clipped",
        "search_union_cache_preserved"
      ],
      "metrics": [
        "adjusted_score_a",
        "completed_fitness",
        "information_score_max",
        "map_parameter_norm",
        "missing_execution_cells",
        "regression_slope",
        "rejected_negative_controls"
      ],
      "gates": [
        {
          "metric": "map_parameter_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "missing_execution_cells",
          "op": "eq",
          "value": 3
        },
        {
          "metric": "regression_slope",
          "op": "ge",
          "value": 0
        },
        {
          "metric": "regression_slope",
          "op": "le",
          "value": 1.5
        },
        {
          "metric": "rejected_negative_controls",
          "op": "eq",
          "value": 3
        }
      ]
    }
  ]
}