Forcing-Consistency Training Constraint / results.json
Mechanism confirmed, baseline not beaten
1{
2 "seed": 1215,
3 "math_check": {
4 "intersection_equivalence": true,
5 "js_zero_equal": 0.0,
6 "js_positive_unequal": 0.27539611524877045,
7 "js_symmetric_error": 0.0,
8 "js_upper_bound_log2": 0.41775106531117484
9 },
10 "safety_structure": {
11 "feasible_intersection": [
12 0
13 ],
14 "empty_intersection": [],
15 "empty_intersection_rate": 1.0
16 },
17 "baseline_task_only": {
18 "kappa": 0.6870363354682922,
19 "feasible_violation": 0.5,
20 "empty_robust_margin": 0.0,
21 "empty_intersection_certificate": true,
22 "task_ce": 0.0007453762227669358,
23 "q": [
24 [
25 0.9993000030517578,
26 0.000699999975040555
27 ],
28 [
29 0.000699999975040555,
30 0.9993000030517578
31 ],
32 [
33 0.9993000030517578,
34 0.000699999975040555
35 ],
36 [
37 0.000699999975040555,
38 0.9993000030517578
39 ]
40 ]
41 },
42 "consistency_sweep": {
43 "0.0": {
44 "kappa": 0.6870363354682922,
45 "feasible_violation": 0.5,
46 "empty_robust_margin": 0.0,
47 "empty_intersection_certificate": true,
48 "task_ce": 0.0007453762227669358,
49 "q": [
50 [
51 0.9993000030517578,
52 0.000699999975040555
53 ],
54 [
55 0.000699999975040555,
56 0.9993000030517578
57 ],
58 [
59 0.9993000030517578,
60 0.000699999975040555
61 ],
62 [
63 0.000699999975040555,
64 0.9993000030517578
65 ]
66 ]
67 },
68 "0.02": {
69 "kappa": 0.6863622069358826,
70 "feasible_violation": 0.5,
71 "empty_robust_margin": 0.0,
72 "empty_intersection_certificate": true,
73 "task_ce": 0.0008398343343287706,
74 "q": [
75 [
76 0.9991999864578247,
77 0.0007999999797903001
78 ],
79 [
80 0.0007999999797903001,
81 0.9991999864578247
82 ],
83 [
84 0.9991999864578247,
85 0.0007999999797903001
86 ],
87 [
88 0.0007999999797903001,
89 0.9991999864578247
90 ]
91 ]
92 },
93 "0.05": {
94 "kappa": 0.6850269436836243,
95 "feasible_violation": 0.5,
96 "empty_robust_margin": 0.0,
97 "empty_intersection_certificate": true,
98 "task_ce": 0.0010314626852050424,
99 "q": [
100 [
101 0.9990000128746033,
102 0.0010000000474974513
103 ],
104 [
105 0.0010000000474974513,
106 0.9990000128746033
107 ],
108 [
109 0.9990000128746033,
110 0.0010000000474974513
111 ],
112 [
113 0.0010000000474974513,
114 0.9990000128746033
115 ]
116 ]
117 },
118 "0.1": {
119 "kappa": 0.6812583208084106,
120 "feasible_violation": 0.5,
121 "empty_robust_margin": 0.0,
122 "empty_intersection_certificate": true,
123 "task_ce": 0.0015997005393728614,
124 "q": [
125 [
126 0.9983999729156494,
127 0.0015999999595806003
128 ],
129 [
130 0.0015999999595806003,
131 0.9983999729156494
132 ],
133 [
134 0.9983999729156494,
135 0.0015999999595806003
136 ],
137 [
138 0.0015999999595806003,
139 0.9983999729156494
140 ]
141 ]
142 },
143 "0.25": {
144 "kappa": 0.6077403426170349,
145 "feasible_violation": 0.5,
146 "empty_robust_margin": 0.0,
147 "empty_intersection_certificate": true,
148 "task_ce": 0.016967138275504112,
149 "q": [
150 [
151 0.9832000136375427,
152 0.01679999940097332
153 ],
154 [
155 0.01679999940097332,
156 0.9832000136375427
157 ],
158 [
159 0.9832000136375427,
160 0.01679999940097332
161 ],
162 [
163 0.01679999940097332,
164 0.9832000136375427
165 ]
166 ]
167 },
168 "0.5": {
169 "kappa": 0.37210187315940857,
170 "feasible_violation": 0.5,
171 "empty_robust_margin": 0.0,
172 "empty_intersection_certificate": true,
173 "task_ce": 0.10333021730184555,
174 "q": [
175 [
176 0.9017999768257141,
177 0.0982000008225441
178 ],
179 [
180 0.0982000008225441,
181 0.9017999768257141
182 ],
183 [
184 0.9017999768257141,
185 0.0982000008225441
186 ],
187 [
188 0.0982000008225441,
189 0.9017999768257141
190 ]
191 ]
192 },
193 "1.0": {
194 "kappa": 0.16902287304401398,
195 "feasible_violation": 0.5,
196 "empty_robust_margin": 0.0,
197 "empty_intersection_certificate": true,
198 "task_ce": 0.24565979838371277,
199 "q": [
200 [
201 0.7821999788284302,
202 0.21780000627040863
203 ],
204 [
205 0.21780000627040863,
206 0.7821999788284302
207 ],
208 [
209 0.7821999788284302,
210 0.21780000627040863
211 ],
212 [
213 0.21780000627040863,
214 0.7821999788284302
215 ]
216 ]
217 },
218 "2.0": {
219 "kappa": 0.06384560465812683,
220 "feasible_violation": 0.5,
221 "empty_robust_margin": 0.0,
222 "empty_intersection_certificate": true,
223 "task_ce": 0.390466570854187,
224 "q": [
225 [
226 0.6766999959945679,
227 0.32330000400543213
228 ],
229 [
230 0.32330000400543213,
231 0.6766999959945679
232 ],
233 [
234 0.6766999959945679,
235 0.32330000400543213
236 ],
237 [
238 0.32330000400543213,
239 0.6766999959945679
240 ]
241 ]
242 },
243 "5.0": {
244 "kappa": 0.014414504170417786,
245 "feasible_violation": 0.5,
246 "empty_robust_margin": 0.0,
247 "empty_intersection_certificate": true,
248 "task_ce": 0.5366716980934143,
249 "q": [
250 [
251 0.5846999883651733,
252 0.41530001163482666
253 ],
254 [
255 0.41530001163482666,
256 0.5846999883651733
257 ],
258 [
259 0.5846999883651733,
260 0.41530001163482666
261 ],
262 [
263 0.41530001163482666,
264 0.5846999883651733
265 ]
266 ]
267 }
268 },
269 "safety_sweep": {
270 "0.0": {
271 "kappa": 0.16902287304401398,
272 "feasible_violation": 0.5,
273 "empty_robust_margin": 0.0,
274 "empty_intersection_certificate": true,
275 "task_ce": 0.24565979838371277,
276 "q": [
277 [
278 0.7821999788284302,
279 0.21780000627040863
280 ],
281 [
282 0.21780000627040863,
283 0.7821999788284302
284 ],
285 [
286 0.7821999788284302,
287 0.21780000627040863
288 ],
289 [
290 0.21780000627040863,
291 0.7821999788284302
292 ]
293 ]
294 },
295 "0.02": {
296 "kappa": 0.16899904608726501,
297 "feasible_violation": 0.48859095573425293,
298 "empty_robust_margin": 0.0,
299 "empty_intersection_certificate": true,
300 "task_ce": 0.24579764902591705,
301 "q": [
302 [
303 0.7935000061988831,
304 0.20649999380111694
305 ],
306 [
307 0.22939999401569366,
308 0.7706000208854675
309 ],
310 [
311 0.7821999788284302,
312 0.21780000627040863
313 ],
314 [
315 0.21780000627040863,
316 0.7821999788284302
317 ]
318 ]
319 },
320 "0.05": {
321 "kappa": 0.16887471079826355,
322 "feasible_violation": 0.47155892848968506,
323 "empty_robust_margin": 0.0,
324 "empty_intersection_certificate": true,
325 "task_ce": 0.24651779234409332,
326 "q": [
327 [
328 0.8098000288009644,
329 0.19020000100135803
330 ],
331 [
332 0.24709999561309814,
333 0.7529000043869019
334 ],
335 [
336 0.7821999788284302,
337 0.21780000627040863
338 ],
339 [
340 0.21780000627040863,
341 0.7821999788284302
342 ]
343 ]
344 },
345 "0.1": {
346 "kappa": 0.16843822598457336,
347 "feasible_violation": 0.44368624687194824,
348 "empty_robust_margin": 0.0,
349 "empty_intersection_certificate": true,
350 "task_ce": 0.24904121458530426,
351 "q": [
352 [
353 0.8353000283241272,
354 0.1647000014781952
355 ],
356 [
357 0.2773999869823456,
358 0.722599983215332
359 ],
360 [
361 0.7821999788284302,
362 0.21780000627040863
363 ],
364 [
365 0.21780000627040863,
366 0.7821999788284302
367 ]
368 ]
369 },
370 "0.25": {
371 "kappa": 0.16566991806030273,
372 "feasible_violation": 0.36782193183898926,
373 "empty_robust_margin": 0.0,
374 "empty_intersection_certificate": true,
375 "task_ce": 0.26491066813468933,
376 "q": [
377 [
378 0.8963000178337097,
379 0.10369999706745148
380 ],
381 [
382 0.36800000071525574,
383 0.6320000290870667
384 ],
385 [
386 0.7821999788284302,
387 0.21780000627040863
388 ],
389 [
390 0.21780000627040863,
391 0.7821999788284302
392 ]
393 ]
394 },
395 "0.5": {
396 "kappa": 0.1583424210548401,
397 "feasible_violation": 0.2753860354423523,
398 "empty_robust_margin": 0.0,
399 "empty_intersection_certificate": true,
400 "task_ce": 0.3060367703437805,
401 "q": [
402 [
403 0.9532999992370605,
404 0.04670000076293945
405 ],
406 [
407 0.4959000051021576,
408 0.5041000247001648
409 ],
410 [
411 0.7821999788284302,
412 0.21780000627040863
413 ],
414 [
415 0.21780000627040863,
416 0.7821999788284302
417 ]
418 ]
419 },
420 "1.0": {
421 "kappa": 0.14298632740974426,
422 "feasible_violation": 0.17822718620300293,
423 "empty_robust_margin": 0.0,
424 "empty_intersection_certificate": true,
425 "task_ce": 0.3909050226211548,
426 "q": [
427 [
428 0.9894000291824341,
429 0.010599999688565731
430 ],
431 [
432 0.6541000008583069,
433 0.3458999991416931
434 ],
435 [
436 0.7821999788284302,
437 0.21780000627040863
438 ],
439 [
440 0.21780000627040863,
441 0.7821999788284302
442 ]
443 ]
444 },
445 "2.0": {
446 "kappa": 0.12127116322517395,
447 "feasible_violation": 0.10490959882736206,
448 "empty_robust_margin": 0.0,
449 "empty_intersection_certificate": true,
450 "task_ce": 0.5157949328422546,
451 "q": [
452 [
453 0.998199999332428,
454 0.0017999999690800905
455 ],
456 [
457 0.7919999957084656,
458 0.20800000429153442
459 ],
460 [
461 0.7821999788284302,
462 0.21780000627040863
463 ],
464 [
465 0.21780000627040863,
466 0.7821999788284302
467 ]
468 ]
469 },
470 "5.0": {
471 "kappa": 0.11885703355073929,
472 "feasible_violation": 0.10082364082336426,
473 "empty_robust_margin": 0.0,
474 "empty_intersection_certificate": true,
475 "task_ce": 0.5267858505249023,
476 "q": [
477 [
478 0.9976000189781189,
479 0.002400000113993883
480 ],
481 [
482 0.8008000254631042,
483 0.19920000433921814
484 ],
485 [
486 0.7821999788284302,
487 0.21780000627040863
488 ],
489 [
490 0.21780000627040863,
491 0.7821999788284302
492 ]
493 ]
494 }
495 },
496 "prediction_checks": {
497 "P1_js_zero_is_exact": true,
498 "P2_kappa_reduced_high_consistency": true,
499 "P2_monotone_kappa_fraction": 1.0,
500 "P3_feasible_violation_reduced_high_safety": true,
501 "P3_empty_intersection_always_certificate": true
502 },
503 "prediction_summary": {
504 "P1": "JS is exactly zero for equal distributions and positive for unequal ones.",
505 "P2": "Increasing the JS coefficient reduces observational disagreement kappa.",
506 "P3": "Increasing robust-safety weight reduces feasible-group unsafe probability; empty intersections stay explicitly infeasible."
507 }
508}