{
  "schema_version": 1,
  "batch": "2026-10-26-priority-components",
  "seeds": [
    42,
    43,
    44
  ],
  "evaluation_tier": "l1_mechanism",
  "components": [
    {
      "key": "detector-scope",
      "source": "src/auto_research/foundation_models/detector_scope.py",
      "device": "cuda",
      "checks": [
        "actual_sgd_momentum_proximal_update",
        "external_sae_and_validation_scores_frozen",
        "folded_raw_weights_match_normalized_logits",
        "max_pool_masks_padding_and_concatenates_layers",
        "proximal_exact_zeros_bias_not_shrunk",
        "spearman_ties_and_global_mask",
        "training_only_rms_no_centering",
        "zero_ablation_is_exact_logit_contribution"
      ],
      "metrics": [
        "counterfactual_attribution_error",
        "fixture_bce_after",
        "fixture_bce_before",
        "gradient_norm",
        "held_out_documents",
        "nonzero_head_weights",
        "optimizer_steps",
        "padding_negative_control_gap",
        "pretrained_model_calls",
        "raw_fold_error",
        "retained_features",
        "rms_leakage_negative_control_gap",
        "training_documents"
      ],
      "gates": [
        {
          "metric": "fixture_bce_after",
          "op": "lt",
          "other_metric": "fixture_bce_before"
        },
        {
          "metric": "raw_fold_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "counterfactual_attribution_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "padding_negative_control_gap",
          "op": "gt",
          "value": 1
        },
        {
          "metric": "rms_leakage_negative_control_gap",
          "op": "gt",
          "value": 1
        },
        {
          "metric": "optimizer_steps",
          "op": "eq",
          "value": 1
        },
        {
          "metric": "pretrained_model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "dt-jepa",
      "source": "src/auto_research/foundation_models/dt_jepa.py",
      "device": "cuda",
      "checks": [
        "actor_q_gradient_not_detached",
        "adaln_formula_reference",
        "bc_sums_time_then_means_batch",
        "ema_future_target_stop_gradient",
        "ema_parameter_update_reference",
        "gaussian_nll_independent_reference",
        "inference_has_no_future_or_labels",
        "multistep_twin_min_target_reference",
        "reparameterized_predictor_receives_gradient"
      ],
      "metrics": [
        "adaln_reference_error",
        "bc_predictor_gradient_norm",
        "behavior_cloning_loss",
        "ema_reference_error",
        "environment_interactions",
        "jepa_nll",
        "nll_reference_error",
        "predictor_gradient_norm",
        "pretraining_loss",
        "td_reference_error"
      ],
      "gates": [
        {
          "metric": "nll_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "adaln_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "ema_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "td_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "predictor_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "bc_predictor_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "environment_interactions",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "editjev",
      "source": "src/auto_research/system_one/editjev.py",
      "device": "cuda",
      "checks": [
        "beta_zero_uses_only_predicted_utility",
        "empty_pair_request_zero_gradient",
        "four_rating_min_geometric_mean",
        "last_nonpadding_hidden_linear_head",
        "per_request_pair_mean_reference",
        "positive_beta_prefers_faster_tied_utility",
        "preference_head_has_actual_gradient",
        "ranking_step_reduces_loss",
        "routing_does_not_accept_quality_labels"
      ],
      "metrics": [
        "candidate_editors",
        "editor_model_calls",
        "gradient_norm",
        "head_gradient_norm",
        "head_pooling_reference_error",
        "loss",
        "loss_after_step",
        "ranking_reference_error",
        "requests"
      ],
      "gates": [
        {
          "metric": "ranking_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "head_pooling_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "loss_after_step",
          "op": "lt",
          "other_metric": "loss"
        },
        {
          "metric": "gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "head_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "editor_model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "frontier-opd",
      "source": "src/auto_research/agent_research/frontier_opd.py",
      "device": "cuda",
      "checks": [
        "distinct_nodes_ranked",
        "equal_turn_not_token_weighting",
        "frozen_scores",
        "invalid_replay_rejected",
        "replay_representative_tie_order",
        "signed_not_absolute_ranking",
        "trajectory_balanced_not_visit_weighted",
        "zero_margin_equality_accepted"
      ],
      "metrics": [
        "agreement_reference_error",
        "matched_window_gain",
        "selected_nodes",
        "signed_score_reference_error",
        "trajectory_balanced_loop_score"
      ],
      "gates": [
        {
          "metric": "signed_score_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "agreement_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "trajectory_balanced_loop_score",
          "op": "eq",
          "value": 5
        },
        {
          "metric": "selected_nodes",
          "op": "eq",
          "value": 2
        },
        {
          "metric": "matched_window_gain",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "hge",
      "source": "src/auto_research/reproductions/hge.py",
      "device": "cuda",
      "checks": [
        "conditional_item_not_softmax",
        "context_injected_after_item_attention",
        "eq5_independent_reference",
        "finite_nonzero_backward",
        "fusion_endpoints",
        "position_negative_control",
        "price_negative_control",
        "residual_negative_control"
      ],
      "metrics": [
        "context_attention_error",
        "examples",
        "items_per_container",
        "parameter_gradient_l1",
        "pll_reference_error",
        "position_negative_control_gap",
        "pretrained_model_calls",
        "price_negative_control_gap",
        "residual_negative_control_gap"
      ],
      "gates": [
        {
          "metric": "pll_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "context_attention_error",
          "op": "eq",
          "value": 0
        },
        {
          "metric": "position_negative_control_gap",
          "op": "gt",
          "value": 1e-8
        },
        {
          "metric": "price_negative_control_gap",
          "op": "gt",
          "value": 1e-8
        },
        {
          "metric": "residual_negative_control_gap",
          "op": "gt",
          "value": 1e-8
        },
        {
          "metric": "parameter_gradient_l1",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "pretrained_model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "mart-grpo",
      "source": "src/auto_research/post_training/mart_grpo.py",
      "device": "cuda",
      "checks": [
        "failure_suffix_zero_gradient",
        "gradient_matches_reference",
        "inclusive_upper_event_boundary",
        "old_policy_and_advantage_detached",
        "original_denominator_reference",
        "recovered_event_full_feedback",
        "success_feedback_preserved"
      ],
      "metrics": [
        "gradient_norm",
        "gradient_reference_error",
        "masked_loss",
        "reference_error",
        "retained_failure_fraction"
      ],
      "gates": [
        {
          "metric": "reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "gradient_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "retained_failure_fraction",
          "op": "eq",
          "value": 0.75
        }
      ]
    },
    {
      "key": "opst",
      "source": "src/auto_research/post_training/opst.py",
      "device": "cuda",
      "checks": [
        "candidate_consensus_rejected",
        "different_answer_consensus_retained",
        "ema_detached",
        "ema_new_student_weight",
        "equal_response_weighting",
        "failed_verification_rejected",
        "gradient_matches_reference",
        "padding_zero_gradient"
      ],
      "metrics": [
        "ema_parameter_delta",
        "external_divergence_loss",
        "gradient_norm",
        "gradient_reference_error",
        "reference_error",
        "retained_questions"
      ],
      "gates": [
        {
          "metric": "reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "gradient_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "retained_questions",
          "op": "eq",
          "value": 1
        },
        {
          "metric": "ema_parameter_delta",
          "op": "gt",
          "value": 0
        }
      ]
    },
    {
      "key": "uniaction",
      "source": "src/auto_research/multimodal/uniaction.py",
      "device": "cuda",
      "checks": [
        "actual_action_key_value_gradients",
        "attention_mass_normalized",
        "duplicated_action_cardinality_invariant",
        "joint_softmax_formula_reference",
        "native_position_interface_receives_coordinates",
        "padding_excluded_from_cardinality_and_output",
        "semantic_identity_reuses_codebook",
        "uncalibrated_duplication_negative_control"
      ],
      "metrics": [
        "action_key_gradient_norm",
        "action_value_gradient_norm",
        "attention_reference_error",
        "duplication_invariance_error",
        "pretrained_video_model_calls",
        "semantic_codebook_gradient_norm",
        "uncalibrated_duplication_gap"
      ],
      "gates": [
        {
          "metric": "attention_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "duplication_invariance_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "uncalibrated_duplication_gap",
          "op": "gt",
          "value": 0.000001
        },
        {
          "metric": "action_key_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "action_value_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "semantic_codebook_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "pretrained_video_model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    }
  ]
}