MI-Guided Latent Protection / bench_report.json
Failed on benchmark
1{
2 "bench_version": 1,
3 "track": "sequence",
4 "model": "transformer_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.001
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.5707684904336929
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.6878322213888168
23 },
24 {
25 "cfg": {
26 "lr": 0.006
27 },
28 "mean": 1.0211272835731506
29 }
30 ],
31 "full": {
32 "mean": 0.5598551481962204,
33 "std": 0.02523705019158861,
34 "per_seed": [
35 0.5729890465736389,
36 0.5925205945968628,
37 0.5500922203063965,
38 0.5674721002578735,
39 0.5691339373588562,
40 0.5163435935974121,
41 0.5841162204742432,
42 0.52617347240448
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.5598490461707115,
49 "std": 0.025273703128400025,
50 "per_seed": [
51 0.5730356574058533,
52 0.5925182104110718,
53 0.5501575469970703,
54 0.5674620866775513,
55 0.5691909790039062,
56 0.516342043876648,
57 0.5841178894042969,
58 0.5259679555892944
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": -6.102025508880615e-06,
64 "idea_wins": 4,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 4.661083221435547e-05,
68 -2.384185791015625e-06,
69 6.532669067382812e-05,
70 -1.0013580322265625e-05,
71 5.704164505004883e-05,
72 -1.5497207641601562e-06,
73 1.6689300537109375e-06,
74 -0.00020551681518554688
75 ],
76 "p_value": 0.8987,
77 "mde": 7.188999984046296e-05,
78 "mde_rel_pct": 0.012840821428914797,
79 "verdict": "no measurable effect",
80 "system_worked": false
81 },
82 "mechanism_signature": {
83 "method": "MI-gradient inverse variance at 64-d token latent",
84 "best_idea_cfg": {
85 "lr": 0.001
86 },
87 "idea_sweep": [
88 {
89 "cfg": {
90 "lr": 0.001
91 },
92 "result": {
93 "mean": 0.5598490461707115,
94 "std": 0.025273703128400025,
95 "per_seed": [
96 0.5730356574058533,
97 0.5925182104110718,
98 0.5501575469970703,
99 0.5674620866775513,
100 0.5691909790039062,
101 0.516342043876648,
102 0.5841178894042969,
103 0.5259679555892944
104 ],
105 "n": 8
106 }
107 },
108 {
109 "cfg": {
110 "lr": 0.003
111 },
112 "result": {
113 "mean": 0.6711730286478996,
114 "std": 0.11012775840681698,
115 "per_seed": [
116 0.5534999966621399,
117 0.57029789686203,
118 0.8852774500846863,
119 0.741381824016571,
120 0.5719724297523499,
121 0.75211101770401,
122 0.6959620714187622,
123 0.5988815426826477
124 ],
125 "n": 8
126 }
127 },
128 {
129 "cfg": {
130 "lr": 0.006
131 },
132 "result": {
133 "mean": 1.010021261870861,
134 "std": 0.1928573253470958,
135 "per_seed": [
136 0.9215983152389526,
137 0.815628170967102,
138 0.754155158996582,
139 0.9170165657997131,
140 1.3768141269683838,
141 0.983137845993042,
142 1.1562342643737793,
143 1.1555856466293335
144 ],
145 "n": 8
146 }
147 }
148 ],
149 "predicted_vs_observed": {
150 "metric": 0.5730356574058533,
151 "q": [
152 1.0012232065200806,
153 0.999777615070343,
154 0.9991872906684875,
155 1.001699447631836,
156 1.0009998083114624,
157 0.9999410510063171,
158 0.9996676445007324,
159 1.0007892847061157,
160 0.9993820190429688,
161 1.0001580715179443,
162 0.9999819993972778,
163 0.9993095397949219,
164 0.9994521141052246,
165 0.9997609853744507,
166 0.9999178647994995,
167 0.9990708827972412,
168 0.9990319609642029,
169 0.9996174573898315,
170 1.0006672143936157,
171 1.000030279159546,
172 1.000243902206421,
173 0.9987349510192871,
174 1.0009965896606445,
175 0.9990262389183044,
176 1.001396894454956,
177 0.9992781281471252,
178 0.99986332654953,
179 1.0000636577606201,
180 1.0003782510757446,
181 0.999755322933197,
182 0.9989909529685974,
183 1.0014028549194336,
184 0.9997116923332214,
185 1.0010855197906494,
186 1.0008490085601807,
187 0.9986003637313843,
188 0.9999604225158691,
189 1.0011999607086182,
190 0.9982680082321167,
191 0.9994734525680542,
192 0.9993472695350647,
193 0.9998814463615417,
194 0.9986863732337952,
195 1.000624656677246,
196 0.9999277591705322,
197 0.9998847246170044,
198 1.000219464302063,
199 0.9992424845695496,
200 0.9991764426231384,
201 0.9996323585510254,
202 0.999434769153595,
203 1.0023554563522339,
204 1.0001237392425537,
205 1.0017119646072388,
206 0.9994488954544067,
207 1.0003513097763062,
208 1.0015864372253418,
209 0.9987058639526367,
210 1.0016520023345947,
211 1.0003422498703003,
212 0.9981855750083923,
213 1.000388741493225,
214 0.999549150466919,
215 1.0005910396575928
216 ],
217 "variances": [
218 0.19975546002388,
219 0.20004430413246155,
220 0.20016248524188995,
221 0.1996604949235916,
222 0.199800044298172,
223 0.20001161098480225,
224 0.20006629824638367,
225 0.19984206557273865,
226 0.20012347400188446,
227 0.19996818900108337,
228 0.20000341534614563,
229 0.20013800263404846,
230 0.20010945200920105,
231 0.20004761219024658,
232 0.20001623034477234,
233 0.20018580555915833,
234 0.2001935988664627,
235 0.20007634162902832,
236 0.19986644387245178,
237 0.19999375939369202,
238 0.19995103776454926,
239 0.20025312900543213,
240 0.19980069994926453,
241 0.20019474625587463,
242 0.1997208297252655,
243 0.20014429092407227,
244 0.20002715289592743,
245 0.19998708367347717,
246 0.19992418587207794,
247 0.2000487595796585,
248 0.2002018243074417,
249 0.19971963763237,
250 0.2000574767589569,
251 0.19978293776512146,
252 0.19983014464378357,
253 0.2002801150083542,
254 0.20000772178173065,
255 0.19976010918617249,
256 0.20034679770469666,
257 0.20010516047477722,
258 0.200130432844162,
259 0.20002353191375732,
260 0.2002628743648529,
261 0.19987496733665466,
262 0.20001426339149475,
263 0.2000228613615036,
264 0.19995592534542084,
265 0.20015141367912292,
266 0.20016464591026306,
267 0.20007336139678955,
268 0.20011290907859802,
269 0.19952982664108276,
270 0.19997505843639374,
271 0.19965800642967224,
272 0.20011010766029358,
273 0.19992956519126892,
274 0.1996830254793167,
275 0.20025897026062012,
276 0.1996699571609497,
277 0.19993138313293457,
278 0.20036335289478302,
279 0.1999220997095108,
280 0.20009002089500427,
281 0.1998816728591919
282 ],
283 "predicted_latent_noise_power": 12.799999237060547,
284 "observed_output_perturbation_mse": 0.06943081552162766,
285 "observed_output_perturbation_std": 0.006431090703341278,
286 "predicted_vs_observed_ratio": 0.00542428278593961,
287 "confirmed": true
288 }
289 },
290 "paired_seed_protocol": {
291 "seeds": [
292 0,
293 1,
294 2,
295 3,
296 4,
297 5,
298 6,
299 7
300 ],
301 "epochs": 8,
302 "batch": 128,
303 "average_variance": 0.2
304 }
305}