{
  "schema_version": "v0",
  "submitted_at": "2026-04-25T01:47:03Z",
  "track": "open-swap",
  "reader": "Llama-3.1-8B-Instruct",
  "model": "meta-llama/Llama-3.1-8B-Instruct",
  "params": "8B",
  "pipeline": "Stack (BGE-large UNION QNDN v0 UNION BM25 -> RRF k=60 -> MixK rerank -> top-10)",
  "evidence": {
    "artifact": "frozen_retrieval_topK_500q.v1.jsonl",
    "artifact_sha256": "0000000000000000000000000000000000000000000000000000000000000000",
    "top_k": 10,
    "r_at_5": 0.962
  },
  "judge": {
    "model": "gpt-4o",
    "n_seeds": 5,
    "threshold": 3,
    "protocol": "K=5 3-of-5 majority vote"
  },
  "overall_acc": 0.612,
  "overall_n": 500,
  "ci_lo": 0.5685,
  "ci_hi": 0.6541,
  "refusal_rate": 0.196,
  "qtype_acc": {
    "SSA": {"acc": 0.821, "n": 56},
    "MSA": {"acc": 0.620, "n": 133},
    "TR":  {"acc": 0.519, "n": 133},
    "KU":  {"acc": 0.718, "n": 78},
    "MR":  {"acc": 0.300, "n": 30}
  },
  "notes": "Example only. Stack pipeline (no qtype routing, no full-context fallback). Frozen retrieval contract; reader is the only swapped variable. SHA-256 of evidence artifact replaced with zeros for example purposes; real submissions must verify against the published manifest hash."
}