{ "schema": "paydirt.model/1", "release": "2026-10", "run_id": "2026-10", "slug": "openai/gpt-6-luna", "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": 0.005380905, "wall_s": 75.5, "tokens_in": 26340, "tokens_out": 6761, "tokens_reasoning": 6435, "turns": 4, "effort": "max", "providers": [], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "ch-02-a", "pair": "ch-02", "family": "challenge", "variant": "accurate", "status": "ok", "failure": null, "correct": false, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": null, "decoys": 0, "decoy_bites": 0, "unmatched": 0, "verdict": "overclaimed", "quote_hit": null, "max_severity": "critical", "max_severity_exact": true, "usd": 0.027889445, "wall_s": 453, "tokens_in": 19113, "tokens_out": 52868, "tokens_reasoning": 51432, "turns": 3, "effort": "max", "providers": [], "stop": "stop", "arm": "report", "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": "unrated", "max_severity_exact": false, "usd": 0.014723745, "wall_s": 236.3, "tokens_in": 23267, "tokens_out": 26240, "tokens_reasoning": 25487, "turns": 4, "effort": "max", "providers": [], "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": 0.032051535, "wall_s": 605.9, "tokens_in": 18775, "tokens_out": 61227, "tokens_reasoning": 58717, "turns": 3, "effort": "max", "providers": [], "stop": "stop", "arm": "report", "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": 0.002415465, "wall_s": 33.7, "tokens_in": 10510, "tokens_out": 2881, "tokens_reasoning": 2647, "turns": 3, "effort": "max", "providers": [], "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": 0.019442025, "wall_s": 287, "tokens_in": 18072, "tokens_out": 36195, "tokens_reasoning": 34696, "turns": 3, "effort": "max", "providers": [], "stop": "stop", "arm": "report", "rep": 1, "infra_retries": 0 }, { "case": "ch-05-o", "pair": "ch-05", "family": "challenge", "variant": "overclaimed", "status": "ok", "failure": null, "correct": true, "findings_read": 2, "hits": [], "planted": 0, "primary_all": null, "bite": null, "decoys": 0, "decoy_bites": 0, "unmatched": 0, "verdict": "overclaimed", "quote_hit": true, "max_severity": "critical", "max_severity_exact": false, "usd": 0.014643195, "wall_s": 303.6, "tokens_in": 10977, "tokens_out": 27225, "tokens_reasoning": 26737, "turns": 3, "effort": "max", "providers": [], "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": 2, "hits": [], "planted": 0, "primary_all": null, "bite": null, "decoys": 0, "decoy_bites": 0, "unmatched": 0, "verdict": "overclaimed", "quote_hit": true, "max_severity": "critical", "max_severity_exact": false, "usd": 0.023416195, "wall_s": 343.2, "tokens_in": 18956, "tokens_out": 43978, "tokens_reasoning": 42054, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "rep": 1, "infra_retries": 0 }, { "case": "ch-06-a", "pair": "ch-06", "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": 0.00580242, "wall_s": 78.2, "tokens_in": 28244, "tokens_out": 7505, "tokens_reasoning": 7192, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "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": false, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": null, "decoys": 0, "decoy_bites": 0, "unmatched": 0, "verdict": "overclaimed", "quote_hit": null, "max_severity": "critical", "max_severity_exact": true, "usd": 0.028138195, "wall_s": 455.5, "tokens_in": 20846, "tokens_out": 52938, "tokens_reasoning": 51287, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "rep": 1, "infra_retries": 0 }, { "case": "ch-06-o", "pair": "ch-06", "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": "unrated", "max_severity_exact": false, "usd": 0.019512895, "wall_s": 333.4, "tokens_in": 13019, "tokens_out": 36447, "tokens_reasoning": 35802, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "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": 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": 0.027669085, "wall_s": 453.4, "tokens_in": 20668, "tokens_out": 52013, "tokens_reasoning": 49661, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "rep": 1, "infra_retries": 0 }, { "case": "ch-07-a", "pair": "ch-07", "family": "challenge", "variant": "accurate", "status": "ok", "failure": null, "correct": false, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": null, "decoys": 0, "decoy_bites": 0, "unmatched": 0, "verdict": "overclaimed", "quote_hit": null, "max_severity": "high", "max_severity_exact": false, "usd": 0.018063595, "wall_s": 365.4, "tokens_in": 11141, "tokens_out": 34011, "tokens_reasoning": 33710, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "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": 0.024374255, "wall_s": 454.7, "tokens_in": 19085, "tokens_out": 45846, "tokens_reasoning": 44417, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "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": "medium", "max_severity_exact": true, "usd": 0.02408238, "wall_s": 561.3, "tokens_in": 11400, "tokens_out": 45999, "tokens_reasoning": 45361, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "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": "medium", "max_severity_exact": true, "usd": 0.033488575, "wall_s": 592, "tokens_in": 70799, "tokens_out": 53625, "tokens_reasoning": 51556, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "rep": 1, "infra_retries": 0 }, { "case": "hard-01-f", "pair": "hard-01", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.011516215, "wall_s": 218.2, "tokens_in": 18839, "tokens_out": 20246, "tokens_reasoning": 19971, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-01-f", "pair": "hard-01", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.018526145, "wall_s": 273.4, "tokens_in": 39981, "tokens_out": 32699, "tokens_reasoning": 31583, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": true, "severity": "high", "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": "high", "max_severity_exact": null, "usd": 0.007231815, "wall_s": 120.4, "tokens_in": 30499, "tokens_out": 10501, "tokens_reasoning": 10093, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": true, "severity": "high", "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": "high", "max_severity_exact": null, "usd": 0.040327495, "wall_s": 606.1, "tokens_in": 147085, "tokens_out": 60934, "tokens_reasoning": 58358, "turns": 7, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "hard-02-f", "pair": "hard-02", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.012326855, "wall_s": 203.5, "tokens_in": 30667, "tokens_out": 19961, "tokens_reasoning": 19730, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-02-f", "pair": "hard-02", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.02754332, "wall_s": 477, "tokens_in": 59395, "tokens_out": 46042, "tokens_reasoning": 44296, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.007568865, "wall_s": 125, "tokens_in": 26755, "tokens_out": 11406, "tokens_reasoning": 11108, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-02-v", "pair": "hard-02", "family": "find-sol", "variant": "vulnerable", "status": "ok", "failure": null, "correct": false, "findings_read": 0, "hits": [], "planted": 1, "primary_all": false, "bite": null, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.02405218, "wall_s": 326.6, "tokens_in": 55486, "tokens_out": 40005, "tokens_reasoning": 38522, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "hard-03-f", "pair": "hard-03", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.0068392, "wall_s": 124.3, "tokens_in": 12997, "tokens_out": 11459, "tokens_reasoning": 11284, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-03-f", "pair": "hard-03", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.016067905, "wall_s": 228.8, "tokens_in": 50699, "tokens_out": 23891, "tokens_reasoning": 22954, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "hard-03-v", "pair": "hard-03", "family": "find-sol", "variant": "vulnerable", "status": "ok", "failure": null, "correct": false, "findings_read": 0, "hits": [], "planted": 1, "primary_all": false, "bite": null, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.004725915, "wall_s": 76.8, "tokens_in": 28203, "tokens_out": 5410, "tokens_reasoning": 5192, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-03-v", "pair": "hard-03", "family": "find-sol", "variant": "vulnerable", "status": "ok", "failure": null, "correct": false, "findings_read": 0, "hits": [], "planted": 1, "primary_all": false, "bite": null, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.01710649, "wall_s": 262.5, "tokens_in": 47600, "tokens_out": 27776, "tokens_reasoning": 26626, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "hard-04-f", "pair": "hard-04", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.00807956, "wall_s": 109.3, "tokens_in": 47221, "tokens_out": 10450, "tokens_reasoning": 10074, "turns": 6, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-04-f", "pair": "hard-04", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.02450982, "wall_s": 414.8, "tokens_in": 44173, "tokens_out": 43746, "tokens_reasoning": 41550, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": 0.00636868, "wall_s": 102.9, "tokens_in": 28767, "tokens_out": 8943, "tokens_reasoning": 8588, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": 0.02313571, "wall_s": 325, "tokens_in": 47301, "tokens_out": 39254, "tokens_reasoning": 37228, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": 0.01078069, "wall_s": 140.8, "tokens_in": 36256, "tokens_out": 16333, "tokens_reasoning": 15783, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "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": 0.02606952, "wall_s": 355.9, "tokens_in": 24248, "tokens_out": 47926, "tokens_reasoning": 46329, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "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": "unrated", "max_severity_exact": false, "usd": 0.033215255, "wall_s": 473.4, "tokens_in": 87262, "tokens_out": 52962, "tokens_reasoning": 51619, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "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": 0.043112345, "wall_s": 518.6, "tokens_in": 102577, "tokens_out": 66090, "tokens_reasoning": 63703, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "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": 0.013777575, "wall_s": 170.5, "tokens_in": 47398, "tokens_out": 19760, "tokens_reasoning": 19315, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "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": false, "findings_read": 1, "hits": [], "planted": 1, "primary_all": null, "bite": null, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": "overclaimed", "quote_hit": null, "max_severity": "medium", "max_severity_exact": true, "usd": 0.02459747, "wall_s": 333.9, "tokens_in": 25173, "tokens_out": 44738, "tokens_reasoning": 43064, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "report", "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": 0.0141048, "wall_s": 241.3, "tokens_in": 32969, "tokens_out": 24070, "tokens_reasoning": 22906, "turns": 4, "effort": "max", "providers": [], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "hard-08-o", "pair": "hard-08", "family": "challenge", "variant": "overclaimed", "status": "failed", "failure": "truncated", "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": 0.03510536, "wall_s": 534.4, "tokens_in": 25263, "tokens_out": 65723, "tokens_reasoning": 65585, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "length", "arm": "report", "rep": 1, "infra_retries": 0 }, { "case": "sol-02-f", "pair": "sol-02", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.009450585, "wall_s": 144.8, "tokens_in": 27685, "tokens_out": 14366, "tokens_reasoning": 13994, "turns": 5, "effort": "max", "providers": [], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "sol-02-f", "pair": "sol-02", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.027263245, "wall_s": 400.1, "tokens_in": 39577, "tokens_out": 48334, "tokens_reasoning": 46477, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.00992372, "wall_s": 151.4, "tokens_in": 34647, "tokens_out": 16193, "tokens_reasoning": 15837, "turns": 6, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.02181775, "wall_s": 301.9, "tokens_in": 40799, "tokens_out": 38140, "tokens_reasoning": 36017, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "sol-03-f", "pair": "sol-03", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 2, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 1, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.02530694, "wall_s": 385.6, "tokens_in": 51243, "tokens_out": 40690, "tokens_reasoning": 40169, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "sol-03-f", "pair": "sol-03", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 1, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.03302077, "wall_s": 542.8, "tokens_in": 50603, "tokens_out": 59050, "tokens_reasoning": 57182, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.014612755, "wall_s": 274.4, "tokens_in": 31575, "tokens_out": 24200, "tokens_reasoning": 23767, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": 0.030081045, "wall_s": 369.8, "tokens_in": 79672, "tokens_out": 45924, "tokens_reasoning": 43806, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "sol-05-f", "pair": "sol-05", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "low", "max_severity_exact": null, "usd": 0.00872582, "wall_s": 163.6, "tokens_in": 17783, "tokens_out": 15567, "tokens_reasoning": 15280, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "sol-05-f", "pair": "sol-05", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.02537419, "wall_s": 418.3, "tokens_in": 26541, "tokens_out": 48081, "tokens_reasoning": 46655, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": 2, "hits": [ { "id": "P1", "primary": true, "finding": 0, "by_line": true, "by_function": false, "severity": "high", "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": "high", "max_severity_exact": null, "usd": 0.011234895, "wall_s": 202.4, "tokens_in": 17731, "tokens_out": 20184, "tokens_reasoning": 19752, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": 2, "hits": [ { "id": "P1", "primary": true, "finding": 0, "by_line": true, "by_function": true, "severity": "high", "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": "high", "max_severity_exact": null, "usd": 0.02546214, "wall_s": 401.3, "tokens_in": 21654, "tokens_out": 48385, "tokens_reasoning": 46031, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "sol-09-f", "pair": "sol-09", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.005740605, "wall_s": 79.1, "tokens_in": 26835, "tokens_out": 8239, "tokens_reasoning": 7891, "turns": 6, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "sol-09-f", "pair": "sol-09", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "critical", "max_severity_exact": null, "usd": 0.024814555, "wall_s": 328.2, "tokens_in": 42960, "tokens_out": 43540, "tokens_reasoning": 41897, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "sol-09-v", "pair": "sol-09", "family": "find-sol", "variant": "vulnerable", "status": "ok", "failure": null, "correct": true, "findings_read": 2, "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": 1, "verdict": null, "quote_hit": null, "max_severity": "critical", "max_severity_exact": null, "usd": 0.00701198, "wall_s": 99.6, "tokens_in": 22151, "tokens_out": 10775, "tokens_reasoning": 10394, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "sol-09-v", "pair": "sol-09", "family": "find-sol", "variant": "vulnerable", "status": "ok", "failure": null, "correct": true, "findings_read": 2, "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": 0.028398775, "wall_s": 412.4, "tokens_in": 20367, "tokens_out": 53982, "tokens_reasoning": 51490, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 1, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "low", "max_severity_exact": null, "usd": 0.013026785, "wall_s": 235.3, "tokens_in": 32206, "tokens_out": 19974, "tokens_reasoning": 19597, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 1, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.024381115, "wall_s": 357.2, "tokens_in": 40084, "tokens_out": 43589, "tokens_reasoning": 41815, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.00866938, "wall_s": 148, "tokens_in": 8567, "tokens_out": 15864, "tokens_reasoning": 15612, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.029426415, "wall_s": 400.5, "tokens_in": 56872, "tokens_out": 49409, "tokens_reasoning": 47217, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "sol-11-f", "pair": "sol-11", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.01316648, "wall_s": 227.4, "tokens_in": 21708, "tokens_out": 23442, "tokens_reasoning": 23151, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "sol-11-f", "pair": "sol-11", "family": "find-sol", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 0, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "unrated", "max_severity_exact": null, "usd": 0.02189106, "wall_s": 327.7, "tokens_in": 16854, "tokens_out": 41453, "tokens_reasoning": 40429, "turns": 3, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.013782415, "wall_s": 186.7, "tokens_in": 36151, "tokens_out": 21996, "tokens_reasoning": 21622, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": 0.02949528, "wall_s": 354.9, "tokens_in": 49420, "tokens_out": 51536, "tokens_reasoning": 49439, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "ts-02-f", "pair": "ts-02", "family": "find-ts", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.020442095, "wall_s": 258.9, "tokens_in": 45407, "tokens_out": 33294, "tokens_reasoning": 32041, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "general", "rep": 1, "infra_retries": 0 }, { "case": "ts-02-f", "pair": "ts-02", "family": "find-ts", "variant": "fixed", "status": "ok", "failure": null, "correct": true, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.010971225, "wall_s": 156.2, "tokens_in": 33330, "tokens_out": 17039, "tokens_reasoning": 16735, "turns": 6, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": 2, "hits": [ { "id": "P1", "primary": true, "finding": 0, "by_line": true, "by_function": true, "severity": "high", "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": "high", "max_severity_exact": null, "usd": 0.02338782, "wall_s": 302.4, "tokens_in": 49073, "tokens_out": 38277, "tokens_reasoning": 36375, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "general", "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": 2, "hits": [ { "id": "P1", "primary": true, "finding": 0, "by_line": true, "by_function": true, "severity": "high", "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": "high", "max_severity_exact": null, "usd": 0.005375205, "wall_s": 68.3, "tokens_in": 27740, "tokens_out": 6762, "tokens_reasoning": 6382, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 }, { "case": "ts-03-f", "pair": "ts-03", "family": "find-ts", "variant": "fixed", "status": "ok", "failure": null, "correct": false, "findings_read": 1, "hits": [], "planted": 0, "primary_all": null, "bite": true, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.0247317, "wall_s": 353.4, "tokens_in": 34718, "tokens_out": 44890, "tokens_reasoning": 42998, "turns": 4, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "general", "rep": 1, "infra_retries": 0 }, { "case": "ts-03-f", "pair": "ts-03", "family": "find-ts", "variant": "fixed", "status": "ok", "failure": null, "correct": false, "findings_read": 2, "hits": [], "planted": 0, "primary_all": null, "bite": true, "decoys": 2, "decoy_bites": 0, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.01495358, "wall_s": 236.9, "tokens_in": 30140, "tokens_out": 26094, "tokens_reasoning": 25661, "turns": 6, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "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": "medium", "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": "medium", "max_severity_exact": null, "usd": 0.031312235, "wall_s": 402.1, "tokens_in": 69401, "tokens_out": 50165, "tokens_reasoning": 48183, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "general", "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": "high", "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": "high", "max_severity_exact": null, "usd": 0.006600045, "wall_s": 91.7, "tokens_in": 35587, "tokens_out": 9368, "tokens_reasoning": 9001, "turns": 5, "effort": "max", "providers": [ "OpenAI" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 } ] }