schema.json

{
  "schema_version": "classical-rl-trace-dataset/privacy-1.0",
  "identity": "record_id and episode_record_id remain ORIGINAL content hashes for stable lineage, not hashes of sanitized payloads. sanitized_trace_sha256 hashes the redacted trace.",
  "canonical_policy": "Select exactly one raw payload per trace ID only when unambiguous; never newest/best reward. Derived-only identities are labelled.",
  "missing": "Unavailable measurements are JSON null and empty CSV cells; unknown protocol is not inferred from date or current code.",
  "dates": "started_at_utc comes only from trace.timing.start. File modification times live only in sources.",
  "tokens": "Prompt/cache/completion/reasoning fields copied or strictly summed across saved calls. No synthetic total: prompt/cache accounting may differ by harness. Extra usage is retained in raw trace, not silently added. Missing call fields make exact sums null; separately named observed lower bounds sum known entries.",
  "scores": "All recorded reward/metric values retained. No candidate executed and no historical score recomputed under newer graders.",
  "status": "recorded_error follows trace.ok=false. It is not a model-quality zero. Completed/scored does not mean successful reproduction.",
  "calls": "All saved calls, including retained retry branches, not just active-branch calls. Original nodes and call links retained.",
  "protocol": "protocol_version plus protocol/task/archive/grader digests and agent configuration are required to assess comparability. One version label alone is insufficient.",
  "source_reconstruction": "Unavailable: source objects and archives are intentionally excluded. Source hashes/IDs describe private originals; paths are anonymized and not executable locators.",
  "tables": {
    "attempts": "one row per distinct trace payload including annotated derivatives",
    "trace_identities": "links all payload variants and optional canonical record",
    "occurrences": "every source/line/trace-index occurrence, including copied and archived records",
    "episodes": "wrapper outcomes and links to trace records, including empty-trace failures",
    "calls/nodes/tool_calls/filesystem_events/telemetry/metrics": "Redacted record_id-linked projections, consistent with redacted traces",
    "sources/archive_members": "Anonymized provenance inventory only; no archived payloads included"
  },
  "measurement_basis": "All historical numerical metrics, counts, token usage, timing and rewards describe ORIGINAL pre-redaction executions. No retokenization or rescoring."
}