{
  "schema": "paydirt.model/1",
  "release": "2026-10",
  "run_id": "2026-10",
  "slug": "google/gemini-3.1-pro-preview",
  "outcomes": [
    {
      "case": "ch-02-a",
      "pair": "ch-02",
      "family": "challenge",
      "variant": "accurate",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "supported",
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": true,
      "usd": 0.053056,
      "wall_s": 92,
      "tokens_in": 11522,
      "tokens_out": 2501,
      "tokens_reasoning": 2290,
      "turns": 4,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-02-o",
      "pair": "ch-02",
      "family": "challenge",
      "variant": "overclaimed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "unsupported",
      "quote_hit": true,
      "max_severity": "critical",
      "max_severity_exact": false,
      "usd": 0.076306,
      "wall_s": 92.1,
      "tokens_in": 11681,
      "tokens_out": 4412,
      "tokens_reasoning": 4090,
      "turns": 4,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-05-a",
      "pair": "ch-05",
      "family": "challenge",
      "variant": "accurate",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "supported",
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": true,
      "usd": 0.108544,
      "wall_s": 57.9,
      "tokens_in": 10280,
      "tokens_out": 7332,
      "tokens_reasoning": 7136,
      "turns": 3,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-05-o",
      "pair": "ch-05",
      "family": "challenge",
      "variant": "overclaimed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "unsupported",
      "quote_hit": true,
      "max_severity": "critical",
      "max_severity_exact": false,
      "usd": 0.318594,
      "wall_s": 401.1,
      "tokens_in": 47718,
      "tokens_out": 20935,
      "tokens_reasoning": 20329,
      "turns": 12,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-06-a",
      "pair": "ch-06",
      "family": "challenge",
      "variant": "accurate",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 3,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "supported",
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": true,
      "usd": 0.15083,
      "wall_s": 72.5,
      "tokens_in": 13387,
      "tokens_out": 10338,
      "tokens_reasoning": 9952,
      "turns": 3,
      "effort": "high",
      "providers": [],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-06-o",
      "pair": "ch-06",
      "family": "challenge",
      "variant": "overclaimed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 3,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "unsupported",
      "quote_hit": true,
      "max_severity": "unrated",
      "max_severity_exact": false,
      "usd": 0.1605388,
      "wall_s": 238,
      "tokens_in": 25745,
      "tokens_out": 10276,
      "tokens_reasoning": 9778,
      "turns": 7,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-07-a",
      "pair": "ch-07",
      "family": "challenge",
      "variant": "accurate",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "supported",
      "quote_hit": null,
      "max_severity": "medium",
      "max_severity_exact": true,
      "usd": 0.1160688,
      "wall_s": 194.4,
      "tokens_in": 20994,
      "tokens_out": 6762,
      "tokens_reasoning": 6519,
      "turns": 6,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ch-07-o",
      "pair": "ch-07",
      "family": "challenge",
      "variant": "overclaimed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 0,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "overclaimed",
      "quote_hit": true,
      "max_severity": "high",
      "max_severity_exact": false,
      "usd": 0.3424492,
      "wall_s": 169.2,
      "tokens_in": 55376,
      "tokens_out": 22831,
      "tokens_reasoning": 22360,
      "turns": 7,
      "effort": "high",
      "providers": [],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-01-f",
      "pair": "hard-01",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 3,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.0386782,
      "wall_s": 1802.6,
      "tokens_in": 34485,
      "tokens_out": 2224,
      "tokens_reasoning": 187,
      "turns": 28,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-01-v",
      "pair": "hard-01",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": false,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 3,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.259954,
      "wall_s": 143.7,
      "tokens_in": 12833,
      "tokens_out": 19524,
      "tokens_reasoning": 19306,
      "turns": 3,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-02-f",
      "pair": "hard-02",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.017202,
      "wall_s": 1805.7,
      "tokens_in": 14286,
      "tokens_out": 245,
      "tokens_reasoning": 122,
      "turns": 28,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-02-v",
      "pair": "hard-02",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.3619158,
      "wall_s": 447,
      "tokens_in": 38178,
      "tokens_out": 26171,
      "tokens_reasoning": 25892,
      "turns": 9,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-03-f",
      "pair": "hard-03",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 3,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.1119208,
      "wall_s": 1801.2,
      "tokens_in": 23796,
      "tokens_out": 9453,
      "tokens_reasoning": 8209,
      "turns": 26,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-03-v",
      "pair": "hard-03",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 1,
      "primary_all": null,
      "bite": null,
      "decoys": 3,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.0135526,
      "wall_s": 1801.3,
      "tokens_in": 39483,
      "tokens_out": 5626,
      "tokens_reasoning": 172,
      "turns": 24,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-04-f",
      "pair": "hard-04",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.043892,
      "wall_s": 1801.3,
      "tokens_in": 49902,
      "tokens_out": 7339,
      "tokens_reasoning": 1190,
      "turns": 24,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-04-v",
      "pair": "hard-04",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.3326786,
      "wall_s": 1401.2,
      "tokens_in": 39844,
      "tokens_out": 28233,
      "tokens_reasoning": 25089,
      "turns": 18,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-07-a",
      "pair": "hard-07",
      "family": "challenge",
      "variant": "accurate",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "supported",
      "quote_hit": null,
      "max_severity": "high",
      "max_severity_exact": true,
      "usd": 0.10668,
      "wall_s": 125.3,
      "tokens_in": 17910,
      "tokens_out": 5905,
      "tokens_reasoning": 5678,
      "turns": 4,
      "effort": "high",
      "providers": [],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-07-o",
      "pair": "hard-07",
      "family": "challenge",
      "variant": "overclaimed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "unsupported",
      "quote_hit": true,
      "max_severity": "high",
      "max_severity_exact": false,
      "usd": 0.294102,
      "wall_s": 145.4,
      "tokens_in": 17493,
      "tokens_out": 21593,
      "tokens_reasoning": 20341,
      "turns": 3,
      "effort": "high",
      "providers": [],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-08-a",
      "pair": "hard-08",
      "family": "challenge",
      "variant": "accurate",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 1,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "supported",
      "quote_hit": null,
      "max_severity": "medium",
      "max_severity_exact": true,
      "usd": 0.055108,
      "wall_s": 76.8,
      "tokens_in": 20516,
      "tokens_out": 1173,
      "tokens_reasoning": 932,
      "turns": 5,
      "effort": "high",
      "providers": [],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "hard-08-o",
      "pair": "hard-08",
      "family": "challenge",
      "variant": "overclaimed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [],
      "planted": 1,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": "overclaimed",
      "quote_hit": true,
      "max_severity": "medium",
      "max_severity_exact": true,
      "usd": 0.3306178,
      "wall_s": 270.3,
      "tokens_in": 202508,
      "tokens_out": 16018,
      "tokens_reasoning": 15406,
      "turns": 15,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-02-f",
      "pair": "sol-02",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.017428,
      "wall_s": 1801.4,
      "tokens_in": 7772,
      "tokens_out": 157,
      "tokens_reasoning": 41,
      "turns": 29,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-02-v",
      "pair": "sol-02",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.1327746,
      "wall_s": 183.2,
      "tokens_in": 13860,
      "tokens_out": 9331,
      "tokens_reasoning": 9101,
      "turns": 6,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-03-f",
      "pair": "sol-03",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 1,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.029098,
      "wall_s": 1801.5,
      "tokens_in": 32940,
      "tokens_out": 4899,
      "tokens_reasoning": 129,
      "turns": 24,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-03-v",
      "pair": "sol-03",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 1,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.1358112,
      "wall_s": 1026,
      "tokens_in": 27084,
      "tokens_out": 7994,
      "tokens_reasoning": 7759,
      "turns": 17,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-05-f",
      "pair": "sol-05",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.0248736,
      "wall_s": 1801.4,
      "tokens_in": 17529,
      "tokens_out": 321,
      "tokens_reasoning": 186,
      "turns": 32,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-05-v",
      "pair": "sol-05",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": false,
          "severity_distance": 1
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.2149194,
      "wall_s": 637.3,
      "tokens_in": 10197,
      "tokens_out": 16798,
      "tokens_reasoning": 16565,
      "turns": 11,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-09-f",
      "pair": "sol-09",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.0486524,
      "wall_s": 1812.2,
      "tokens_in": 33239,
      "tokens_out": 5396,
      "tokens_reasoning": 199,
      "turns": 27,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-09-v",
      "pair": "sol-09",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "failed",
      "failure": "error",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 1,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.006038,
      "wall_s": 354.2,
      "tokens_in": 6984,
      "tokens_out": 1131,
      "tokens_reasoning": 102,
      "turns": 7,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "error",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-10-f",
      "pair": "sol-10",
      "family": "find-sol",
      "variant": "fixed",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": false,
      "decoys": 1,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "unrated",
      "max_severity_exact": null,
      "usd": 0.3227196,
      "wall_s": 1379.4,
      "tokens_in": 10323,
      "tokens_out": 25762,
      "tokens_reasoning": 25664,
      "turns": 21,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-10-v",
      "pair": "sol-10",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 1,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.239522,
      "wall_s": 141.9,
      "tokens_in": 8167,
      "tokens_out": 18599,
      "tokens_reasoning": 18136,
      "turns": 3,
      "effort": "high",
      "providers": [],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-11-f",
      "pair": "sol-11",
      "family": "find-sol",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.1147524,
      "wall_s": 1810.5,
      "tokens_in": 15639,
      "tokens_out": 7537,
      "tokens_reasoning": 7398,
      "turns": 28,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "sol-11-v",
      "pair": "sol-11",
      "family": "find-sol",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": false,
          "severity_distance": 1
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.2221972,
      "wall_s": 1030.2,
      "tokens_in": 17441,
      "tokens_out": 16779,
      "tokens_reasoning": 16563,
      "turns": 16,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ts-02-f",
      "pair": "ts-02",
      "family": "find-ts",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.025558,
      "wall_s": 1801.3,
      "tokens_in": 19514,
      "tokens_out": 2671,
      "tokens_reasoning": 143,
      "turns": 27,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ts-02-v",
      "pair": "ts-02",
      "family": "find-ts",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": true,
          "severity": "critical",
          "severity_exact": true,
          "severity_distance": 0
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.2431192,
      "wall_s": 408.3,
      "tokens_in": 40253,
      "tokens_out": 15904,
      "tokens_reasoning": 15626,
      "turns": 11,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ts-03-f",
      "pair": "ts-03",
      "family": "find-ts",
      "variant": "fixed",
      "status": "failed",
      "failure": "timeout",
      "correct": false,
      "findings_read": 0,
      "hits": [],
      "planted": 0,
      "primary_all": null,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": null,
      "max_severity_exact": null,
      "usd": 0.0169454,
      "wall_s": 1802.3,
      "tokens_in": 14161,
      "tokens_out": 244,
      "tokens_reasoning": 133,
      "turns": 26,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "aborted",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    },
    {
      "case": "ts-03-v",
      "pair": "ts-03",
      "family": "find-ts",
      "variant": "vulnerable",
      "status": "ok",
      "failure": null,
      "correct": true,
      "findings_read": 1,
      "hits": [
        {
          "id": "P1",
          "primary": true,
          "finding": 0,
          "by_line": true,
          "by_function": false,
          "severity": "critical",
          "severity_exact": false,
          "severity_distance": 1
        }
      ],
      "planted": 1,
      "primary_all": true,
      "bite": null,
      "decoys": 2,
      "decoy_bites": 0,
      "unmatched": 0,
      "verdict": null,
      "quote_hit": null,
      "max_severity": "critical",
      "max_severity_exact": null,
      "usd": 0.221496,
      "wall_s": 531,
      "tokens_in": 13800,
      "tokens_out": 17349,
      "tokens_reasoning": 17112,
      "turns": 8,
      "effort": "high",
      "providers": [
        "Google"
      ],
      "stop": "stop",
      "arm": "raw",
      "rep": 1,
      "infra_retries": 0
    }
  ]
}