{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "SemiHS-Bench Record",
  "version": "2.0.0",
  "description": "Schema for a single SemiHS-Bench benchmark record. Each record carries two model-facing input tiers (tier1_description and tier2_minimal), the gold HS6/HS4/HS2 labels, a fixed candidate slate, source-type metadata, hardness tags, and a human-readable justification.",
  "type": "object",
  "required": [
    "id",
    "frozen_id",
    "split",
    "tier1_description",
    "tier1_source",
    "tier2_minimal",
    "tier2_classifiable",
    "hs6_label",
    "hs4_label",
    "hs2_label",
    "segment",
    "confidence_tier",
    "difficulty_tags",
    "boundary_note",
    "justification_text",
    "bol_metadata",
    "candidate_set",
    "reference_ruling_ids"
  ],
  "properties": {
    "id": {
      "type": "string",
      "comment": "Released ids are SH-<digits>. Records still moving through ingest carry a provisional source-prefixed form SH-<source>-<digits> until the release build renumbers them.",
      "pattern": "^SH-([a-z]+-)?\\d{4,}(\\.[a-z_]+)?$"
    },
    "frozen_id": {
      "type": "string",
      "pattern": "^v2\\.0\\.(eval|train)\\.\\d{4,}$",
      "description": "Stable release identifier: v2.0.<split>.<ordinal>."
    },
    "split": {
      "type": "string",
      "enum": [
        "eval",
        "train"
      ]
    },
    "tier1_description": {
      "type": "string",
      "minLength": 10,
      "description": "Full natural-language product description, written to read like real trade or catalog text (typically 60-500 chars, median around 180)."
    },
    "tier1_source": {
      "type": "string",
      "enum": [
        "BOL",
        "catalog"
      ],
      "description": "Only BOL and manufacturer-catalog sources contribute benchmark records; customs rulings live in reference_corpus.jsonl as supporting evidence."
    },
    "tier2_minimal": {
      "type": "object",
      "required": [
        "part_name",
        "manufacturer"
      ],
      "additionalProperties": false,
      "properties": {
        "part_name": {
          "type": "string",
          "minLength": 1
        },
        "manufacturer": {
          "type": "string"
        }
      },
      "description": "The sparse ERP-style Tier 2 input: a manufacturer part number or short descriptor plus the manufacturer name."
    },
    "tier2_classifiable": {
      "type": "string",
      "enum": [
        "yes",
        "no",
        "partial"
      ],
      "description": "Expert judgment of whether the gold HS6 can be reached from the Tier 2 input alone."
    },
    "hs6_label": {
      "type": "string",
      "pattern": "^\\d{6}$"
    },
    "hs4_label": {
      "type": "string",
      "pattern": "^\\d{4}$"
    },
    "hs2_label": {
      "type": "string",
      "pattern": "^\\d{2}$"
    },
    "segment": {
      "type": "string",
      "enum": [
        "material",
        "equipment",
        "metrology",
        "component",
        "end_product"
      ],
      "description": "Supply-chain segment by OECD production stage. 'material' = process inputs, substrates, gases, chemicals, photoresist; 'equipment' = fabrication machinery, furnaces, dedicated equipment parts; 'metrology' = measurement, inspection, test instrumentation; 'component' = semiconductor devices, electronic components, optics; 'end_product' = finished, semiconductor-intensive downstream goods. See data/taxonomy.csv for the per-HS6 mapping."
    },
    "confidence_tier": {
      "type": "string",
      "enum": [
        "high",
        "medium",
        "low"
      ],
      "description": "Confidence in the gold label. For records grounded in cited rulings: high = \u22652 cross-jurisdiction citations, medium = \u22651, low = none. For catalog records the tier reflects the expert reviewer's assessed confidence."
    },
    "difficulty_tags": {
      "type": "array",
      "uniqueItems": true,
      "items": {
        "type": "string",
        "enum": [
          "8541_siblings",
          "8542_ic_function",
          "8486_process_stage",
          "8504_power_splits",
          "8536_connection_splits",
          "9030_measurement_splits",
          "9031_inspection_splits",
          "2804_gas_purity",
          "3707_photochemical",
          "9027_analysis_splits",
          "8471_adp_splits",
          "8541_vs_8542",
          "populated_board_boundary",
          "storage_boundary",
          "doped_vs_undoped",
          "process_vs_metrology",
          "furnace_boundary",
          "machine_with_function",
          "parts_attribution",
          "led_device_vs_luminaire",
          "display_module_boundary",
          "amplifier_boundary",
          "cable_vs_connector",
          "sensor_boundary",
          "crystal_substrate_boundary"
        ]
      },
      "description": "IDs of the confusable HS6 boundaries this record sits on, drawn from a closed vocabulary of within-family sibling splits and cross-family frontiers (see construction/configs/boundary_tags.yaml). Empty when the record is not on a contested boundary."
    },
    "boundary_note": {
      "type": "string",
      "description": "Human-readable comment stating the deciding criterion for each boundary in difficulty_tags. Empty string exactly when difficulty_tags is empty."
    },
    "justification_text": {
      "type": "string",
      "description": "Long-form justification carried over from the source ruling or expert annotation; not size-bounded. For released records, this is typically derived from the cited reference-corpus rationale_excerpt or expert-written extension."
    },
    "bol_metadata": {
      "type": [
        "object",
        "null"
      ],
      "description": "BOL-source metadata (null for catalog-source records).",
      "properties": {
        "shipper": {
          "type": "string"
        },
        "consignee": {
          "type": "string"
        },
        "port_origin": {
          "type": "string"
        },
        "port_dest": {
          "type": "string"
        },
        "bol_description": {
          "type": "string"
        },
        "declared_hs": {
          "type": "string"
        },
        "hs_verified": {
          "type": "boolean"
        },
        "arrival_date": {
          "type": "string"
        },
        "origin_country": {
          "type": "string"
        }
      }
    },
    "candidate_set": {
      "type": "object",
      "required": [
        "size",
        "codes",
        "gold_rank_in_candidates"
      ],
      "additionalProperties": true,
      "properties": {
        "size": {
          "type": "integer",
          "minimum": 2,
          "maximum": 10
        },
        "codes": {
          "type": "array",
          "items": {
            "type": "string",
            "pattern": "^\\d{6}$"
          },
          "minItems": 2,
          "maxItems": 10,
          "uniqueItems": true
        },
        "gold_rank_in_candidates": {
          "type": "integer",
          "minimum": 0
        }
      }
    },
    "reference_ruling_ids": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "Evidence IDs of rulings in data/reference_corpus.jsonl that classify comparable goods under this record's gold HS6, kept when they share descriptive keywords with the record text and ordered by that overlap. When no same-HS6 ruling shares a keyword (common for rulings in another language), all same-HS6 rulings are listed instead. Empty when the corpus has no ruling for that code."
    }
  }
}