{
  "schema": "asn.public-cuad-llama-repeatability.v1",
  "study_id": "cuad-llama-six-contract-output-repeatability-2026-10-27",
  "title": "Llama 3.2 3B · three-run CUAD output repeatability",
  "source": {
    "name": "Contract Understanding Atticus Dataset (CUAD) v1",
    "license": "CC BY 4.0",
    "task_count": 6,
    "distinct_contracts": 6,
    "suite_sha256": "611a4e3fbfacc9dc67f16d462c5d8fcb1953e30d140926f0226d0357f614150c"
  },
  "model": {
    "name": "llama3.2:3b",
    "digest": "a80c4f17acd55265feec403c7aef86be0c25983ab279d83f3bcd3abbcb5b8b72",
    "quantization": "Q4_K_M"
  },
  "protocol": {
    "prompt_variants": [
      "canonical",
      "paraphrase"
    ],
    "repetitions_per_task_per_variant": 2,
    "attempts_per_run": 24,
    "parallelism": 1,
    "temperature": 0,
    "seed": 42,
    "request_timeout_ms": 60000,
    "runtime": "Ollama 0.33.3 on local loopback only; no external provider calls or provider spend."
  },
  "runs": [
    {
      "run_artifact_sha256": "5ab98cdef9607e797f62766f5db697e57af3cf21338d4cecd194fe1d6a4b7847",
      "captured_at": "2026-10-27T00:07:18.184Z",
      "attempts": 24,
      "exact_matches": 2,
      "verified_mismatches": 22,
      "execution_errors": 0
    },
    {
      "run_artifact_sha256": "00e3e9ef339976ec0dafff9d375bc20ca57b8317cc425c79786623ef52250390",
      "captured_at": "2026-10-27T07:22:46.022Z",
      "attempts": 24,
      "exact_matches": 2,
      "verified_mismatches": 22,
      "execution_errors": 0
    },
    {
      "run_artifact_sha256": "9398275d5accb152b51e2f14424a0051a40b2ef87d63f740d483acb61deb62a7",
      "captured_at": "2026-10-27T21:55:16.446Z",
      "attempts": 24,
      "exact_matches": 2,
      "verified_mismatches": 22,
      "execution_errors": 0
    }
  ],
  "pairwise_repeatability": [
    {
      "left_run": "5ab98cdef9607e797f62766f5db697e57af3cf21338d4cecd194fe1d6a4b7847",
      "right_run": "00e3e9ef339976ec0dafff9d375bc20ca57b8317cc425c79786623ef52250390",
      "matching_output_hashes": 24,
      "matching_verdicts": 24
    },
    {
      "left_run": "5ab98cdef9607e797f62766f5db697e57af3cf21338d4cecd194fe1d6a4b7847",
      "right_run": "9398275d5accb152b51e2f14424a0051a40b2ef87d63f740d483acb61deb62a7",
      "matching_output_hashes": 24,
      "matching_verdicts": 24
    },
    {
      "left_run": "00e3e9ef339976ec0dafff9d375bc20ca57b8317cc425c79786623ef52250390",
      "right_run": "9398275d5accb152b51e2f14424a0051a40b2ef87d63f740d483acb61deb62a7",
      "matching_output_hashes": 24,
      "matching_verdicts": 24
    }
  ],
  "interpretation": {
    "statement": "Three separate local executions of the same six-task protocol returned identical output hashes and exact-verification verdicts for all 24 slots. Each run produced 2 exact annotation matches, 22 verified mismatches and no execution errors.",
    "scope": "This is output repeatability for one quantized local model build, prompt suite and runtime configuration. It adds no new task coverage and does not establish legal correctness, general model capability, hosted behavior or customer utility.",
    "public_ranking_eligible": false,
    "production_index_eligible": false
  },
  "evidence": {
    "verification_method": "independent_exact_holdout_oracle.v1",
    "all_runs_independently_verified": true,
    "all_pairwise_output_hashes_match": true,
    "all_pairwise_verdicts_match": true,
    "raw_prompts_published": false,
    "contract_excerpts_published": false,
    "raw_model_outputs_published": false
  },
  "semantics": {
    "local_research_only": true,
    "external_provider_calls": false,
    "provider_spend_usd": 0,
    "production_measurement_written": false,
    "production_routing_changed": false,
    "ranking_eligible": false,
    "authority_expansion": false
  },
  "receipt_sha256": "0bef28c8ae6564c0c689982025fabce96673917e48944577dcb91c0195e034b9"
}