d3-mini-agentcode / results.json
stonefish-labs's picture
d3-mini-agentcode v0.1-preview (pre-release)
21a9138 verified
Raw History Blame Contribute Delete
27.8 kB
{
"thresholds": {
"seed1": [
0.9,
0.9
],
"seed2": [
0.9,
0.9
],
"untuned_same_context": [
0.5,
0.7
],
"untuned_task_change": [
0.5,
0.9
]
},
"multilingual": {
"B1s1": {
"c_cpp": {
"n": {
"positive": 38,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
35,
38
],
"fa_real": [
1,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.977,
"auroc_pos_vs_real_ci": [
0.9434,
1.0
]
},
"csharp": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
34,
37
],
"fa_real": [
0,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9975,
"auroc_pos_vs_real_ci": [
0.9923,
1.0
]
},
"go": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
36,
38
],
"fa_real": [
1,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9991,
"auroc_pos_vs_real_ci": [
0.9966,
1.0
]
},
"java": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
36,
37
],
"fa_real": [
1,
149
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.9982,
"auroc_pos_vs_real_ci": [
0.994,
1.0
]
},
"js_ts": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
33,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9991,
"auroc_pos_vs_real_ci": [
0.9969,
1.0
]
},
"php": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
34,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9837,
"auroc_pos_vs_real_ci": [
0.9519,
1.0
]
},
"rust": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
33,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.962,
"auroc_pos_vs_real_ci": [
0.8927,
1.0
]
},
"swift_objc": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
37,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9995,
"auroc_pos_vs_real_ci": [
0.9975,
1.0
]
},
"cm_bash_sh": {
"n": {
"positive": 33,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
24,
33
],
"fa_real": [
1,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9345,
"auroc_pos_vs_real_ci": [
0.924,
0.942
]
},
"cm_cmd_batch": {
"n": {
"positive": 33,
"real": 130,
"hard_control": 30,
"easy_control": 10
},
"recall": [
28,
33
],
"fa_real": [
0,
130
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9963,
"auroc_pos_vs_real_ci": [
0.9862,
1.0
]
},
"cm_zsh_macos": {
"n": {
"positive": 36,
"real": 144,
"hard_control": 30,
"easy_control": 10
},
"recall": [
29,
36
],
"fa_real": [
3,
144
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9745,
"auroc_pos_vs_real_ci": [
0.9387,
1.0
]
},
"cm_powershell": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
35,
37
],
"fa_real": [
3,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9917,
"auroc_pos_vs_real_ci": [
0.984,
0.9974
]
},
"cm_python": {
"n": {
"positive": 35,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
30,
35
],
"fa_real": [
1,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9859,
"auroc_pos_vs_real_ci": [
0.9848,
1.0
]
},
"cm_js_ts": {
"n": {
"positive": 36,
"real": 129,
"hard_control": 30,
"easy_control": 10
},
"recall": [
34,
36
],
"fa_real": [
0,
129
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9978,
"auroc_pos_vs_real_ci": [
0.9678,
1.0
]
},
"ALL_diff": {
"n": {
"positive": 300,
"real": 1197,
"hard_control": 240,
"easy_control": 80
},
"recall": [
278,
300
],
"fa_real": [
3,
1197
],
"fa_controls": [
1,
320
],
"auroc_pos_vs_real": 0.989,
"auroc_pos_vs_real_ci": [
0.9782,
0.9974
]
},
"ALL_codemode": {
"n": {
"positive": 210,
"real": 953,
"hard_control": 180,
"easy_control": 60
},
"recall": [
180,
210
],
"fa_real": [
8,
953
],
"fa_controls": [
0,
240
],
"auroc_pos_vs_real": 0.9789,
"auroc_pos_vs_real_ci": [
0.9703,
0.9957
]
},
"ALL": {
"n": {
"positive": 510,
"real": 2150,
"hard_control": 420,
"easy_control": 140
},
"recall": [
458,
510
],
"fa_real": [
11,
2150
],
"fa_controls": [
1,
560
],
"auroc_pos_vs_real": 0.9849,
"auroc_pos_vs_real_ci": [
0.9775,
0.9929
]
}
},
"B1s2": {
"c_cpp": {
"n": {
"positive": 38,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
29,
38
],
"fa_real": [
1,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9693,
"auroc_pos_vs_real_ci": [
0.9134,
0.9998
]
},
"csharp": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
33,
37
],
"fa_real": [
0,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9808,
"auroc_pos_vs_real_ci": [
0.9498,
0.9991
]
},
"go": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
35,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9972,
"auroc_pos_vs_real_ci": [
0.9883,
1.0
]
},
"java": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
34,
37
],
"fa_real": [
2,
149
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.9978,
"auroc_pos_vs_real_ci": [
0.993,
1.0
]
},
"js_ts": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
35,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9996,
"auroc_pos_vs_real_ci": [
0.9982,
1.0
]
},
"php": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
33,
38
],
"fa_real": [
1,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9979,
"auroc_pos_vs_real_ci": [
0.9938,
1.0
]
},
"rust": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
32,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9746,
"auroc_pos_vs_real_ci": [
0.9268,
1.0
]
},
"swift_objc": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
37,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 1.0,
"auroc_pos_vs_real_ci": [
1.0,
1.0
]
},
"cm_bash_sh": {
"n": {
"positive": 33,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
22,
33
],
"fa_real": [
0,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9547,
"auroc_pos_vs_real_ci": [
0.9539,
0.956
]
},
"cm_cmd_batch": {
"n": {
"positive": 33,
"real": 130,
"hard_control": 30,
"easy_control": 10
},
"recall": [
29,
33
],
"fa_real": [
0,
130
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9939,
"auroc_pos_vs_real_ci": [
0.9775,
1.0
]
},
"cm_zsh_macos": {
"n": {
"positive": 36,
"real": 144,
"hard_control": 30,
"easy_control": 10
},
"recall": [
31,
36
],
"fa_real": [
3,
144
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9807,
"auroc_pos_vs_real_ci": [
0.9538,
1.0
]
},
"cm_powershell": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
33,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.9973,
"auroc_pos_vs_real_ci": [
0.9904,
1.0
]
},
"cm_python": {
"n": {
"positive": 35,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
29,
35
],
"fa_real": [
0,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9659,
"auroc_pos_vs_real_ci": [
0.9573,
1.0
]
},
"cm_js_ts": {
"n": {
"positive": 36,
"real": 129,
"hard_control": 30,
"easy_control": 10
},
"recall": [
34,
36
],
"fa_real": [
0,
129
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9882,
"auroc_pos_vs_real_ci": [
0.9211,
1.0
]
},
"ALL_diff": {
"n": {
"positive": 300,
"real": 1197,
"hard_control": 240,
"easy_control": 80
},
"recall": [
268,
300
],
"fa_real": [
4,
1197
],
"fa_controls": [
1,
320
],
"auroc_pos_vs_real": 0.99,
"auroc_pos_vs_real_ci": [
0.9794,
0.9977
]
},
"ALL_codemode": {
"n": {
"positive": 210,
"real": 953,
"hard_control": 180,
"easy_control": 60
},
"recall": [
178,
210
],
"fa_real": [
3,
953
],
"fa_controls": [
1,
240
],
"auroc_pos_vs_real": 0.9787,
"auroc_pos_vs_real_ci": [
0.9691,
0.9955
]
},
"ALL": {
"n": {
"positive": 510,
"real": 2150,
"hard_control": 420,
"easy_control": 140
},
"recall": [
446,
510
],
"fa_real": [
7,
2150
],
"fa_controls": [
2,
560
],
"auroc_pos_vs_real": 0.9853,
"auroc_pos_vs_real_ci": [
0.9778,
0.9938
]
}
},
"A1": {
"c_cpp": {
"n": {
"positive": 38,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
20,
38
],
"fa_real": [
2,
149
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.8347,
"auroc_pos_vs_real_ci": [
0.7673,
0.902
]
},
"csharp": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
24,
37
],
"fa_real": [
0,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.8395,
"auroc_pos_vs_real_ci": [
0.7621,
0.9139
]
},
"go": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
20,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.8644,
"auroc_pos_vs_real_ci": [
0.8064,
0.9181
]
},
"java": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
21,
37
],
"fa_real": [
2,
149
],
"fa_controls": [
2,
40
],
"auroc_pos_vs_real": 0.8104,
"auroc_pos_vs_real_ci": [
0.7316,
0.8962
]
},
"js_ts": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
24,
37
],
"fa_real": [
1,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.8393,
"auroc_pos_vs_real_ci": [
0.7548,
0.909
]
},
"php": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
21,
38
],
"fa_real": [
1,
150
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.8077,
"auroc_pos_vs_real_ci": [
0.7153,
0.885
]
},
"rust": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
15,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.7335,
"auroc_pos_vs_real_ci": [
0.6261,
0.8453
]
},
"swift_objc": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
27,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9247,
"auroc_pos_vs_real_ci": [
0.875,
0.9719
]
},
"cm_bash_sh": {
"n": {
"positive": 33,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
24,
33
],
"fa_real": [
14,
200
],
"fa_controls": [
3,
40
],
"auroc_pos_vs_real": 0.8518,
"auroc_pos_vs_real_ci": [
0.8275,
0.906
]
},
"cm_cmd_batch": {
"n": {
"positive": 33,
"real": 130,
"hard_control": 30,
"easy_control": 10
},
"recall": [
28,
33
],
"fa_real": [
1,
130
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.9583,
"auroc_pos_vs_real_ci": [
0.9241,
0.9852
]
},
"cm_zsh_macos": {
"n": {
"positive": 36,
"real": 144,
"hard_control": 30,
"easy_control": 10
},
"recall": [
26,
36
],
"fa_real": [
5,
144
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9037,
"auroc_pos_vs_real_ci": [
0.8728,
0.9705
]
},
"cm_powershell": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
29,
37
],
"fa_real": [
8,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9575,
"auroc_pos_vs_real_ci": [
0.9455,
0.9669
]
},
"cm_python": {
"n": {
"positive": 35,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
31,
35
],
"fa_real": [
12,
200
],
"fa_controls": [
3,
40
],
"auroc_pos_vs_real": 0.9184,
"auroc_pos_vs_real_ci": [
0.8827,
0.998
]
},
"cm_js_ts": {
"n": {
"positive": 36,
"real": 129,
"hard_control": 30,
"easy_control": 10
},
"recall": [
33,
36
],
"fa_real": [
2,
129
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.9246,
"auroc_pos_vs_real_ci": [
0.8554,
0.9765
]
},
"ALL_diff": {
"n": {
"positive": 300,
"real": 1197,
"hard_control": 240,
"easy_control": 80
},
"recall": [
172,
300
],
"fa_real": [
6,
1197
],
"fa_controls": [
4,
320
],
"auroc_pos_vs_real": 0.8309,
"auroc_pos_vs_real_ci": [
0.8021,
0.859
]
},
"ALL_codemode": {
"n": {
"positive": 210,
"real": 953,
"hard_control": 180,
"easy_control": 60
},
"recall": [
171,
210
],
"fa_real": [
42,
953
],
"fa_controls": [
8,
240
],
"auroc_pos_vs_real": 0.9222,
"auroc_pos_vs_real_ci": [
0.8954,
0.961
]
},
"ALL": {
"n": {
"positive": 510,
"real": 2150,
"hard_control": 420,
"easy_control": 140
},
"recall": [
343,
510
],
"fa_real": [
48,
2150
],
"fa_controls": [
12,
560
],
"auroc_pos_vs_real": 0.8258,
"auroc_pos_vs_real_ci": [
0.7989,
0.8647
]
}
},
"A0": {
"c_cpp": {
"n": {
"positive": 38,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
5,
38
],
"fa_real": [
2,
149
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.8108,
"auroc_pos_vs_real_ci": [
0.7353,
0.8849
]
},
"csharp": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
5,
37
],
"fa_real": [
0,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.8132,
"auroc_pos_vs_real_ci": [
0.7154,
0.9009
]
},
"go": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
5,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.8486,
"auroc_pos_vs_real_ci": [
0.7859,
0.9017
]
},
"java": {
"n": {
"positive": 37,
"real": 149,
"hard_control": 30,
"easy_control": 10
},
"recall": [
3,
37
],
"fa_real": [
0,
149
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.7818,
"auroc_pos_vs_real_ci": [
0.7056,
0.8676
]
},
"js_ts": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
1,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.7735,
"auroc_pos_vs_real_ci": [
0.683,
0.8583
]
},
"php": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
5,
38
],
"fa_real": [
1,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.7544,
"auroc_pos_vs_real_ci": [
0.6627,
0.8424
]
},
"rust": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
2,
37
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.6823,
"auroc_pos_vs_real_ci": [
0.5802,
0.7969
]
},
"swift_objc": {
"n": {
"positive": 38,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
10,
38
],
"fa_real": [
0,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9077,
"auroc_pos_vs_real_ci": [
0.8558,
0.9609
]
},
"cm_bash_sh": {
"n": {
"positive": 33,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
17,
33
],
"fa_real": [
6,
200
],
"fa_controls": [
1,
40
],
"auroc_pos_vs_real": 0.8736,
"auroc_pos_vs_real_ci": [
0.8672,
0.894
]
},
"cm_cmd_batch": {
"n": {
"positive": 33,
"real": 130,
"hard_control": 30,
"easy_control": 10
},
"recall": [
23,
33
],
"fa_real": [
0,
130
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9688,
"auroc_pos_vs_real_ci": [
0.9396,
0.9917
]
},
"cm_zsh_macos": {
"n": {
"positive": 36,
"real": 144,
"hard_control": 30,
"easy_control": 10
},
"recall": [
23,
36
],
"fa_real": [
1,
144
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9323,
"auroc_pos_vs_real_ci": [
0.9026,
0.9886
]
},
"cm_powershell": {
"n": {
"positive": 37,
"real": 150,
"hard_control": 30,
"easy_control": 10
},
"recall": [
22,
37
],
"fa_real": [
3,
150
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9548,
"auroc_pos_vs_real_ci": [
0.9471,
0.9669
]
},
"cm_python": {
"n": {
"positive": 35,
"real": 200,
"hard_control": 30,
"easy_control": 10
},
"recall": [
25,
35
],
"fa_real": [
5,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9164,
"auroc_pos_vs_real_ci": [
0.8851,
0.998
]
},
"cm_js_ts": {
"n": {
"positive": 36,
"real": 129,
"hard_control": 30,
"easy_control": 10
},
"recall": [
26,
36
],
"fa_real": [
4,
129
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9076,
"auroc_pos_vs_real_ci": [
0.7878,
0.962
]
},
"ALL_diff": {
"n": {
"positive": 300,
"real": 1197,
"hard_control": 240,
"easy_control": 80
},
"recall": [
36,
300
],
"fa_real": [
3,
1197
],
"fa_controls": [
1,
320
],
"auroc_pos_vs_real": 0.796,
"auroc_pos_vs_real_ci": [
0.7671,
0.8273
]
},
"ALL_codemode": {
"n": {
"positive": 210,
"real": 953,
"hard_control": 180,
"easy_control": 60
},
"recall": [
136,
210
],
"fa_real": [
19,
953
],
"fa_controls": [
1,
240
],
"auroc_pos_vs_real": 0.924,
"auroc_pos_vs_real_ci": [
0.9028,
0.9563
]
},
"ALL": {
"n": {
"positive": 510,
"real": 2150,
"hard_control": 420,
"easy_control": 140
},
"recall": [
172,
510
],
"fa_real": [
22,
2150
],
"fa_controls": [
2,
560
],
"auroc_pos_vs_real": 0.8009,
"auroc_pos_vs_real_ci": [
0.7743,
0.8374
]
}
}
},
"email_in_headers": {
"B1s1": {
"theta": [
0.9,
0.9
],
"benign|all": "0/60",
"benign|lang:bash": "0/18",
"benign|lang:js": "0/18",
"benign|lang:python": "0/24",
"benign|writer:claude": "0/30",
"benign|writer:openai": "0/30",
"leak|all": "1/40",
"leak|lang:bash": "0/12",
"leak|lang:js": "1/12",
"leak|lang:python": "0/16",
"leak|writer:claude": "1/20",
"leak|writer:openai": "0/20",
"variant:C1": "0/15",
"variant:C2": "0/15",
"variant:C3": "0/15",
"variant:C4": "0/15",
"variant:P1": "0/14",
"variant:P2": "0/13",
"variant:P3": "1/13",
"bases_fully_right": "1/40"
},
"B1s2": {
"theta": [
0.9,
0.9
],
"benign|all": "0/60",
"benign|lang:bash": "0/18",
"benign|lang:js": "0/18",
"benign|lang:python": "0/24",
"benign|writer:claude": "0/30",
"benign|writer:openai": "0/30",
"leak|all": "0/40",
"leak|lang:bash": "0/12",
"leak|lang:js": "0/12",
"leak|lang:python": "0/16",
"leak|writer:claude": "0/20",
"leak|writer:openai": "0/20",
"variant:C1": "0/15",
"variant:C2": "0/15",
"variant:C3": "0/15",
"variant:C4": "0/15",
"variant:P1": "0/14",
"variant:P2": "0/13",
"variant:P3": "0/13",
"bases_fully_right": "0/40"
},
"A1": {
"theta": [
0.5,
0.7
],
"benign|all": "4/60",
"benign|lang:bash": "0/18",
"benign|lang:js": "4/18",
"benign|lang:python": "0/24",
"benign|writer:claude": "2/30",
"benign|writer:openai": "2/30",
"leak|all": "3/40",
"leak|lang:bash": "0/12",
"leak|lang:js": "2/12",
"leak|lang:python": "1/16",
"leak|writer:claude": "1/20",
"leak|writer:openai": "2/20",
"variant:C1": "2/15",
"variant:C2": "0/15",
"variant:C3": "0/15",
"variant:C4": "2/15",
"variant:P1": "1/14",
"variant:P2": "2/13",
"variant:P3": "0/13",
"bases_fully_right": "1/40"
}
},
"placeholder_summary": {
"seed": 1,
"theta": [
0.9,
0.9
],
"multilingual": {
"ALL": {
"extractor": {
"recall": [
458,
510
],
"fa_real": [
11,
2150
],
"fa_controls": [
1,
560
],
"auroc_pos_vs_real": 0.9849
},
"placeholder": {
"recall": [
448,
510
],
"fa_real": [
5,
2150
],
"fa_controls": [
1,
560
],
"auroc_pos_vs_real": 0.9837
}
},
"ALL_diff": {
"extractor": {
"recall": [
278,
300
],
"fa_real": [
3,
1197
],
"fa_controls": [
1,
320
],
"auroc_pos_vs_real": 0.989
},
"placeholder": {
"recall": [
277,
300
],
"fa_real": [
2,
1197
],
"fa_controls": [
1,
320
],
"auroc_pos_vs_real": 0.9888
}
},
"ALL_codemode": {
"extractor": {
"recall": [
180,
210
],
"fa_real": [
8,
953
],
"fa_controls": [
0,
240
],
"auroc_pos_vs_real": 0.9789
},
"placeholder": {
"recall": [
171,
210
],
"fa_real": [
3,
953
],
"fa_controls": [
0,
240
],
"auroc_pos_vs_real": 0.9755
}
},
"cm_bash_sh": {
"extractor": {
"recall": [
24,
33
],
"fa_real": [
1,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.9345
},
"placeholder": {
"recall": [
20,
33
],
"fa_real": [
1,
200
],
"fa_controls": [
0,
40
],
"auroc_pos_vs_real": 0.927
}
}
},
"independent_agent_diffs": {
"extractor": {
"found": [
21,
22
],
"flagged": [
0,
172
]
},
"placeholder": {
"found": [
20,
22
],
"flagged": [
0,
172
]
}
}
}
}