{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://sediment.so/schemas/training-rows/dpo-metadata/v4.json",
  "title": "DPOMetadata",
  "description": "Sediment evidence for one DPO pair.",
  "type": "object",
  "properties": {
    "org_id": {
      "type": "string",
      "description": "The deployment's tenant id, normalized. Bound to the server."
    },
    "source_model": {
      "type": "string",
      "description": "The model that served the source inference call."
    },
    "chosen_completion_id": {
      "type": "string",
      "description": "The preferred inference call."
    },
    "rejected_completion_id": {
      "type": "string",
      "description": "The dispreferred inference call from the same prompt and model."
    },
    "recipe_id": {
      "enum": [
        "dpo_human",
        "dpo_outcome"
      ],
      "type": "string",
      "description": "The closed evidence recipe that produced this training row."
    },
    "recipe_version": {
      "const": 2,
      "type": "integer",
      "description": "The version of the evidence recipe; increment it when recipe semantics change."
    },
    "chosen_label_source": {
      "enum": [
        "explicit_accept",
        "resolved_ci_pass"
      ],
      "type": "string",
      "description": "The closed evidence source that labeled the preferred member."
    },
    "rejected_label_source": {
      "enum": [
        "explicit_reject",
        "resolved_ci_fail"
      ],
      "type": "string",
      "description": "The closed evidence source that labeled the dispreferred member."
    },
    "label_confidence": {
      "type": "number",
      "description": "The weaker member confidence — how much to trust the pair."
    },
    "ci_reliability": {
      "anyOf": [
        {
          "type": "number"
        },
        {
          "type": "null"
        }
      ],
      "description": "Trust in the resolved CI evidence, separate from the categorical label."
    },
    "confidence_margin": {
      "type": "number",
      "description": "Chosen confidence minus rejected. Can be negative when attribution discounts or configured confidence factors outweigh direction."
    },
    "provenance": {
      "$ref": "#/$defs/DPOProvenance",
      "description": "The structured provenance for this artifact."
    },
    "split": {
      "enum": [
        "train",
        "eval"
      ],
      "type": "string",
      "description": "Which side of the deterministic eval split this row is on, hashed on `session_id` so a session never straddles both."
    },
    "chosen_repository_identity": {
      "anyOf": [
        {
          "$ref": "#/$defs/RepositoryIdentity"
        },
        {
          "type": "null"
        }
      ],
      "description": "Qualified repository identity of the chosen member; null for an unambiguous legacy repository."
    },
    "rejected_repository_identity": {
      "anyOf": [
        {
          "$ref": "#/$defs/RepositoryIdentity"
        },
        {
          "type": "null"
        }
      ],
      "description": "Qualified repository identity of the rejected member; independent of the chosen member."
    },
    "chosen_attribution_source": {
      "anyOf": [
        {
          "enum": [
            "git_notes",
            "jaccard"
          ],
          "type": "string"
        },
        {
          "type": "null"
        }
      ],
      "description": "The selected preferred member's inferred Attribution method, or null without Attribution."
    },
    "rejected_attribution_source": {
      "anyOf": [
        {
          "enum": [
            "git_notes",
            "jaccard"
          ],
          "type": "string"
        },
        {
          "type": "null"
        }
      ],
      "description": "The selected dispreferred member's inferred Attribution method, or null without Attribution."
    },
    "chosen_session_commit_observation_ids": {
      "type": "array",
      "items": {
        "type": "string",
        "pattern": "^[^\\u0000\\uD800-\\uDFFF]*$"
      },
      "description": "Sorted unique observation IDs for the preferred member's exact Session-to-commit edge; empty when absent."
    },
    "rejected_session_commit_observation_ids": {
      "type": "array",
      "items": {
        "type": "string",
        "pattern": "^[^\\u0000\\uD800-\\uDFFF]*$"
      },
      "description": "Sorted unique observation IDs for the dispreferred member's exact Session-to-commit edge; empty when absent."
    },
    "schema_id": {
      "const": "https://sediment.so/schemas/training-rows/dpo-pair/v4.json",
      "type": "string",
      "description": "The canonical wire-shape contract for this row."
    },
    "schema_version": {
      "const": 4,
      "type": "integer",
      "description": "The positive integer version of the canonical row schema."
    }
  },
  "additionalProperties": false,
  "required": [
    "org_id",
    "source_model",
    "chosen_completion_id",
    "rejected_completion_id",
    "recipe_id",
    "recipe_version",
    "chosen_label_source",
    "rejected_label_source",
    "label_confidence",
    "ci_reliability",
    "confidence_margin",
    "provenance",
    "split",
    "chosen_repository_identity",
    "rejected_repository_identity",
    "chosen_attribution_source",
    "rejected_attribution_source",
    "chosen_session_commit_observation_ids",
    "rejected_session_commit_observation_ids",
    "schema_id",
    "schema_version"
  ],
  "$defs": {
    "DPOProvenance": {
      "type": "object",
      "properties": {
        "chosen": {
          "$ref": "#/$defs/Provenance",
          "description": "Provenance for the preferred member."
        },
        "rejected": {
          "$ref": "#/$defs/Provenance",
          "description": "Provenance for the dispreferred member."
        }
      },
      "additionalProperties": false,
      "required": [
        "chosen",
        "rejected"
      ]
    },
    "Provenance": {
      "type": "object",
      "properties": {
        "policy_version": {
          "type": "string",
          "description": "The derivation implementation policy version."
        },
        "quarantine_revision": {
          "type": "integer",
          "description": "The integer quarantine-log high-water mark used for the derivation."
        },
        "policy_digest": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "description": "SHA-256 of the fully resolved derivation policy when available."
        }
      },
      "additionalProperties": false,
      "required": [
        "policy_version",
        "quarantine_revision",
        "policy_digest"
      ]
    },
    "RepositoryIdentity": {
      "type": "object",
      "properties": {
        "provider": {
          "const": "github",
          "type": "string",
          "description": "The forge provider that assigns the immutable repository ID."
        },
        "host": {
          "type": "string",
          "pattern": "^[^\\u0000\\uD800-\\uDFFF]*$",
          "description": "The lowercase configured forge host."
        },
        "repository_id": {
          "type": "string",
          "pattern": "^[^\\u0000\\uD800-\\uDFFF]*$",
          "description": "Immutable provider repository ID within its forge host."
        }
      },
      "additionalProperties": false,
      "required": [
        "provider",
        "host",
        "repository_id"
      ]
    }
  }
}