Hugging Face's logo

XcodeAddy
/
sentinel-grpo-qwen15

Model card Files
xet
 Community
XcodeAddy commited on
Commit
ca9c749
·
verified ·
1 Parent(s): bec7b74

Upload folder using huggingface_hub

Browse files
Files changed (20) hide show
  1. .gitattributes +5 -0
  2. outputs/baseline_scores.json +531 -0
  3. outputs/charts/ablation.png +0 -0
  4. outputs/charts/baseline_delta_lines.png +3 -0
  5. outputs/charts/baseline_grouped_bars.png +0 -0
  6. outputs/charts/cluster_health_policy_lines.png +0 -0
  7. outputs/charts/cluster_health_timeline.png +0 -0
  8. outputs/charts/detection_vs_poisoning.png +0 -0
  9. outputs/charts/failure_fishbone_map.png +3 -0
  10. outputs/charts/grpo_reward_curve.png +0 -0
  11. outputs/charts/reward_component_stacked_area.png +3 -0
  12. outputs/charts/task_radar.png +3 -0
  13. outputs/charts/trust_evolution.png +0 -0
  14. outputs/charts/trust_gap_over_time.png +3 -0
  15. outputs/cluster_health_history.json +128 -0
  16. outputs/eval_post.json +0 -0
  17. outputs/eval_pre.json +0 -0
  18. outputs/evaluation_results.json +0 -0
  19. outputs/reward_report_task3_seed42.json +526 -0
  20. outputs/trained_policy_replay.jsonl +0 -0
.gitattributes CHANGED
@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ outputs/charts/baseline_delta_lines.png filter=lfs diff=lfs merge=lfs -text
37
+ outputs/charts/failure_fishbone_map.png filter=lfs diff=lfs merge=lfs -text
38
+ outputs/charts/reward_component_stacked_area.png filter=lfs diff=lfs merge=lfs -text
39
+ outputs/charts/task_radar.png filter=lfs diff=lfs merge=lfs -text
40
+ outputs/charts/trust_gap_over_time.png filter=lfs diff=lfs merge=lfs -text
outputs/baseline_scores.json ADDED
@@ -0,0 +1,531 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "heuristic-baseline",
3
+ "total_episodes": 30,
4
+ "avg_score": 0.7942,
5
+ "by_task": {
6
+ "task1": {
7
+ "episodes": 10,
8
+ "avg_score": 0.8706
9
+ },
10
+ "task2": {
11
+ "episodes": 10,
12
+ "avg_score": 0.7475
13
+ },
14
+ "task3": {
15
+ "episodes": 10,
16
+ "avg_score": 0.7646
17
+ }
18
+ },
19
+ "episodes": [
20
+ {
21
+ "scenario_id": "SCN-TASK1-001",
22
+ "task_type": "task1",
23
+ "steps": 13,
24
+ "score": 0.765,
25
+ "total_reward": 10.71,
26
+ "completion_rate": 0.8,
27
+ "adversarial_detections": 0,
28
+ "adversarial_poisonings": 0,
29
+ "final_trust": {
30
+ "S0": 0.473,
31
+ "S1": 0.743,
32
+ "S2": 0.5,
33
+ "S3": 0.5,
34
+ "S4": 0.5
35
+ }
36
+ },
37
+ {
38
+ "scenario_id": "SCN-TASK1-002",
39
+ "task_type": "task1",
40
+ "steps": 12,
41
+ "score": 0.7962,
42
+ "total_reward": 10.35,
43
+ "completion_rate": 0.8,
44
+ "adversarial_detections": 0,
45
+ "adversarial_poisonings": 0,
46
+ "final_trust": {
47
+ "S0": 0.473,
48
+ "S1": 0.888,
49
+ "S2": 0.5,
50
+ "S3": 0.5,
51
+ "S4": 0.5
52
+ }
53
+ },
54
+ {
55
+ "scenario_id": "SCN-TASK1-003",
56
+ "task_type": "task1",
57
+ "steps": 11,
58
+ "score": 0.885,
59
+ "total_reward": 10.62,
60
+ "completion_rate": 0.9,
61
+ "adversarial_detections": 0,
62
+ "adversarial_poisonings": 0,
63
+ "final_trust": {
64
+ "S0": 0.296,
65
+ "S1": 0.296,
66
+ "S2": 0.94,
67
+ "S3": 0.5,
68
+ "S4": 0.5
69
+ }
70
+ },
71
+ {
72
+ "scenario_id": "SCN-TASK1-004",
73
+ "task_type": "task1",
74
+ "steps": 8,
75
+ "score": 0.99,
76
+ "total_reward": 8.91,
77
+ "completion_rate": 0.8,
78
+ "adversarial_detections": 0,
79
+ "adversarial_poisonings": 0,
80
+ "final_trust": {
81
+ "S0": 0.931,
82
+ "S1": 0.5,
83
+ "S2": 0.5,
84
+ "S3": 0.5,
85
+ "S4": 0.5
86
+ }
87
+ },
88
+ {
89
+ "scenario_id": "SCN-TASK1-005",
90
+ "task_type": "task1",
91
+ "steps": 11,
92
+ "score": 0.9375,
93
+ "total_reward": 11.25,
94
+ "completion_rate": 1.0,
95
+ "adversarial_detections": 0,
96
+ "adversarial_poisonings": 0,
97
+ "final_trust": {
98
+ "S0": 0.86,
99
+ "S1": 0.5,
100
+ "S2": 0.5,
101
+ "S3": 0.5,
102
+ "S4": 0.5
103
+ }
104
+ },
105
+ {
106
+ "scenario_id": "SCN-TASK1-006",
107
+ "task_type": "task1",
108
+ "steps": 8,
109
+ "score": 0.85,
110
+ "total_reward": 7.65,
111
+ "completion_rate": 0.6,
112
+ "adversarial_detections": 0,
113
+ "adversarial_poisonings": 0,
114
+ "final_trust": {
115
+ "S0": 0.71,
116
+ "S1": 0.5,
117
+ "S2": 0.5,
118
+ "S3": 0.5,
119
+ "S4": 0.5
120
+ }
121
+ },
122
+ {
123
+ "scenario_id": "SCN-TASK1-007",
124
+ "task_type": "task1",
125
+ "steps": 10,
126
+ "score": 0.99,
127
+ "total_reward": 10.89,
128
+ "completion_rate": 1.0,
129
+ "adversarial_detections": 0,
130
+ "adversarial_poisonings": 0,
131
+ "final_trust": {
132
+ "S0": 0.943,
133
+ "S1": 0.5,
134
+ "S2": 0.5,
135
+ "S3": 0.5,
136
+ "S4": 0.5
137
+ }
138
+ },
139
+ {
140
+ "scenario_id": "SCN-TASK1-008",
141
+ "task_type": "task1",
142
+ "steps": 11,
143
+ "score": 0.8325,
144
+ "total_reward": 9.99,
145
+ "completion_rate": 0.8,
146
+ "adversarial_detections": 0,
147
+ "adversarial_poisonings": 0,
148
+ "final_trust": {
149
+ "S0": 0.482,
150
+ "S1": 0.9,
151
+ "S2": 0.5,
152
+ "S3": 0.5,
153
+ "S4": 0.5
154
+ }
155
+ },
156
+ {
157
+ "scenario_id": "SCN-TASK1-009",
158
+ "task_type": "task1",
159
+ "steps": 9,
160
+ "score": 0.864,
161
+ "total_reward": 8.64,
162
+ "completion_rate": 0.7,
163
+ "adversarial_detections": 0,
164
+ "adversarial_poisonings": 0,
165
+ "final_trust": {
166
+ "S0": 0.492,
167
+ "S1": 0.801,
168
+ "S2": 0.5,
169
+ "S3": 0.5,
170
+ "S4": 0.5
171
+ }
172
+ },
173
+ {
174
+ "scenario_id": "SCN-TASK1-010",
175
+ "task_type": "task1",
176
+ "steps": 12,
177
+ "score": 0.7962,
178
+ "total_reward": 10.35,
179
+ "completion_rate": 0.8,
180
+ "adversarial_detections": 0,
181
+ "adversarial_poisonings": 0,
182
+ "final_trust": {
183
+ "S0": 0.494,
184
+ "S1": 0.885,
185
+ "S2": 0.5,
186
+ "S3": 0.5,
187
+ "S4": 0.5
188
+ }
189
+ },
190
+ {
191
+ "scenario_id": "SCN-TASK2-001",
192
+ "task_type": "task2",
193
+ "steps": 19,
194
+ "score": 0.6054,
195
+ "total_reward": 12.1087,
196
+ "completion_rate": 0.8,
197
+ "adversarial_detections": 0,
198
+ "adversarial_poisonings": 0,
199
+ "final_trust": {
200
+ "S0": 0.476,
201
+ "S1": 0.26,
202
+ "S2": 0.717,
203
+ "S3": 0.5,
204
+ "S4": 0.5
205
+ }
206
+ },
207
+ {
208
+ "scenario_id": "SCN-TASK2-002",
209
+ "task_type": "task2",
210
+ "steps": 17,
211
+ "score": 0.7762,
212
+ "total_reward": 13.9711,
213
+ "completion_rate": 0.933,
214
+ "adversarial_detections": 0,
215
+ "adversarial_poisonings": 0,
216
+ "final_trust": {
217
+ "S0": 0.478,
218
+ "S1": 0.958,
219
+ "S2": 0.5,
220
+ "S3": 0.5,
221
+ "S4": 0.5
222
+ }
223
+ },
224
+ {
225
+ "scenario_id": "SCN-TASK2-003",
226
+ "task_type": "task2",
227
+ "steps": 17,
228
+ "score": 0.7377,
229
+ "total_reward": 13.2781,
230
+ "completion_rate": 0.867,
231
+ "adversarial_detections": 0,
232
+ "adversarial_poisonings": 0,
233
+ "final_trust": {
234
+ "S0": 0.289,
235
+ "S1": 0.289,
236
+ "S2": 0.818,
237
+ "S3": 0.5,
238
+ "S4": 0.5
239
+ }
240
+ },
241
+ {
242
+ "scenario_id": "SCN-TASK2-004",
243
+ "task_type": "task2",
244
+ "steps": 15,
245
+ "score": 0.7783,
246
+ "total_reward": 12.4521,
247
+ "completion_rate": 0.933,
248
+ "adversarial_detections": 0,
249
+ "adversarial_poisonings": 0,
250
+ "final_trust": {
251
+ "S0": 0.9,
252
+ "S1": 0.5,
253
+ "S2": 0.5,
254
+ "S3": 0.5,
255
+ "S4": 0.5
256
+ }
257
+ },
258
+ {
259
+ "scenario_id": "SCN-TASK2-005",
260
+ "task_type": "task2",
261
+ "steps": 17,
262
+ "score": 0.8174,
263
+ "total_reward": 14.7129,
264
+ "completion_rate": 1.0,
265
+ "adversarial_detections": 0,
266
+ "adversarial_poisonings": 0,
267
+ "final_trust": {
268
+ "S0": 0.849,
269
+ "S1": 0.5,
270
+ "S2": 0.5,
271
+ "S3": 0.5,
272
+ "S4": 0.5
273
+ }
274
+ },
275
+ {
276
+ "scenario_id": "SCN-TASK2-006",
277
+ "task_type": "task2",
278
+ "steps": 15,
279
+ "score": 0.6476,
280
+ "total_reward": 10.3617,
281
+ "completion_rate": 0.733,
282
+ "adversarial_detections": 0,
283
+ "adversarial_poisonings": 0,
284
+ "final_trust": {
285
+ "S0": 0.708,
286
+ "S1": 0.5,
287
+ "S2": 0.5,
288
+ "S3": 0.5,
289
+ "S4": 0.5
290
+ }
291
+ },
292
+ {
293
+ "scenario_id": "SCN-TASK2-007",
294
+ "task_type": "task2",
295
+ "steps": 15,
296
+ "score": 0.8967,
297
+ "total_reward": 14.3478,
298
+ "completion_rate": 1.0,
299
+ "adversarial_detections": 0,
300
+ "adversarial_poisonings": 0,
301
+ "final_trust": {
302
+ "S0": 0.967,
303
+ "S1": 0.5,
304
+ "S2": 0.5,
305
+ "S3": 0.5,
306
+ "S4": 0.5
307
+ }
308
+ },
309
+ {
310
+ "scenario_id": "SCN-TASK2-008",
311
+ "task_type": "task2",
312
+ "steps": 17,
313
+ "score": 0.7442,
314
+ "total_reward": 13.3953,
315
+ "completion_rate": 0.933,
316
+ "adversarial_detections": 0,
317
+ "adversarial_poisonings": 0,
318
+ "final_trust": {
319
+ "S0": 0.49,
320
+ "S1": 0.959,
321
+ "S2": 0.5,
322
+ "S3": 0.5,
323
+ "S4": 0.5
324
+ }
325
+ },
326
+ {
327
+ "scenario_id": "SCN-TASK2-009",
328
+ "task_type": "task2",
329
+ "steps": 16,
330
+ "score": 0.7525,
331
+ "total_reward": 12.792,
332
+ "completion_rate": 0.933,
333
+ "adversarial_detections": 0,
334
+ "adversarial_poisonings": 0,
335
+ "final_trust": {
336
+ "S0": 0.492,
337
+ "S1": 0.906,
338
+ "S2": 0.5,
339
+ "S3": 0.5,
340
+ "S4": 0.5
341
+ }
342
+ },
343
+ {
344
+ "scenario_id": "SCN-TASK2-010",
345
+ "task_type": "task2",
346
+ "steps": 18,
347
+ "score": 0.7191,
348
+ "total_reward": 13.6622,
349
+ "completion_rate": 0.933,
350
+ "adversarial_detections": 0,
351
+ "adversarial_poisonings": 0,
352
+ "final_trust": {
353
+ "S0": 0.474,
354
+ "S1": 0.955,
355
+ "S2": 0.5,
356
+ "S3": 0.5,
357
+ "S4": 0.5
358
+ }
359
+ },
360
+ {
361
+ "scenario_id": "SCN-TASK3-001",
362
+ "task_type": "task3",
363
+ "steps": 25,
364
+ "score": 0.7354,
365
+ "total_reward": 19.1204,
366
+ "completion_rate": 0.85,
367
+ "adversarial_detections": 0,
368
+ "adversarial_poisonings": 0,
369
+ "final_trust": {
370
+ "S0": 0.456,
371
+ "S1": 0.258,
372
+ "S2": 0.76,
373
+ "S3": 0.5,
374
+ "S4": 0.5
375
+ }
376
+ },
377
+ {
378
+ "scenario_id": "SCN-TASK3-002",
379
+ "task_type": "task3",
380
+ "steps": 25,
381
+ "score": 0.7054,
382
+ "total_reward": 18.341,
383
+ "completion_rate": 0.85,
384
+ "adversarial_detections": 3,
385
+ "adversarial_poisonings": 5,
386
+ "final_trust": {
387
+ "S0": 0.458,
388
+ "S1": 0.473,
389
+ "S2": 0.868,
390
+ "S3": 0.5,
391
+ "S4": 0.5
392
+ }
393
+ },
394
+ {
395
+ "scenario_id": "SCN-TASK3-003",
396
+ "task_type": "task3",
397
+ "steps": 19,
398
+ "score": 0.6438,
399
+ "total_reward": 12.8767,
400
+ "completion_rate": 0.6,
401
+ "adversarial_detections": 0,
402
+ "adversarial_poisonings": 5,
403
+ "final_trust": {
404
+ "S0": 0.299,
405
+ "S1": 0.299,
406
+ "S2": 0.633,
407
+ "S3": 0.5,
408
+ "S4": 0.5
409
+ }
410
+ },
411
+ {
412
+ "scenario_id": "SCN-TASK3-004",
413
+ "task_type": "task3",
414
+ "steps": 21,
415
+ "score": 0.8954,
416
+ "total_reward": 19.6992,
417
+ "completion_rate": 1.0,
418
+ "adversarial_detections": 0,
419
+ "adversarial_poisonings": 0,
420
+ "final_trust": {
421
+ "S0": 0.93,
422
+ "S1": 0.5,
423
+ "S2": 0.5,
424
+ "S3": 0.5,
425
+ "S4": 0.5
426
+ }
427
+ },
428
+ {
429
+ "scenario_id": "SCN-TASK3-005",
430
+ "task_type": "task3",
431
+ "steps": 24,
432
+ "score": 0.7134,
433
+ "total_reward": 17.8339,
434
+ "completion_rate": 0.85,
435
+ "adversarial_detections": 3,
436
+ "adversarial_poisonings": 6,
437
+ "final_trust": {
438
+ "S0": 0.491,
439
+ "S1": 0.797,
440
+ "S2": 0.5,
441
+ "S3": 0.5,
442
+ "S4": 0.5
443
+ }
444
+ },
445
+ {
446
+ "scenario_id": "SCN-TASK3-006",
447
+ "task_type": "task3",
448
+ "steps": 23,
449
+ "score": 0.7857,
450
+ "total_reward": 18.8578,
451
+ "completion_rate": 0.9,
452
+ "adversarial_detections": 0,
453
+ "adversarial_poisonings": 0,
454
+ "final_trust": {
455
+ "S0": 0.774,
456
+ "S1": 0.5,
457
+ "S2": 0.5,
458
+ "S3": 0.5,
459
+ "S4": 0.5
460
+ }
461
+ },
462
+ {
463
+ "scenario_id": "SCN-TASK3-007",
464
+ "task_type": "task3",
465
+ "steps": 24,
466
+ "score": 0.7045,
467
+ "total_reward": 17.6133,
468
+ "completion_rate": 0.85,
469
+ "adversarial_detections": 3,
470
+ "adversarial_poisonings": 7,
471
+ "final_trust": {
472
+ "S0": 0.498,
473
+ "S1": 0.5,
474
+ "S2": 0.5,
475
+ "S3": 0.5,
476
+ "S4": 0.5
477
+ }
478
+ },
479
+ {
480
+ "scenario_id": "SCN-TASK3-008",
481
+ "task_type": "task3",
482
+ "steps": 24,
483
+ "score": 0.8057,
484
+ "total_reward": 20.1435,
485
+ "completion_rate": 0.95,
486
+ "adversarial_detections": 0,
487
+ "adversarial_poisonings": 0,
488
+ "final_trust": {
489
+ "S0": 0.479,
490
+ "S1": 0.856,
491
+ "S2": 0.5,
492
+ "S3": 0.5,
493
+ "S4": 0.5
494
+ }
495
+ },
496
+ {
497
+ "scenario_id": "SCN-TASK3-009",
498
+ "task_type": "task3",
499
+ "steps": 23,
500
+ "score": 0.8456,
501
+ "total_reward": 20.2932,
502
+ "completion_rate": 1.0,
503
+ "adversarial_detections": 0,
504
+ "adversarial_poisonings": 0,
505
+ "final_trust": {
506
+ "S0": 0.488,
507
+ "S1": 0.891,
508
+ "S2": 0.5,
509
+ "S3": 0.5,
510
+ "S4": 0.5
511
+ }
512
+ },
513
+ {
514
+ "scenario_id": "SCN-TASK3-010",
515
+ "task_type": "task3",
516
+ "steps": 24,
517
+ "score": 0.8106,
518
+ "total_reward": 20.2645,
519
+ "completion_rate": 0.95,
520
+ "adversarial_detections": 0,
521
+ "adversarial_poisonings": 0,
522
+ "final_trust": {
523
+ "S0": 0.473,
524
+ "S1": 0.91,
525
+ "S2": 0.5,
526
+ "S3": 0.5,
527
+ "S4": 0.5
528
+ }
529
+ }
530
+ ]
531
+ }
outputs/charts/ablation.png ADDED
outputs/charts/baseline_delta_lines.png ADDED

Git LFS Details

  • SHA256: 734831a79790cc2f3e67d8ed7bd60acc8441ec80cf4c8df611da81ec7206acfb
  • Pointer size: 131 Bytes
  • Size of remote file: 264 kB
outputs/charts/baseline_grouped_bars.png ADDED
outputs/charts/cluster_health_policy_lines.png ADDED
outputs/charts/cluster_health_timeline.png ADDED
outputs/charts/detection_vs_poisoning.png ADDED
outputs/charts/failure_fishbone_map.png ADDED

Git LFS Details

  • SHA256: a6c1e20f5596eabae55b982ef046a1a3abd769155d8f6e74ce489c338691663e
  • Pointer size: 131 Bytes
  • Size of remote file: 193 kB
outputs/charts/grpo_reward_curve.png ADDED
outputs/charts/reward_component_stacked_area.png ADDED

Git LFS Details

  • SHA256: 868bd7d365f5c0e66c2a9cf32ef43175deb8f23f86fa139fb00618b2aa41ae2f
  • Pointer size: 131 Bytes
  • Size of remote file: 170 kB
outputs/charts/task_radar.png ADDED

Git LFS Details

  • SHA256: f3aab1649fe59d411b77f5bfac99aa7b2ca1650663ba0580d5cbdfd366db969d
  • Pointer size: 131 Bytes
  • Size of remote file: 249 kB
outputs/charts/trust_evolution.png ADDED
outputs/charts/trust_gap_over_time.png ADDED

Git LFS Details

  • SHA256: 860bd1703452fb7bb060d40c285ef4442dc9366f473ba4898285729c327ebd5b
  • Pointer size: 131 Bytes
  • Size of remote file: 121 kB
outputs/cluster_health_history.json ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task_type": "task3",
3
+ "seed": 42,
4
+ "series": {
5
+ "random": [
6
+ 0.5,
7
+ 0.5,
8
+ 0.5,
9
+ 0.5,
10
+ 0.5,
11
+ 0.5,
12
+ 0.5,
13
+ 0.5,
14
+ 0.5,
15
+ 0.5,
16
+ 0.5,
17
+ 0.5,
18
+ 0.5,
19
+ 0.5,
20
+ 0.5,
21
+ 0.5,
22
+ 0.5,
23
+ 0.5,
24
+ 0.5,
25
+ 0.5,
26
+ 0.5,
27
+ 0.5,
28
+ 0.5,
29
+ 0.5,
30
+ 0.5,
31
+ 0.5,
32
+ 0.5,
33
+ 0.5,
34
+ 0.5,
35
+ 0.5,
36
+ 0.5,
37
+ 0.5,
38
+ 0.5,
39
+ 0.5,
40
+ 0.5,
41
+ 0.5,
42
+ 0.5,
43
+ 0.5,
44
+ 0.5,
45
+ 0.5,
46
+ 0.5,
47
+ 0.5,
48
+ 0.5,
49
+ 0.5,
50
+ 0.5,
51
+ 0.5,
52
+ 0.5,
53
+ 0.5,
54
+ 0.5,
55
+ 0.5,
56
+ 0.5
57
+ ],
58
+ "heuristic": [
59
+ 0.5,
60
+ 0.5,
61
+ 0.5,
62
+ 0.5,
63
+ 0.5,
64
+ 0.5,
65
+ 0.5,
66
+ 0.5,
67
+ 0.5,
68
+ 0.5,
69
+ 0.5,
70
+ 0.5,
71
+ 0.5,
72
+ 0.5,
73
+ 0.5,
74
+ 0.5,
75
+ 0.5,
76
+ 0.5,
77
+ 0.5,
78
+ 0.5,
79
+ 0.5,
80
+ 0.5,
81
+ 0.5,
82
+ 0.5,
83
+ 0.5,
84
+ 0.5,
85
+ 0.5,
86
+ 0.5,
87
+ 0.5,
88
+ 0.5,
89
+ 0.5,
90
+ 0.5,
91
+ 0.5,
92
+ 0.5,
93
+ 0.5,
94
+ 0.5,
95
+ 0.5,
96
+ 0.5,
97
+ 0.5,
98
+ 0.5,
99
+ 0.5,
100
+ 0.5,
101
+ 0.5,
102
+ 0.5,
103
+ 0.5,
104
+ 0.5,
105
+ 0.5,
106
+ 0.5,
107
+ 0.5,
108
+ 0.5,
109
+ 0.5,
110
+ 0.5,
111
+ 0.5,
112
+ 0.5,
113
+ 0.5,
114
+ 0.5,
115
+ 0.5,
116
+ 0.5,
117
+ 0.5,
118
+ 0.5,
119
+ 0.5,
120
+ 0.5,
121
+ 0.5,
122
+ 0.5,
123
+ 0.5,
124
+ 0.5,
125
+ 0.5
126
+ ]
127
+ }
128
+ }
outputs/eval_post.json ADDED
The diff for this file is too large to render. See raw diff
 
outputs/eval_pre.json ADDED
The diff for this file is too large to render. See raw diff
 
outputs/evaluation_results.json ADDED
The diff for this file is too large to render. See raw diff
 
outputs/reward_report_task3_seed42.json ADDED
@@ -0,0 +1,526 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task_type": "task3",
3
+ "seed": 42,
4
+ "events": [
5
+ {
6
+ "step_count": 1,
7
+ "trust_snapshot": {
8
+ "S0": 0.706,
9
+ "S1": 0.5,
10
+ "S2": 0.5,
11
+ "S3": 0.5,
12
+ "S4": 0.5
13
+ },
14
+ "signal_breakdown": {
15
+ "task_accuracy": 1.0,
16
+ "stakes_awareness": 1.0,
17
+ "efficiency": 0.978,
18
+ "confidence_alignment": 1.0,
19
+ "verification_quality": 0.55,
20
+ "domain_routing": 0.5
21
+ },
22
+ "specialist_id": null,
23
+ "trust_after": null
24
+ },
25
+ {
26
+ "step_count": 2,
27
+ "trust_snapshot": {
28
+ "S0": 0.795,
29
+ "S1": 0.5,
30
+ "S2": 0.5,
31
+ "S3": 0.5,
32
+ "S4": 0.5
33
+ },
34
+ "signal_breakdown": {
35
+ "task_accuracy": 1.0,
36
+ "stakes_awareness": 1.0,
37
+ "efficiency": 0.956,
38
+ "confidence_alignment": 1.0,
39
+ "verification_quality": 0.55,
40
+ "domain_routing": 0.5
41
+ },
42
+ "specialist_id": null,
43
+ "trust_after": null
44
+ },
45
+ {
46
+ "step_count": 3,
47
+ "trust_snapshot": {
48
+ "S0": 0.847,
49
+ "S1": 0.5,
50
+ "S2": 0.5,
51
+ "S3": 0.5,
52
+ "S4": 0.5
53
+ },
54
+ "signal_breakdown": {
55
+ "task_accuracy": 1.0,
56
+ "stakes_awareness": 1.0,
57
+ "efficiency": 0.933,
58
+ "confidence_alignment": 1.0,
59
+ "verification_quality": 0.55,
60
+ "domain_routing": 0.5
61
+ },
62
+ "specialist_id": null,
63
+ "trust_after": null
64
+ },
65
+ {
66
+ "step_count": 4,
67
+ "trust_snapshot": {
68
+ "S0": 0.878,
69
+ "S1": 0.5,
70
+ "S2": 0.5,
71
+ "S3": 0.5,
72
+ "S4": 0.5
73
+ },
74
+ "signal_breakdown": {
75
+ "task_accuracy": 1.0,
76
+ "stakes_awareness": 1.0,
77
+ "efficiency": 0.911,
78
+ "confidence_alignment": 1.0,
79
+ "verification_quality": 0.55,
80
+ "domain_routing": 0.5
81
+ },
82
+ "specialist_id": null,
83
+ "trust_after": null
84
+ },
85
+ {
86
+ "step_count": 5,
87
+ "trust_snapshot": {
88
+ "S0": 0.9,
89
+ "S1": 0.5,
90
+ "S2": 0.5,
91
+ "S3": 0.5,
92
+ "S4": 0.5
93
+ },
94
+ "signal_breakdown": {
95
+ "task_accuracy": 1.0,
96
+ "stakes_awareness": 1.0,
97
+ "efficiency": 0.889,
98
+ "confidence_alignment": 1.0,
99
+ "verification_quality": 0.55,
100
+ "domain_routing": 0.5
101
+ },
102
+ "specialist_id": null,
103
+ "trust_after": null
104
+ },
105
+ {
106
+ "step_count": 6,
107
+ "trust_snapshot": {
108
+ "S0": 0.913,
109
+ "S1": 0.5,
110
+ "S2": 0.5,
111
+ "S3": 0.5,
112
+ "S4": 0.5
113
+ },
114
+ "signal_breakdown": {
115
+ "task_accuracy": 1.0,
116
+ "stakes_awareness": 1.0,
117
+ "efficiency": 0.867,
118
+ "confidence_alignment": 1.0,
119
+ "verification_quality": 0.55,
120
+ "domain_routing": 0.5
121
+ },
122
+ "specialist_id": null,
123
+ "trust_after": null
124
+ },
125
+ {
126
+ "step_count": 7,
127
+ "trust_snapshot": {
128
+ "S0": 0.786,
129
+ "S1": 0.5,
130
+ "S2": 0.5,
131
+ "S3": 0.5,
132
+ "S4": 0.5
133
+ },
134
+ "signal_breakdown": {
135
+ "task_accuracy": 0.0,
136
+ "stakes_awareness": 0.5,
137
+ "efficiency": 0.844,
138
+ "confidence_alignment": 0.12,
139
+ "verification_quality": 0.55,
140
+ "domain_routing": 0.5
141
+ },
142
+ "specialist_id": null,
143
+ "trust_after": null
144
+ },
145
+ {
146
+ "step_count": 8,
147
+ "trust_snapshot": {
148
+ "S0": 0.812,
149
+ "S1": 0.5,
150
+ "S2": 0.5,
151
+ "S3": 0.5,
152
+ "S4": 0.5
153
+ },
154
+ "signal_breakdown": {
155
+ "task_accuracy": 1.0,
156
+ "stakes_awareness": 1.0,
157
+ "efficiency": 0.822,
158
+ "confidence_alignment": 1.0,
159
+ "verification_quality": 0.55,
160
+ "domain_routing": 0.5
161
+ },
162
+ "specialist_id": null,
163
+ "trust_after": null
164
+ },
165
+ {
166
+ "step_count": 9,
167
+ "trust_snapshot": {
168
+ "S0": 0.834,
169
+ "S1": 0.5,
170
+ "S2": 0.5,
171
+ "S3": 0.5,
172
+ "S4": 0.5
173
+ },
174
+ "signal_breakdown": {
175
+ "task_accuracy": 1.0,
176
+ "stakes_awareness": 1.0,
177
+ "efficiency": 0.8,
178
+ "confidence_alignment": 1.0,
179
+ "verification_quality": 0.55,
180
+ "domain_routing": 0.5
181
+ },
182
+ "specialist_id": null,
183
+ "trust_after": null
184
+ },
185
+ {
186
+ "step_count": 10,
187
+ "trust_snapshot": {
188
+ "S0": 0.852,
189
+ "S1": 0.5,
190
+ "S2": 0.5,
191
+ "S3": 0.5,
192
+ "S4": 0.5
193
+ },
194
+ "signal_breakdown": {
195
+ "task_accuracy": 1.0,
196
+ "stakes_awareness": 1.0,
197
+ "efficiency": 0.778,
198
+ "confidence_alignment": 1.0,
199
+ "verification_quality": 0.55,
200
+ "domain_routing": 0.5
201
+ },
202
+ "specialist_id": null,
203
+ "trust_after": null
204
+ },
205
+ {
206
+ "step_count": 11,
207
+ "trust_snapshot": {
208
+ "S0": 0.865,
209
+ "S1": 0.5,
210
+ "S2": 0.5,
211
+ "S3": 0.5,
212
+ "S4": 0.5
213
+ },
214
+ "signal_breakdown": {
215
+ "task_accuracy": 1.0,
216
+ "stakes_awareness": 1.0,
217
+ "efficiency": 0.756,
218
+ "confidence_alignment": 1.0,
219
+ "verification_quality": 0.55,
220
+ "domain_routing": 0.5
221
+ },
222
+ "specialist_id": null,
223
+ "trust_after": null
224
+ },
225
+ {
226
+ "step_count": 12,
227
+ "trust_snapshot": {
228
+ "S0": 0.878,
229
+ "S1": 0.5,
230
+ "S2": 0.5,
231
+ "S3": 0.5,
232
+ "S4": 0.5
233
+ },
234
+ "signal_breakdown": {
235
+ "task_accuracy": 1.0,
236
+ "stakes_awareness": 1.0,
237
+ "efficiency": 0.733,
238
+ "confidence_alignment": 1.0,
239
+ "verification_quality": 0.55,
240
+ "domain_routing": 0.5
241
+ },
242
+ "specialist_id": null,
243
+ "trust_after": null
244
+ },
245
+ {
246
+ "step_count": 13,
247
+ "trust_snapshot": {
248
+ "S0": 0.888,
249
+ "S1": 0.5,
250
+ "S2": 0.5,
251
+ "S3": 0.5,
252
+ "S4": 0.5
253
+ },
254
+ "signal_breakdown": {
255
+ "task_accuracy": 1.0,
256
+ "stakes_awareness": 1.0,
257
+ "efficiency": 0.711,
258
+ "confidence_alignment": 1.0,
259
+ "verification_quality": 0.55,
260
+ "domain_routing": 0.5
261
+ },
262
+ "specialist_id": null,
263
+ "trust_after": null
264
+ },
265
+ {
266
+ "step_count": 14,
267
+ "trust_snapshot": {
268
+ "S0": 0.897,
269
+ "S1": 0.5,
270
+ "S2": 0.5,
271
+ "S3": 0.5,
272
+ "S4": 0.5
273
+ },
274
+ "signal_breakdown": {
275
+ "task_accuracy": 1.0,
276
+ "stakes_awareness": 1.0,
277
+ "efficiency": 0.689,
278
+ "confidence_alignment": 1.0,
279
+ "verification_quality": 0.55,
280
+ "domain_routing": 0.5
281
+ },
282
+ "specialist_id": null,
283
+ "trust_after": null
284
+ },
285
+ {
286
+ "step_count": 15,
287
+ "trust_snapshot": {
288
+ "S0": 0.825,
289
+ "S1": 0.5,
290
+ "S2": 0.5,
291
+ "S3": 0.5,
292
+ "S4": 0.5
293
+ },
294
+ "signal_breakdown": {
295
+ "task_accuracy": 0.0,
296
+ "stakes_awareness": 0.0,
297
+ "efficiency": 0.667,
298
+ "confidence_alignment": 0.08,
299
+ "verification_quality": 0.0,
300
+ "domain_routing": 0.5
301
+ },
302
+ "specialist_id": null,
303
+ "trust_after": null
304
+ },
305
+ {
306
+ "step_count": 16,
307
+ "trust_snapshot": {
308
+ "S0": 0.763,
309
+ "S1": 0.5,
310
+ "S2": 0.5,
311
+ "S3": 0.5,
312
+ "S4": 0.5
313
+ },
314
+ "signal_breakdown": {
315
+ "task_accuracy": 0.0,
316
+ "stakes_awareness": 0.0,
317
+ "efficiency": 0.644,
318
+ "confidence_alignment": 0.08,
319
+ "verification_quality": 0.0,
320
+ "domain_routing": 0.5
321
+ },
322
+ "specialist_id": null,
323
+ "trust_after": null
324
+ },
325
+ {
326
+ "step_count": 17,
327
+ "trust_snapshot": {
328
+ "S0": 0.708,
329
+ "S1": 0.5,
330
+ "S2": 0.5,
331
+ "S3": 0.5,
332
+ "S4": 0.5
333
+ },
334
+ "signal_breakdown": {
335
+ "task_accuracy": 0.0,
336
+ "stakes_awareness": 0.0,
337
+ "efficiency": 0.622,
338
+ "confidence_alignment": 0.08,
339
+ "verification_quality": 0.0,
340
+ "domain_routing": 0.5
341
+ },
342
+ "specialist_id": null,
343
+ "trust_after": null
344
+ },
345
+ {
346
+ "step_count": 18,
347
+ "trust_snapshot": {
348
+ "S0": 0.659,
349
+ "S1": 0.5,
350
+ "S2": 0.5,
351
+ "S3": 0.5,
352
+ "S4": 0.5
353
+ },
354
+ "signal_breakdown": {
355
+ "task_accuracy": 0.0,
356
+ "stakes_awareness": 0.0,
357
+ "efficiency": 0.6,
358
+ "confidence_alignment": 0.08,
359
+ "verification_quality": 0.0,
360
+ "domain_routing": 0.5
361
+ },
362
+ "specialist_id": null,
363
+ "trust_after": null
364
+ },
365
+ {
366
+ "step_count": 19,
367
+ "trust_snapshot": {
368
+ "S0": 0.619,
369
+ "S1": 0.5,
370
+ "S2": 0.5,
371
+ "S3": 0.5,
372
+ "S4": 0.5
373
+ },
374
+ "signal_breakdown": {
375
+ "task_accuracy": 0.0,
376
+ "stakes_awareness": 0.0,
377
+ "efficiency": 0.578,
378
+ "confidence_alignment": 0.08,
379
+ "verification_quality": 0.0,
380
+ "domain_routing": 0.5
381
+ },
382
+ "specialist_id": null,
383
+ "trust_after": null
384
+ },
385
+ {
386
+ "step_count": 21,
387
+ "trust_snapshot": {
388
+ "S0": 0.583,
389
+ "S1": 0.5,
390
+ "S2": 0.5,
391
+ "S3": 0.5,
392
+ "S4": 0.5
393
+ },
394
+ "signal_breakdown": {
395
+ "task_accuracy": 1.0,
396
+ "stakes_awareness": 0.99,
397
+ "efficiency": 0.533,
398
+ "confidence_alignment": 1.0,
399
+ "verification_quality": 1.0,
400
+ "domain_routing": 0.5
401
+ },
402
+ "specialist_id": null,
403
+ "trust_after": null
404
+ },
405
+ {
406
+ "step_count": 23,
407
+ "trust_snapshot": {
408
+ "S0": 0.551,
409
+ "S1": 0.5,
410
+ "S2": 0.5,
411
+ "S3": 0.5,
412
+ "S4": 0.5
413
+ },
414
+ "signal_breakdown": {
415
+ "task_accuracy": 1.0,
416
+ "stakes_awareness": 0.99,
417
+ "efficiency": 0.489,
418
+ "confidence_alignment": 1.0,
419
+ "verification_quality": 1.0,
420
+ "domain_routing": 0.5
421
+ },
422
+ "specialist_id": null,
423
+ "trust_after": null
424
+ },
425
+ {
426
+ "step_count": 25,
427
+ "trust_snapshot": {
428
+ "S0": 0.52,
429
+ "S1": 0.5,
430
+ "S2": 0.5,
431
+ "S3": 0.5,
432
+ "S4": 0.5
433
+ },
434
+ "signal_breakdown": {
435
+ "task_accuracy": 1.0,
436
+ "stakes_awareness": 0.99,
437
+ "efficiency": 0.444,
438
+ "confidence_alignment": 1.0,
439
+ "verification_quality": 1.0,
440
+ "domain_routing": 0.5
441
+ },
442
+ "specialist_id": null,
443
+ "trust_after": null
444
+ },
445
+ {
446
+ "step_count": 27,
447
+ "trust_snapshot": {
448
+ "S0": 0.492,
449
+ "S1": 0.5,
450
+ "S2": 0.5,
451
+ "S3": 0.5,
452
+ "S4": 0.5
453
+ },
454
+ "signal_breakdown": {
455
+ "task_accuracy": 1.0,
456
+ "stakes_awareness": 0.99,
457
+ "efficiency": 0.4,
458
+ "confidence_alignment": 1.0,
459
+ "verification_quality": 1.0,
460
+ "domain_routing": 0.5
461
+ },
462
+ "specialist_id": null,
463
+ "trust_after": null
464
+ },
465
+ {
466
+ "step_count": 29,
467
+ "trust_snapshot": {
468
+ "S0": 0.492,
469
+ "S1": 0.207,
470
+ "S2": 0.5,
471
+ "S3": 0.5,
472
+ "S4": 0.5
473
+ },
474
+ "signal_breakdown": {
475
+ "task_accuracy": 0.0,
476
+ "stakes_awareness": 0.5,
477
+ "efficiency": 0.356,
478
+ "confidence_alignment": 0.05,
479
+ "verification_quality": 0.75,
480
+ "domain_routing": 0.5
481
+ },
482
+ "specialist_id": null,
483
+ "trust_after": null
484
+ },
485
+ {
486
+ "step_count": 31,
487
+ "trust_snapshot": {
488
+ "S0": 0.492,
489
+ "S1": 0.207,
490
+ "S2": 0.207,
491
+ "S3": 0.5,
492
+ "S4": 0.5
493
+ },
494
+ "signal_breakdown": {
495
+ "task_accuracy": 0.0,
496
+ "stakes_awareness": 0.5,
497
+ "efficiency": 0.311,
498
+ "confidence_alignment": 0.55,
499
+ "verification_quality": 0.75,
500
+ "domain_routing": 0.1
501
+ },
502
+ "specialist_id": null,
503
+ "trust_after": null
504
+ },
505
+ {
506
+ "step_count": 31,
507
+ "trust_snapshot": {
508
+ "S0": 0.492,
509
+ "S1": 0.207,
510
+ "S2": 0.207,
511
+ "S3": 0.5,
512
+ "S4": 0.5
513
+ },
514
+ "signal_breakdown": {
515
+ "completion_rate": 0.85,
516
+ "detection_rate": 0.444,
517
+ "trust_calibration": 0.442,
518
+ "efficiency": 0.311,
519
+ "adversarial_detections": 4,
520
+ "adversarial_poisonings": 5
521
+ },
522
+ "specialist_id": null,
523
+ "trust_after": null
524
+ }
525
+ ]
526
+ }
outputs/trained_policy_replay.jsonl ADDED
The diff for this file is too large to render. See raw diff