{ "schema": "paydirt.model/1", "release": "2026-10", "run_id": "2026-10", "slug": "deepseek/deepseek-v4.1-flash", "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": 0.020736278, "wall_s": 166.5, "tokens_in": 49393, "tokens_out": 28479, "tokens_reasoning": 27165, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.01159636, "wall_s": 188.2, "tokens_in": 28598, "tokens_out": 18374, "tokens_reasoning": 17012, "turns": 4, "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": "info", "max_severity_exact": true, "usd": 0.019153891, "wall_s": 146.9, "tokens_in": 50956, "tokens_out": 25369, "tokens_reasoning": 23876, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.02153916, "wall_s": 197.6, "tokens_in": 28800, "tokens_out": 34734, "tokens_reasoning": 32155, "turns": 4, "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": 0.00831142, "wall_s": 225.7, "tokens_in": 12551, "tokens_out": 13434, "tokens_reasoning": 12699, "turns": 3, "effort": "max", "providers": [ "Relace" ], "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": 0.023083103, "wall_s": 202.2, "tokens_in": 27627, "tokens_out": 35981, "tokens_reasoning": 34047, "turns": 4, "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": 1, "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": 0.073782492, "wall_s": 313.8, "tokens_in": 85729, "tokens_out": 58681, "tokens_reasoning": 57508, "turns": 5, "effort": "max", "providers": [ "Together" ], "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": "overclaimed", "quote_hit": true, "max_severity": "critical", "max_severity_exact": false, "usd": 0.077225124, "wall_s": 215.9, "tokens_in": 158617, "tokens_out": 59158, "tokens_reasoning": 56719, "turns": 9, "effort": "max", "providers": [], "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": 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": 0.03531516, "wall_s": 142.1, "tokens_in": 51595, "tokens_out": 26138, "tokens_reasoning": 24600, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.026956536, "wall_s": 71.9, "tokens_in": 73498, "tokens_out": 19048, "tokens_reasoning": 16794, "turns": 6, "effort": "max", "providers": [], "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": 3, "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": 0.042062105, "wall_s": 387.6, "tokens_in": 21040, "tokens_out": 71507, "tokens_reasoning": 70225, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 2, "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": 0.074513628, "wall_s": 197.9, "tokens_in": 36213, "tokens_out": 59188, "tokens_reasoning": 55790, "turns": 5, "effort": "max", "providers": [], "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": 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": 0.064257948, "wall_s": 277.8, "tokens_in": 13349, "tokens_out": 51340, "tokens_reasoning": 50141, "turns": 3, "effort": "max", "providers": [ "Together" ], "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": 0.015473848, "wall_s": 153.2, "tokens_in": 34598, "tokens_out": 24460, "tokens_reasoning": 22300, "turns": 5, "effort": "max", "providers": [], "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": 0.032259982, "wall_s": 312.4, "tokens_in": 19447, "tokens_out": 54612, "tokens_reasoning": 53205, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.021049963, "wall_s": 184.8, "tokens_in": 29892, "tokens_out": 33502, "tokens_reasoning": 29985, "turns": 4, "effort": "max", "providers": [], "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": 5, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.047466973, "wall_s": 453.3, "tokens_in": 24514, "tokens_out": 81172, "tokens_reasoning": 79692, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.048276578, "wall_s": 710.4, "tokens_in": 126048, "tokens_out": 91705, "tokens_reasoning": 88753, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 3, "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": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.040675003, "wall_s": 307.1, "tokens_in": 198640, "tokens_out": 53102, "tokens_reasoning": 51364, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "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": 3, "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": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.027303127, "wall_s": 260.5, "tokens_in": 39464, "tokens_out": 44995, "tokens_reasoning": 41569, "turns": 5, "effort": "max", "providers": [], "stop": "stop", "arm": "solidity", "rep": 1, "infra_retries": 0 }, { "case": "hard-02-f", "pair": "hard-02", "family": "find-sol", "variant": "fixed", "status": "failed", "failure": "truncated", "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": 0.213710316, "wall_s": 813.2, "tokens_in": 69733, "tokens_out": 175147, "tokens_reasoning": 174716, "turns": 4, "effort": "max", "providers": [ "Together" ], "stop": "length", "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": 8, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 1, "unmatched": 2, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.085178954, "wall_s": 697.8, "tokens_in": 247289, "tokens_out": 120527, "tokens_reasoning": 116618, "turns": 6, "effort": "max", "providers": [ "InferenceNet" ], "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": 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": 0.04564906, "wall_s": 329.9, "tokens_in": 165041, "tokens_out": 60961, "tokens_reasoning": 59232, "turns": 7, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.03515932, "wall_s": 956.9, "tokens_in": 76498, "tokens_out": 56635, "tokens_reasoning": 54004, "turns": 4, "effort": "max", "providers": [ "InferenceNet" ], "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": 1, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "info", "max_severity_exact": null, "usd": 0.119227513, "wall_s": 924.4, "tokens_in": 448658, "tokens_out": 165731, "tokens_reasoning": 163965, "turns": 7, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.086061268, "wall_s": 666.7, "tokens_in": 294570, "tokens_out": 119758, "tokens_reasoning": 117595, "turns": 7, "effort": "max", "providers": [ "AtlasCloud" ], "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": 1, "hits": [], "planted": 1, "primary_all": false, "bite": null, "decoys": 3, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "low", "max_severity_exact": null, "usd": 0.120875466, "wall_s": 764.6, "tokens_in": 244711, "tokens_out": 96258, "tokens_reasoning": 94495, "turns": 9, "effort": "max", "providers": [ "AtlasCloud", "Together" ], "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": 0.073033996, "wall_s": 718.7, "tokens_in": 157314, "tokens_out": 123342, "tokens_reasoning": 121233, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "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": 5, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 0, "unmatched": 2, "verdict": null, "quote_hit": null, "max_severity": "critical", "max_severity_exact": null, "usd": 0.130701375, "wall_s": 455.6, "tokens_in": 128311, "tokens_out": 90015, "tokens_reasoning": 88384, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "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": 2, "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": 0.04689237, "wall_s": 397.4, "tokens_in": 192874, "tokens_out": 65596, "tokens_reasoning": 62110, "turns": 7, "effort": "max", "providers": [ "AtlasCloud" ], "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": 4, "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": 0.127128843, "wall_s": 475.8, "tokens_in": 234901, "tokens_out": 101228, "tokens_reasoning": 99765, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.043684366, "wall_s": 432.7, "tokens_in": 23111, "tokens_out": 74384, "tokens_reasoning": 71968, "turns": 3, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.057278364, "wall_s": 218.7, "tokens_in": 26909, "tokens_out": 44078, "tokens_reasoning": 42754, "turns": 4, "effort": "max", "providers": [ "Together" ], "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": 0.04906452, "wall_s": 164.2, "tokens_in": 42142, "tokens_out": 37094, "tokens_reasoning": 34947, "turns": 5, "effort": "max", "providers": [ "Together" ], "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": "high", "max_severity_exact": false, "usd": 0.031101357, "wall_s": 220.3, "tokens_in": 79441, "tokens_out": 41188, "tokens_reasoning": 39471, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": "overclaimed", "quote_hit": true, "max_severity": "high", "max_severity_exact": false, "usd": 0.059589768, "wall_s": 278.9, "tokens_in": 36694, "tokens_out": 46004, "tokens_reasoning": 43315, "turns": 4, "effort": "max", "providers": [ "Together" ], "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": 0.030868199, "wall_s": 731.4, "tokens_in": 41057, "tokens_out": 50745, "tokens_reasoning": 49596, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.078633816, "wall_s": 380.7, "tokens_in": 42998, "tokens_out": 61427, "tokens_reasoning": 58749, "turns": 5, "effort": "max", "providers": [ "Together" ], "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": 0.055946784, "wall_s": 191.1, "tokens_in": 38983, "tokens_out": 42396, "tokens_reasoning": 40877, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.030860642, "wall_s": 269, "tokens_in": 59585, "tokens_out": 49354, "tokens_reasoning": 46069, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "stop": "stop", "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": 2, "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": 0.058528536, "wall_s": 1270.4, "tokens_in": 194760, "tokens_out": 86044, "tokens_reasoning": 84510, "turns": 10, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.106011504, "wall_s": 527.4, "tokens_in": 111576, "tokens_out": 85631, "tokens_reasoning": 83707, "turns": 4, "effort": "max", "providers": [ "Together" ], "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": "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": 0.018180371, "wall_s": 147.5, "tokens_in": 39546, "tokens_out": 25113, "tokens_reasoning": 23761, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": "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": 0.021028684, "wall_s": 172.4, "tokens_in": 53646, "tokens_out": 28539, "tokens_reasoning": 25871, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 6, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 1, "decoy_bites": 1, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.076493825, "wall_s": 612.8, "tokens_in": 213067, "tokens_out": 104335, "tokens_reasoning": 102737, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 2, "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": 0.06338984, "wall_s": 884.6, "tokens_in": 105983, "tokens_out": 99493, "tokens_reasoning": 96446, "turns": 4, "effort": "max", "providers": [ "InferenceNet" ], "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": "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": 0.059993028, "wall_s": 195.6, "tokens_in": 103315, "tokens_out": 45992, "tokens_reasoning": 44727, "turns": 7, "effort": "max", "providers": [ "Together" ], "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": 0.076653552, "wall_s": 362.8, "tokens_in": 88632, "tokens_out": 54703, "tokens_reasoning": 51987, "turns": 4, "effort": "max", "providers": [ "Together" ], "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": 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": 0.142804524, "wall_s": 432.2, "tokens_in": 220213, "tokens_out": 114691, "tokens_reasoning": 113227, "turns": 6, "effort": "max", "providers": [ "Together" ], "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": 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": 0.059876581, "wall_s": 458, "tokens_in": 281771, "tokens_out": 79555, "tokens_reasoning": 77162, "turns": 7, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.03089876, "wall_s": 834, "tokens_in": 128045, "tokens_out": 59291, "tokens_reasoning": 57724, "turns": 7, "effort": "max", "providers": [ "InferenceNet" ], "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": 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": 0.021444577, "wall_s": 213.9, "tokens_in": 19402, "tokens_out": 35879, "tokens_reasoning": 33053, "turns": 3, "effort": "max", "providers": [ "AtlasCloud" ], "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": 2, "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": 0.09383514, "wall_s": 301.2, "tokens_in": 125575, "tokens_out": 75183, "tokens_reasoning": 73645, "turns": 6, "effort": "max", "providers": [ "Together" ], "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": "medium", "max_severity_exact": null, "usd": 0.034273631, "wall_s": 332.6, "tokens_in": 23201, "tokens_out": 58799, "tokens_reasoning": 56512, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0, "verdict": null, "quote_hit": null, "max_severity": "critical", "max_severity_exact": null, "usd": 0.025603188, "wall_s": 84.5, "tokens_in": 37347, "tokens_out": 19557, "tokens_reasoning": 18058, "turns": 5, "effort": "max", "providers": [ "Together" ], "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": 1, "verdict": null, "quote_hit": null, "max_severity": "critical", "max_severity_exact": null, "usd": 0.017424667, "wall_s": 172.8, "tokens_in": 17956, "tokens_out": 29113, "tokens_reasoning": 26533, "turns": 3, "effort": "max", "providers": [], "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": 4, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 1, "decoy_bites": 1, "unmatched": 1, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.056391963, "wall_s": 453.6, "tokens_in": 76575, "tokens_out": 83146, "tokens_reasoning": 81644, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": "medium", "max_severity_exact": null, "usd": 0.059086022, "wall_s": 540.1, "tokens_in": 205629, "tokens_out": 88808, "tokens_reasoning": 85829, "turns": 7, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.021281215, "wall_s": 169.8, "tokens_in": 41977, "tokens_out": 29600, "tokens_reasoning": 28475, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 0.035594928, "wall_s": 170.8, "tokens_in": 18480, "tokens_out": 28053, "tokens_reasoning": 25506, "turns": 3, "effort": "max", "providers": [], "stop": "stop", "arm": "solidity", "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": 0.03632054, "wall_s": 1801.5, "tokens_in": 14257, "tokens_out": 60059, "tokens_reasoning": 54913, "turns": 5, "effort": "max", "providers": [ "Relace" ], "stop": "aborted", "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": 3, "hits": [], "planted": 0, "primary_all": null, "bite": false, "decoys": 2, "decoy_bites": 1, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "medium", "max_severity_exact": null, "usd": 0.089224548, "wall_s": 486.6, "tokens_in": 82327, "tokens_out": 71365, "tokens_reasoning": 68735, "turns": 7, "effort": "max", "providers": [ "Together" ], "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": 3, "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": 0.041158661, "wall_s": 858.8, "tokens_in": 16619, "tokens_out": 71932, "tokens_reasoning": 70337, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "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": "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": 0.052709021, "wall_s": 420, "tokens_in": 106904, "tokens_out": 70581, "tokens_reasoning": 68144, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.020055784, "wall_s": 171.4, "tokens_in": 40042, "tokens_out": 29639, "tokens_reasoning": 27943, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 5, "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": 0.087189961, "wall_s": 583.3, "tokens_in": 191897, "tokens_out": 94265, "tokens_reasoning": 92416, "turns": 8, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.04340196, "wall_s": 207.5, "tokens_in": 18462, "tokens_out": 33748, "tokens_reasoning": 31575, "turns": 3, "effort": "max", "providers": [ "Together" ], "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": 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": 0.012072646, "wall_s": 85.5, "tokens_in": 37000, "tokens_out": 15093, "tokens_reasoning": 14059, "turns": 4, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.033918678, "wall_s": 274, "tokens_in": 101038, "tokens_out": 46976, "tokens_reasoning": 45111, "turns": 5, "effort": "max", "providers": [ "AtlasCloud" ], "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": 7, "hits": [], "planted": 0, "primary_all": null, "bite": true, "decoys": 2, "decoy_bites": 0, "unmatched": 0, "verdict": null, "quote_hit": null, "max_severity": "high", "max_severity_exact": null, "usd": 0.119515197, "wall_s": 291.7, "tokens_in": 210195, "tokens_out": 93219, "tokens_reasoning": 91043, "turns": 6, "effort": "max", "providers": [ "AtlasCloud" ], "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": 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": 0.054168816, "wall_s": 251.3, "tokens_in": 75828, "tokens_out": 42240, "tokens_reasoning": 39228, "turns": 5, "effort": "max", "providers": [ "Together" ], "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": 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": 0.072961752, "wall_s": 190.8, "tokens_in": 44066, "tokens_out": 58064, "tokens_reasoning": 57117, "turns": 4, "effort": "max", "providers": [ "Together" ], "stop": "stop", "arm": "raw", "rep": 1, "infra_retries": 0 } ] }