{
  "schema_version": 1,
  "batch": "2026-10-29-priority-components",
  "seeds": [
    42,
    43,
    44
  ],
  "evaluation_tier": "l1_mechanism",
  "components": [
    {
      "key": "rescom",
      "source": "src/auto_research/foundation_models/rescom.py",
      "device": "cuda",
      "checks": [
        "calibration_factors_frozen",
        "full_rank_ridge_independent_reference",
        "heldout_activation_residual_error_reduced",
        "intercept_matches_calibration_mean",
        "rank_budget_retained",
        "rank_zero_is_mean_only",
        "same_quantized_bin_retains_distinct_residual",
        "streamed_centered_moments_reference"
      ],
      "metrics": [
        "correction_rank",
        "heldout_error_after",
        "heldout_error_before",
        "mean_reference_error",
        "model_calls",
        "ridge_reference_error",
        "stream_reference_error"
      ],
      "gates": [
        {
          "metric": "stream_reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "ridge_reference_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "heldout_error_after",
          "op": "lt",
          "other_metric": "heldout_error_before"
        },
        {
          "metric": "correction_rank",
          "op": "eq",
          "value": 2
        },
        {
          "metric": "model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "select-ttt",
      "source": "src/auto_research/foundation_models/select_ttt.py",
      "device": "cuda",
      "checks": [
        "apply_precedes_current_chunk_write",
        "blockwise_lse_dense_query_ce_reference",
        "clip_not_silently_renormalized",
        "fast_weight_update_independent_reference",
        "lower_ce_receives_stronger_write",
        "query_multiplicity_retained",
        "real_fast_weight_state_changes",
        "unclipped_budget_sums_to_chunk_count"
      ],
      "metrics": [
        "clipped_write_budget",
        "model_calls",
        "score_reference_error",
        "state_reference_error",
        "unclipped_write_budget",
        "weight_delta"
      ],
      "gates": [
        {
          "metric": "score_reference_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "state_reference_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "weight_delta",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "clipped_write_budget",
          "op": "eq",
          "value": 1.2
        },
        {
          "metric": "model_calls",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "gui-jepa",
      "source": "src/auto_research/foundation_models/gui_jepa.py",
      "device": "cuda",
      "checks": [
        "candidate_conditioned_latents_differ",
        "candidate_permutation_equivariant",
        "frozen_external_screen_receives_no_gradient",
        "invalid_candidates_cannot_be_selected",
        "predictor_and_structural_head_receive_gradient",
        "three_real_transformer_layers",
        "trainable_policy_parameters_change",
        "weighted_listwise_independent_reference"
      ],
      "metrics": [
        "environment_interactions",
        "listwise_loss",
        "loss_reference_error",
        "parameter_delta",
        "permutation_reference_error",
        "predictor_gradient_norm",
        "structural_gradient_norm"
      ],
      "gates": [
        {
          "metric": "permutation_reference_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "loss_reference_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "parameter_delta",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "predictor_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "structural_gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "environment_interactions",
          "op": "eq",
          "value": 0
        }
      ]
    },
    {
      "key": "kl-mopd",
      "source": "src/auto_research/reproductions/kl_mopd.py",
      "device": "cuda",
      "checks": [
        "calibration_frozen",
        "dynamic_recalibration_negative_control",
        "padding_zero_gradient",
        "probability_reference",
        "real_student_update",
        "reverse_kl_direction",
        "teachers_frozen"
      ],
      "metrics": [
        "gradient_norm",
        "loss_after",
        "loss_before",
        "reference_error",
        "weight_sum_error"
      ],
      "gates": [
        {
          "metric": "reference_error",
          "op": "lt",
          "value": 1e-12
        },
        {
          "metric": "loss_after",
          "op": "lt",
          "other_metric": "loss_before"
        },
        {
          "metric": "gradient_norm",
          "op": "gt",
          "value": 0
        },
        {
          "metric": "weight_sum_error",
          "op": "lt",
          "value": 1e-12
        }
      ]
    },
    {
      "key": "sleep-lti",
      "source": "src/auto_research/foundation_models/sleep_lti.py",
      "device": "cpu",
      "checks": [
        "distance_symmetry",
        "inclusive_threshold",
        "lexical_not_semantic_proxy",
        "max_length_normalizer",
        "semantic_branch_survives"
      ],
      "metrics": [
        "exact_lexical_without_semantic",
        "insertion_absolute_reward",
        "over_threshold_reward",
        "single_edit_reward"
      ],
      "gates": [
        {
          "metric": "single_edit_reward",
          "op": "eq",
          "value": 0.9
        },
        {
          "metric": "over_threshold_reward",
          "op": "eq",
          "value": 0.4
        },
        {
          "metric": "exact_lexical_without_semantic",
          "op": "eq",
          "value": 0.6
        }
      ]
    },
    {
      "key": "hope-hilbert",
      "source": "src/auto_research/foundation_models/hope_hilbert.py",
      "device": "cpu",
      "checks": [
        "biased_cross_matches_conditional_integral",
        "centering_is_not_silently_dropped",
        "dense_bn_mean_compensated",
        "positive_neuron_gauge_invariant",
        "single_step_undo_exact",
        "static_footprint_changes_selection",
        "two_neuron_gram_psd"
      ],
      "metrics": [
        "bn_mean_compensation_error",
        "capacity_only_negative_control_rate_gap",
        "cross_kernel_reference_error",
        "gram_min_eigenvalue",
        "mean_removal_negative_control_gap",
        "restore_error",
        "scale_invariance_error",
        "selected_rate"
      ],
      "gates": [
        {
          "metric": "cross_kernel_reference_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "scale_invariance_error",
          "op": "lt",
          "value": 1e-10
        },
        {
          "metric": "restore_error",
          "op": "eq",
          "value": 0
        },
        {
          "metric": "capacity_only_negative_control_rate_gap",
          "op": "gt",
          "value": 0
        }
      ]
    },
    {
      "key": "discernmem",
      "source": "src/auto_research/agent_research/discernmem.py",
      "device": "cpu",
      "checks": [
        "avoidance_precedes_content",
        "complete_failure_guard_not_tolerance_only",
        "historical_regression_rejected",
        "identity_owner_and_corrections_preserved",
        "isolated_effect_not_team_credit",
        "negative_use_requires_supported_defect",
        "new_version_resets_old_evidence",
        "repair_passes_both_references"
      ],
      "metrics": [
        "accepted_trigger_gain",
        "failure_count_negative_control",
        "new_version_effects",
        "paired_effect",
        "parent_version_effects",
        "regression_worst_margin"
      ],
      "gates": [
        {
          "metric": "paired_effect",
          "op": "eq",
          "value": -0.4
        },
        {
          "metric": "new_version_effects",
          "op": "eq",
          "value": 0
        },
        {
          "metric": "failure_count_negative_control",
          "op": "eq",
          "value": 1
        }
      ]
    },
    {
      "key": "ruharness",
      "source": "src/auto_research/agent_research/ruharness.py",
      "device": "cpu",
      "checks": [
        "duplicate_sets_have_unit_jaccard",
        "finite_conservative_scores",
        "fixed_seed_reproducible",
        "formula_reference_counts",
        "inactive_not_selected",
        "no_candidate_preserves_incumbent",
        "strict_epsilon_boundary",
        "unverified_candidate_excluded"
      ],
      "metrics": [
        "best_lower_quantile",
        "duplicate_criterion_redundancy",
        "frontier_size",
        "posterior_samples_per_criterion",
        "reference_discrimination",
        "reference_failure_coverage",
        "reference_predictive_validity"
      ],
      "gates": [
        {
          "metric": "reference_discrimination",
          "op": "eq",
          "value": 1
        },
        {
          "metric": "reference_predictive_validity",
          "op": "eq",
          "value": 0.5
        },
        {
          "metric": "posterior_samples_per_criterion",
          "op": "eq",
          "value": 2000
        },
        {
          "metric": "frontier_size",
          "op": "eq",
          "value": 3
        }
      ]
    }
  ]
}