{
  "rule_accuracy": [
    {
      "label": "Microsoft Decision 1",
      "correct": 285,
      "assigned": 529,
      "valid": 529
    },
    {
      "label": "Clef Flash",
      "correct": 291,
      "assigned": 529,
      "valid": 529
    },
    {
      "label": "Clef",
      "correct": 304,
      "assigned": 529,
      "valid": 529
    },
    {
      "label": "Perplexity Decider V1.1 27B",
      "correct": 459,
      "assigned": 529,
      "valid": 529
    },
    {
      "label": "Jev 1.13 (historical)",
      "correct": 377,
      "assigned": 529,
      "valid": 529
    },
    {
      "label": "Jev 1.14 (native; later collection)",
      "correct": 356,
      "assigned": 529,
      "valid": 527
    }
  ],
  "news_equal_budget": [
    {
      "label": "Historical stored significance path",
      "tp": 64,
      "review_slots": 110,
      "positives": 79,
      "provenance": "recomputed existing accepted endpoint"
    },
    {
      "label": "Stored Phase-A heuristic",
      "tp": 57,
      "review_slots": 110,
      "positives": 79,
      "provenance": "recomputed existing accepted endpoint"
    },
    {
      "label": "Perplexity Decider V1.1 27B",
      "tp": 56,
      "review_slots": 110,
      "positives": 79,
      "provenance": "recomputed existing accepted endpoint"
    },
    {
      "label": "Clef",
      "tp": 54,
      "review_slots": 110,
      "positives": 79,
      "provenance": "new post-hoc offline extension; existing raw outputs, fixed original budget and tie-break"
    },
    {
      "label": "Clef Flash",
      "tp": 52,
      "review_slots": 110,
      "positives": 79,
      "provenance": "new post-hoc offline extension; existing raw outputs, fixed original budget and tie-break"
    },
    {
      "label": "Jev 1.13 semantic v3 (historical)",
      "tp": 49,
      "review_slots": 110,
      "positives": 79,
      "provenance": "recomputed existing accepted endpoint"
    },
    {
      "label": "Microsoft Decision 1",
      "tp": 49,
      "review_slots": 110,
      "positives": 79,
      "provenance": "new post-hoc offline extension; existing raw outputs, fixed original budget and tie-break"
    }
  ],
  "news_call_latency": [
    {
      "label": "Microsoft Decision 1",
      "median_s": 0.45412917574867606,
      "p90_s": 0.5847318702377379,
      "n": 220,
      "valid": 220
    },
    {
      "label": "Clef Flash",
      "median_s": 0.47392814490012825,
      "p90_s": 1.0066096589900553,
      "n": 220,
      "valid": 220
    },
    {
      "label": "Clef",
      "median_s": 0.5642088504973799,
      "p90_s": 2.6498728608712554,
      "n": 220,
      "valid": 220
    },
    {
      "label": "Perplexity Decider V1.1 27B",
      "median_s": 0.4438094780780375,
      "p90_s": 0.5215188710950315,
      "n": 220,
      "valid": 220
    },
    {
      "label": "Jev 1.13 semantic v3 (historical)",
      "median_s": 0.141,
      "p90_s": 0.18109999999999998,
      "n": 220,
      "valid": 220
    }
  ],
  "exit_matrix": [
    {
      "label": "Microsoft Decision 1",
      "counts": [
        12,
        12,
        12
      ],
      "assigned_per_condition": 12
    },
    {
      "label": "Clef Flash",
      "counts": [
        12,
        12,
        12
      ],
      "assigned_per_condition": 12
    },
    {
      "label": "Clef",
      "counts": [
        12,
        4,
        12
      ],
      "assigned_per_condition": 12
    },
    {
      "label": "Perplexity Decider V1.1 27B",
      "counts": [
        12,
        9,
        12
      ],
      "assigned_per_condition": 12
    }
  ],
  "source_pins": {
    "accepted_plot_data_sha256": "91b2b2f041939fd04c53249d00878583098742e1b6f5605f196d1c5c1c737f97",
    "accepted_live_content_sha256": "79c95b3ec328101af06c19713101ac2a3e7c6f1ebedd2d4dce39442d9f814b83",
    "accepted_commit": "51740bbef1a1fe031d8cb38a127eb40b468b64f7"
  },
  "scope": "Aggregates only; no raw identifiers. Reading 60/60 on all six tracks. News historical cohort: 220, 79 positives, 110 slots. Exit follow-up: 12 fixtures per condition; not the earlier eight-seed design. Timing: 220 valid/timed each; October hosted HTTP versus September historical SDK; no controlled speed comparison."
}