{
  "schema_version": 2,
  "code_commit": "3f6f422472dc6ed486d97160328560e22ce03cb9",
  "source_sha256": {
    "scripts/run_nov04_components.py": "61d35d574d9561e4611d2647174b8a6af07fabf8c30b6909e21e253542cf4336",
    "docs/research-audits/nov04-components/preflight.json": "443ad9d8319c1c32ee31f8ec1f99f046c6b3091d427512db071248fc8673dd47",
    "scripts/oct18_frozen_sources.py": "da08355049c1f6b5184ee659643c9be9decedf2041ef6a87a3dce2e4582e385a",
    "src/auto_research/agent_research/agentheal.py": "34c7a3137073075333d16d186185e439e3e5538a6172c33f686e2ce912e57699",
    "src/auto_research/agent_research/strategist.py": "f496322604ae6e54ffa74ff459bfdace27cc673514f86ac14dbfbd4678706e6a",
    "src/auto_research/foundation_models/gavagai.py": "ad8389f6d94d104bb9f76740381ec40dadb3005cde6d437c232e0345ab8e132f",
    "src/auto_research/foundation_models/proland.py": "d98512b9af75517dc5522f060244412bc2ef3535c732042d8eece14b6ca39f9d",
    "src/auto_research/post_training/rlid.py": "d349bdc43097604151e755f1ebed761f7b2f8136c0a76bf728ae95f266266805",
    "src/auto_research/post_training/specopd.py": "5602752775ce8cfeb7ca1045a5d5168be7e3f99820234f56fdf109f4a89d03de",
    "src/auto_research/reproductions/wesearch_components.py": "e516559381a2f3e0d96abd1a39e2075b78d3cee2b6b97468e4724a7642d9064c"
  },
  "device": "cuda",
  "accelerator": "NVIDIA A100-SXM4-80GB",
  "seeds": [
    42,
    43,
    44
  ],
  "components": {
    "proland": {
      "per_seed": [
        {
          "seed": 42,
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "dense_reference_max_error": 2.7755575615628914e-17,
            "horizontal_error": 6.351593401586919e-17,
            "initial_gram_violation": 0.10075400736447204,
            "final_gram_violation": 0.06534586793644286,
            "initial_loss": 1.3162159359053247,
            "final_loss": 1.3022552688596924,
            "momentum_width": 4
          },
          "checks": {
            "dense_reference_agrees": true,
            "horizontal_direction": true,
            "landing_recovers_feasibility": true,
            "actual_objective_descends": true,
            "thin_state": true,
            "zero_gradient_no_landing_identity": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 43,
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "dense_reference_max_error": 1.1102230246251565e-16,
            "horizontal_error": 1.202545072173777e-16,
            "initial_gram_violation": 0.10075400736447186,
            "final_gram_violation": 0.06534579027706716,
            "initial_loss": 0.8409383751499768,
            "final_loss": 0.8321902841440918,
            "momentum_width": 4
          },
          "checks": {
            "dense_reference_agrees": true,
            "horizontal_direction": true,
            "landing_recovers_feasibility": true,
            "actual_objective_descends": true,
            "thin_state": true,
            "zero_gradient_no_landing_identity": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 44,
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "dense_reference_max_error": 1.1102230246251565e-16,
            "horizontal_error": 1.2195379109726777e-16,
            "initial_gram_violation": 0.1007540073644724,
            "final_gram_violation": 0.06534371925697756,
            "initial_loss": 1.5075069305948812,
            "final_loss": 1.493280115130914,
            "momentum_width": 4
          },
          "checks": {
            "dense_reference_agrees": true,
            "horizontal_direction": true,
            "landing_recovers_feasibility": true,
            "actual_objective_descends": true,
            "thin_state": true,
            "zero_gradient_no_landing_identity": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        }
      ]
    },
    "gavagai": {
      "per_seed": [
        {
          "seed": 42,
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "scalar_dual_max_error": 1.1102230246251565e-16,
            "inactive_cap_softmax_error": 1.1102230246251565e-16,
            "envelope_gradient_error": 0.0,
            "eight_iteration_row_residual": 0.0021412147736108444,
            "eight_iteration_capacity_excess": 2.220446049250313e-16,
            "converged_row_residual": 0.0,
            "nothing_mass": 0.7
          },
          "checks": {
            "scalar_dual_reference": true,
            "inactive_cap_is_row_softmax": true,
            "zero_capacity_goes_to_nothing": true,
            "no_gradient_through_plan": true,
            "envelope_gradient": true,
            "caption_gradient_retained": true,
            "finite_eight_step_plan": true,
            "capacity_bound": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 43,
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "scalar_dual_max_error": 1.6653345369377348e-16,
            "inactive_cap_softmax_error": 1.1102230246251565e-16,
            "envelope_gradient_error": 0.0,
            "eight_iteration_row_residual": 0.000527098141325899,
            "eight_iteration_capacity_excess": 0.0,
            "converged_row_residual": 2.220446049250313e-16,
            "nothing_mass": 0.7
          },
          "checks": {
            "scalar_dual_reference": true,
            "inactive_cap_is_row_softmax": true,
            "zero_capacity_goes_to_nothing": true,
            "no_gradient_through_plan": true,
            "envelope_gradient": true,
            "caption_gradient_retained": true,
            "finite_eight_step_plan": true,
            "capacity_bound": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 44,
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "scalar_dual_max_error": 1.1102230246251565e-16,
            "inactive_cap_softmax_error": 1.1102230246251565e-16,
            "envelope_gradient_error": 0.0,
            "eight_iteration_row_residual": 0.0003125999529579593,
            "eight_iteration_capacity_excess": 0.0,
            "converged_row_residual": 0.0,
            "nothing_mass": 0.7
          },
          "checks": {
            "scalar_dual_reference": true,
            "inactive_cap_is_row_softmax": true,
            "zero_capacity_goes_to_nothing": true,
            "no_gradient_through_plan": true,
            "envelope_gradient": true,
            "caption_gradient_retained": true,
            "finite_eight_step_plan": true,
            "capacity_bound": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        }
      ]
    },
    "specopd": {
      "per_seed": [
        {
          "seed": 42,
          "device": "cuda",
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "loss": 0.842613011777569,
            "loss_after": 0.8425797639304065,
            "gradient_norm": 0.057656912540409645,
            "causal_scoring_max_error": 0.0,
            "outside_support_loss": 0.19999999999999996,
            "model_calls": 0
          },
          "checks": {
            "packed_matches_independent_causal_scoring": true,
            "predecessor_readout_not_own_token": true,
            "no_future_or_sibling_visibility": true,
            "reference_loss_matches": true,
            "teacher_frozen": true,
            "draft_gradient_finite_nonzero": true,
            "excluded_support_zero_gradient": true,
            "teacher_outside_support_mass_retained": true,
            "small_update_decreases_component_loss": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 43,
          "device": "cuda",
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "loss": 0.6682167888589112,
            "loss_after": 0.6681019537577053,
            "gradient_norm": 0.10715972903561662,
            "causal_scoring_max_error": 0.0,
            "outside_support_loss": 0.19999999999999996,
            "model_calls": 0
          },
          "checks": {
            "packed_matches_independent_causal_scoring": true,
            "predecessor_readout_not_own_token": true,
            "no_future_or_sibling_visibility": true,
            "reference_loss_matches": true,
            "teacher_frozen": true,
            "draft_gradient_finite_nonzero": true,
            "excluded_support_zero_gradient": true,
            "teacher_outside_support_mass_retained": true,
            "small_update_decreases_component_loss": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 44,
          "device": "cuda",
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "loss": 0.8841500809412923,
            "loss_after": 0.8841215544703034,
            "gradient_norm": 0.05340531761948606,
            "causal_scoring_max_error": 0.0,
            "outside_support_loss": 0.19999999999999996,
            "model_calls": 0
          },
          "checks": {
            "packed_matches_independent_causal_scoring": true,
            "predecessor_readout_not_own_token": true,
            "no_future_or_sibling_visibility": true,
            "reference_loss_matches": true,
            "teacher_frozen": true,
            "draft_gradient_finite_nonzero": true,
            "excluded_support_zero_gradient": true,
            "teacher_outside_support_mass_retained": true,
            "small_update_decreases_component_loss": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        }
      ]
    },
    "rlid": {
      "per_seed": [
        {
          "seed": 42,
          "device": "cuda",
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "loss": -0.225,
            "loss_after": -0.2254309822478238,
            "gradient_norm": 0.20760333876546286,
            "mean_update_max_error": 2.7755575615628914e-17,
            "text_variance_trace": 0.3391835423243354,
            "decision_variance_trace": 0.1013288752211105,
            "realization_second_moment": 0.2378546671032249,
            "variance_identity_error": 0.0,
            "model_calls": 0
          },
          "checks": {
            "only_menu_actions_have_mass": true,
            "menu_mass_one": true,
            "gradient_finite_nonzero": true,
            "padding_gradient_zero": true,
            "old_advantage_calibration_frozen": true,
            "small_update_decreases_surrogate": true,
            "same_policy_expected_update_preserved": true,
            "realization_variance_identity": true,
            "realization_variance_positive": true,
            "uniform_label_penalty_zero": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 43,
          "device": "cuda",
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "loss": -0.225,
            "loss_after": -0.2266003116552401,
            "gradient_norm": 0.39998153278921156,
            "mean_update_max_error": 2.7755575615628914e-17,
            "text_variance_trace": 0.1893407601246771,
            "decision_variance_trace": 0.10217774765702331,
            "realization_second_moment": 0.08716301246765376,
            "variance_identity_error": 2.7755575615628914e-17,
            "model_calls": 0
          },
          "checks": {
            "only_menu_actions_have_mass": true,
            "menu_mass_one": true,
            "gradient_finite_nonzero": true,
            "padding_gradient_zero": true,
            "old_advantage_calibration_frozen": true,
            "small_update_decreases_surrogate": true,
            "same_policy_expected_update_preserved": true,
            "realization_variance_identity": true,
            "realization_variance_positive": true,
            "uniform_label_penalty_zero": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        },
        {
          "seed": 44,
          "device": "cuda",
          "diagnostic_only": true,
          "evaluation_tier": "l1_mechanism",
          "metrics": {
            "loss": -0.225,
            "loss_after": -0.22569514902081098,
            "gradient_norm": 0.26371676555470075,
            "mean_update_max_error": 2.7755575615628914e-17,
            "text_variance_trace": 0.21199545394414548,
            "decision_variance_trace": 0.16690655185562606,
            "realization_second_moment": 0.04508890208851941,
            "variance_identity_error": 2.0816681711721685e-17,
            "model_calls": 0
          },
          "checks": {
            "only_menu_actions_have_mass": true,
            "menu_mass_one": true,
            "gradient_finite_nonzero": true,
            "padding_gradient_zero": true,
            "old_advantage_calibration_frozen": true,
            "small_update_decreases_surrogate": true,
            "same_policy_expected_update_preserved": true,
            "realization_variance_identity": true,
            "realization_variance_positive": true,
            "uniform_label_penalty_zero": true
          },
          "execution_device": "cuda",
          "execution_accelerator": "NVIDIA A100-SXM4-80GB"
        }
      ]
    }
  },
  "elapsed_seconds": 0.9178347699344158,
  "diagnostic_only": true,
  "evaluation_tier": "l1_mechanism",
  "formal_comparison": false,
  "promotion_eligible": false,
  "evolve_integrated": false,
  "pretrained_model_calls": 0,
  "full_experiments_executed": false
}