{
  "roft": {
    "schema_version": 2,
    "manifest_ref": "post-training:roft",
    "method": "roft",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "masked_ce": {
        "mean": 4.723277568817139,
        "std": 0.3799676184522421
      },
      "retrospection_tokens": {
        "mean": 3.0,
        "std": 0.0
      },
      "context_target_tokens": {
        "mean": 0.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "masked_ce": 5.138889789581299,
        "retrospection_tokens": 3,
        "context_target_tokens": 0
      },
      {
        "seed": 43,
        "masked_ce": 4.810451030731201,
        "retrospection_tokens": 3,
        "context_target_tokens": 0
      },
      {
        "seed": 44,
        "masked_ce": 4.220491886138916,
        "retrospection_tokens": 3,
        "context_target_tokens": 0
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/post-training/2609.35741-roft/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "lspd": {
    "schema_version": 2,
    "manifest_ref": "post-training:lspd",
    "method": "lspd",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "loss": {
        "mean": -0.01866374599436919,
        "std": 0.0002326233939281733
      },
      "mean_penalty": {
        "mean": 0.02250002883374691,
        "std": 0.0
      },
      "mean_entropy": {
        "mean": 4.116377512613933,
        "std": 0.02326232790400811
      },
      "tail_fraction": {
        "mean": 0.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "loss": -0.018448857590556145,
        "mean_penalty": 0.02250002883374691,
        "mean_entropy": 4.094888687133789,
        "tail_fraction": 0.0
      },
      {
        "seed": 43,
        "loss": -0.018986916169524193,
        "mean_penalty": 0.02250002883374691,
        "mean_entropy": 4.1486945152282715,
        "tail_fraction": 0.0
      },
      {
        "seed": 44,
        "loss": -0.01855546422302723,
        "mean_penalty": 0.02250002883374691,
        "mean_entropy": 4.105549335479736,
        "tail_fraction": 0.0
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/post-training/2609.35505-lspd/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "harness-learning": {
    "schema_version": 2,
    "manifest_ref": "agent-research:harness-learning",
    "method": "harness-learning",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "rounds": {
        "mean": 3.0,
        "std": 0.0
      },
      "final_harness_size": {
        "mean": 3.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "rounds": 3,
        "final_harness_size": 3,
        "uses_execution_reports": true,
        "gold_fields_exposed": false
      },
      {
        "seed": 43,
        "rounds": 3,
        "final_harness_size": 3,
        "uses_execution_reports": true,
        "gold_fields_exposed": false
      },
      {
        "seed": 44,
        "rounds": 3,
        "final_harness_size": 3,
        "uses_execution_reports": true,
        "gold_fields_exposed": false
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/agent-research/2609.35738-harness-learning/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "ms-gla": {
    "schema_version": 2,
    "manifest_ref": "foundation-models:ms-gla",
    "method": "ms-gla",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "loss": {
        "mean": 0.10698820402224858,
        "std": 0.030546831479288786
      },
      "routing_sum_error": {
        "mean": 1.1920928955078125e-07,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "loss": 0.06381452828645706,
        "routing_sum_error": 1.1920928955078125e-07,
        "finite_gradient": true
      },
      {
        "seed": 43,
        "loss": 0.12987452745437622,
        "routing_sum_error": 1.1920928955078125e-07,
        "finite_gradient": true
      },
      {
        "seed": 44,
        "loss": 0.12727555632591248,
        "routing_sum_error": 1.1920928955078125e-07,
        "finite_gradient": true
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/foundation-models/2609.35664-ms-gla/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "graphhca": {
    "schema_version": 2,
    "manifest_ref": "agent-research:graphhca",
    "method": "graphhca",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "start_value": {
        "mean": 0.57,
        "std": 0.0
      },
      "iterations": {
        "mean": 3.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "start_value": 0.57,
        "iterations": 3,
        "successful_credit_above_failure": true,
        "gold_fields_exposed": false
      },
      {
        "seed": 43,
        "start_value": 0.57,
        "iterations": 3,
        "successful_credit_above_failure": true,
        "gold_fields_exposed": false
      },
      {
        "seed": 44,
        "start_value": 0.57,
        "iterations": 3,
        "successful_credit_above_failure": true,
        "gold_fields_exposed": false
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/agent-research/2609.35084-graphhca/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "stepquant": {
    "schema_version": 2,
    "manifest_ref": "foundation-models:stepquant",
    "method": "stepquant",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "state_mse": {
        "mean": 0.025560750315586727,
        "std": 0.0006991335282533404
      },
      "average_allocated_bits": {
        "mean": 5.333333333333333,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "state_mse": 0.026501668617129326,
        "average_allocated_bits": 5.333333333333333,
        "finite_scales": true
      },
      {
        "seed": 43,
        "state_mse": 0.0253533236682415,
        "average_allocated_bits": 5.333333333333333,
        "finite_scales": true
      },
      {
        "seed": 44,
        "state_mse": 0.02482725866138935,
        "average_allocated_bits": 5.333333333333333,
        "finite_scales": true
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/foundation-models/2609.38169-stepquant/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "leapquant": {
    "schema_version": 2,
    "manifest_ref": "foundation-models:leapquant",
    "method": "leapquant",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "state_mse": {
        "mean": 4.499370576619791e-05,
        "std": 6.758646927730397e-06
      },
      "window_steps": {
        "mean": 4.0,
        "std": 0.0
      },
      "compensator_rank": {
        "mean": 2.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "state_mse": 3.667683631647378e-05,
        "window_steps": 4,
        "finite_boundary": true,
        "compensator_rank": 2
      },
      {
        "seed": 43,
        "state_mse": 4.507277481025085e-05,
        "window_steps": 4,
        "finite_boundary": true,
        "compensator_rank": 2
      },
      {
        "seed": 44,
        "state_mse": 5.3231506171869114e-05,
        "window_steps": 4,
        "finite_boundary": true,
        "compensator_rank": 2
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/foundation-models/2609.38166-leapquant/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "chinese-jev": {
    "schema_version": 2,
    "manifest_ref": "foundation-models:chinese-jev",
    "method": "chinese-jev",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "loss": {
        "mean": 1.5964340766270955,
        "std": 0.12320634264566757
      },
      "reward_std": {
        "mean": 0.41116487979888916,
        "std": 0.04942429696418572
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "loss": 1.728244662284851,
        "reward_std": 0.36087170243263245,
        "finite_gradient": true
      },
      {
        "seed": 43,
        "loss": 1.431841492652893,
        "reward_std": 0.3942747116088867,
        "finite_gradient": true
      },
      {
        "seed": 44,
        "loss": 1.6292160749435425,
        "reward_std": 0.4783482253551483,
        "finite_gradient": true
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/foundation-models/2609.36965-chinese-jev/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "dr-opd": {
    "schema_version": 2,
    "manifest_ref": "post-training:dr-opd",
    "method": "dr-opd",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "credit_rms": {
        "mean": 0.20002265771230063,
        "std": 0.021288841537483238
      },
      "weight_range": {
        "mean": 1.7405274932583172,
        "std": 0.34956036327241213
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "credit_rms": 0.2263151854276657,
        "weight_range": 2.0063036903738976,
        "finite_gradient": true
      },
      {
        "seed": 43,
        "credit_rms": 0.19957873225212097,
        "weight_range": 1.2466533407568932,
        "finite_gradient": true
      },
      {
        "seed": 44,
        "credit_rms": 0.17417405545711517,
        "weight_range": 1.9686254486441612,
        "finite_gradient": true
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/post-training/2609.38025-dr-opd/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "sipo": {
    "schema_version": 2,
    "manifest_ref": "post-training:sipo",
    "method": "sipo",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "mean_abs_evidence": {
        "mean": 0.11643468091885249,
        "std": 0.007528721869503832
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "mean_abs_evidence": 0.12633292376995087,
        "uniform_failure_signal": true,
        "finite_gradient": true
      },
      {
        "seed": 43,
        "mean_abs_evidence": 0.11488277465105057,
        "uniform_failure_signal": true,
        "finite_gradient": true
      },
      {
        "seed": 44,
        "mean_abs_evidence": 0.10808834433555603,
        "uniform_failure_signal": true,
        "finite_gradient": true
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/post-training/2609.36742-sipo/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "remem": {
    "schema_version": 2,
    "manifest_ref": "agent-research:remem",
    "method": "remem",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "memory_tokens": {
        "mean": 4.0,
        "std": 0.0
      },
      "history_chunks": {
        "mean": 4.0,
        "std": 0.0
      },
      "memory_objective": {
        "mean": 0.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "memory_tokens": 4,
        "history_chunks": 4,
        "memory_objective": 0.0,
        "gold_fields_exposed": false
      },
      {
        "seed": 43,
        "memory_tokens": 4,
        "history_chunks": 4,
        "memory_objective": 0.0,
        "gold_fields_exposed": false
      },
      {
        "seed": 44,
        "memory_tokens": 4,
        "history_chunks": 4,
        "memory_objective": 0.0,
        "gold_fields_exposed": false
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/agent-research/2609.37311-remem/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "video-rsi": {
    "schema_version": 2,
    "manifest_ref": "agent-research:video-rsi",
    "method": "video-rsi",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "accepted_branches": {
        "mean": 2.0,
        "std": 0.0
      },
      "rejected_branches": {
        "mean": 1.0,
        "std": 0.0
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "accepted_branches": 2,
        "rejected_branches": 1,
        "accuracy_branch_accepts": true,
        "cost_branch_accepts": true,
        "non_pareto_candidate_rejected": true,
        "private_example_scores_exposed": false
      },
      {
        "seed": 43,
        "accepted_branches": 2,
        "rejected_branches": 1,
        "accuracy_branch_accepts": true,
        "cost_branch_accepts": true,
        "non_pareto_candidate_rejected": true,
        "private_example_scores_exposed": false
      },
      {
        "seed": 44,
        "accepted_branches": 2,
        "rejected_branches": 1,
        "accuracy_branch_accepts": true,
        "cost_branch_accepts": true,
        "non_pareto_candidate_rejected": true,
        "private_example_scores_exposed": false
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/agent-research/2609.37950-video-rsi/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "helix": {
    "schema_version": 2,
    "diagnostic_only": true,
    "manifest_ref": "reproductions:helix",
    "method": "helix",
    "dataset": "deterministic public mechanism mini-suite",
    "seeds": [
      42,
      43,
      44
    ],
    "metrics": {
      "candidate_conditioned_delta": {
        "mean": 0.11570185174544652,
        "std": 0.008884967214959426
      }
    },
    "seed_results": [
      {
        "seed": 42,
        "finite_loss": true,
        "candidate_conditioned_delta": 0.1282322257757187,
        "user_cache_unchanged": true,
        "user_gradient_finite": true
      },
      {
        "seed": 43,
        "finite_loss": true,
        "candidate_conditioned_delta": 0.10862656682729721,
        "user_cache_unchanged": true,
        "user_gradient_finite": true
      },
      {
        "seed": 44,
        "finite_loss": true,
        "candidate_conditioned_delta": 0.11024676263332367,
        "user_cache_unchanged": true,
        "user_gradient_finite": true
      }
    ],
    "evaluation_protocol": {
      "tier": "l1_mechanism",
      "seeds": [
        42,
        43,
        44
      ],
      "formal_comparison": false,
      "diagnostic_only": true,
      "claim_policy": "executable mechanism and invariant checks only"
    },
    "provenance": {
      "artifact_path": "docs/reproductions/2609.37183-helix/metrics/mechanism-seeds42-44.json",
      "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
      "original_code_commit": "working tree"
    }
  },
  "schema_version": 2,
  "manifest_ref": "experiments:sep30-p0-p1-mechanisms-seeds42-44",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "diagnostic_only": true,
    "claim_policy": "cross-paper mechanism summary; not a capability comparison"
  },
  "provenance": {
    "artifact_path": "docs/experiments/sep30-p0-p1-mechanisms-seeds42-44.json",
    "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
    "original_code_commit": "working tree"
  }
}