metadata/schema.json

{
  "format": {
    "with_answers_json": "JSON array of public task records with answer/clue artifacts",
    "with_answers_jsonl": "one public task record per line; ground_truth and clues are JSON strings for Arrow compatibility",
    "questions_only_json": "JSON array of model-facing task records",
    "questions_only_jsonl": "one model-facing task record per line"
  },
  "split": "test",
  "primary_files": [
    "data/wildtrace_strict481.questions_only.jsonl",
    "data/wildtrace_strict481.with_answers.jsonl"
  ],
  "task_fields": {
    "question_id": "opaque stable public task id",
    "paradigm": "primary evidence geometry label",
    "context_tier": "coarse document-length tier",
    "doc_token_estimate": "document token estimate used for release stratification",
    "corpus_id": "opaque stable public source id",
    "corpus_file": "relative path to source document text",
    "question_text": "model-facing public question",
    "freeze_id": "public release freeze id",
    "release_version": "public release version",
    "ground_truth": "with-answers only: question_text, summary_answer, key_facts, scoring_rubric",
    "clues": "with-answers only: source evidence spans and passages"
  },
  "evaluation_note": "Models under test receive only corpus_file and question_text. ground_truth and clues are withheld during evaluation."
}