oof-baroomf commited on
Commit
26a2ef2
·
verified ·
1 Parent(s): 93d8ee1

Upload qwen3-soft-switch-fixed-controller

Browse files
qwen3-soft-switch-fixed-controller/SUMMARY.md ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Qwen3-4B held-out fixed-controller audit
2
+
3
+ - Model: `mlx-community/Qwen3-4B-4bit` at `4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25`.
4
+ - Both conditions use `enable_thinking=True` and the public `/think` versus `/no_think` soft switch.
5
+ - The controller is one fit-set mean `think - off` residual vector at layer boundary `28` and the shared `<think>` token. It is added for off-to-think and subtracted for think-to-off.
6
+ - The fit-set rule selected alpha `1.0` before held-out evaluation.
7
+ - Fit prompts: `12`. Held-out same-category prompts: `12`. Held-out category-shift prompts: `24`.
8
+ - Controller norm: `112.586914`; mean fit-prompt cosine to the mean controller: `0.9749`.
9
+
10
+ ## Held-out result
11
+
12
+ ### Held Out Same Category
13
+
14
+ - `off_to_think`: donor win `1.000` (95% prompt-bootstrap CI `[1.000, 1.000]`), donor-token match `1.000`, mean JS-to-donor `0.0000`, mean donor margin `0.6931`.
15
+ - `think_to_off`: donor win `1.000` (95% prompt-bootstrap CI `[1.000, 1.000]`), donor-token match `1.000`, mean JS-to-donor `0.0000`, mean donor margin `0.6931`.
16
+
17
+ ### Held Out Category Shift
18
+
19
+ - `off_to_think`: donor win `1.000` (95% prompt-bootstrap CI `[1.000, 1.000]`), donor-token match `1.000`, mean JS-to-donor `0.0000`, mean donor margin `0.6931`.
20
+ - `think_to_off`: donor win `1.000` (95% prompt-bootstrap CI `[1.000, 1.000]`), donor-token match `1.000`, mean JS-to-donor `0.0000`, mean donor margin `0.6931`.
21
+
22
+ ## Controls
23
+
24
+ The artifact reports the unchanged target, the sign-reversed controller, three norm-matched random vectors, and the same controller moved four positions earlier. Random-vector rows are averaged within prompt before uncertainty is computed.
25
+
26
+ ## Interpretation boundary
27
+
28
+ This is a held-out response-entry audit. It tests whether one train-learned activation vector transfers at the first soft-switch branch decision. It does not test answer accuracy, reasoning quality, or continued residual control after that one decision.
qwen3-soft-switch-fixed-controller/alpha_selection.json ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "alpha": 0.25,
4
+ "count": 24,
5
+ "worse_direction_donor_win": 0.0,
6
+ "combined_donor_win": 0.0,
7
+ "combined_donor_token_match": 0.0,
8
+ "mean_donor_margin": -0.6931458687434165,
9
+ "mean_js_to_donor": 0.693145921225986,
10
+ "directional_donor_win": {
11
+ "off_to_think": 0.0,
12
+ "think_to_off": 0.0
13
+ }
14
+ },
15
+ {
16
+ "alpha": 0.5,
17
+ "count": 24,
18
+ "worse_direction_donor_win": 0.0,
19
+ "combined_donor_win": 0.5,
20
+ "combined_donor_token_match": 0.5,
21
+ "mean_donor_margin": -0.08758681809626101,
22
+ "mean_js_to_donor": 0.36801834077284395,
23
+ "directional_donor_win": {
24
+ "off_to_think": 1.0,
25
+ "think_to_off": 0.0
26
+ }
27
+ },
28
+ {
29
+ "alpha": 1.0,
30
+ "count": 24,
31
+ "worse_direction_donor_win": 1.0,
32
+ "combined_donor_win": 1.0,
33
+ "combined_donor_token_match": 1.0,
34
+ "mean_donor_margin": 0.6931444624197308,
35
+ "mean_js_to_donor": 1.2803966101695965e-07,
36
+ "directional_donor_win": {
37
+ "off_to_think": 1.0,
38
+ "think_to_off": 1.0
39
+ }
40
+ },
41
+ {
42
+ "alpha": 2.0,
43
+ "count": 24,
44
+ "worse_direction_donor_win": 1.0,
45
+ "combined_donor_win": 1.0,
46
+ "combined_donor_token_match": 0.875,
47
+ "mean_donor_margin": 0.6076891112758444,
48
+ "mean_js_to_donor": 0.08545691789823312,
49
+ "directional_donor_win": {
50
+ "off_to_think": 1.0,
51
+ "think_to_off": 1.0
52
+ }
53
+ },
54
+ {
55
+ "alpha": 4.0,
56
+ "count": 24,
57
+ "worse_direction_donor_win": 1.0,
58
+ "combined_donor_win": 1.0,
59
+ "combined_donor_token_match": 0.5,
60
+ "mean_donor_margin": 0.3460064210531057,
61
+ "mean_js_to_donor": 0.34713919171735846,
62
+ "directional_donor_win": {
63
+ "off_to_think": 1.0,
64
+ "think_to_off": 1.0
65
+ }
66
+ }
67
+ ]
qwen3-soft-switch-fixed-controller/audit_aggregate.json ADDED
@@ -0,0 +1,642 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "split": "held_out_category_shift",
4
+ "direction": "off_to_think",
5
+ "method": "fixed_earlier_position",
6
+ "n_prompts": 24,
7
+ "n_rows": 24,
8
+ "mean_donor_win": 0.0,
9
+ "donor_win_ci95": [
10
+ 0.0,
11
+ 0.0
12
+ ],
13
+ "mean_donor_token_match": 0.0,
14
+ "donor_token_match_ci95": [
15
+ 0.0,
16
+ 0.0
17
+ ],
18
+ "mean_donor_margin": -0.6931469310236746,
19
+ "donor_margin_ci95": [
20
+ -0.6931470053683065,
21
+ -0.6931468527762618
22
+ ],
23
+ "mean_js_to_donor": 0.6931469313403071,
24
+ "js_to_donor_ci95": [
25
+ 0.6931468529970968,
26
+ 0.6931470056915587
27
+ ],
28
+ "mean_js_to_target": 3.166325131574721e-10,
29
+ "js_to_target_ci95": [
30
+ 1.7989732507867666e-10,
31
+ 4.92151503506408e-10
32
+ ]
33
+ },
34
+ {
35
+ "split": "held_out_category_shift",
36
+ "direction": "off_to_think",
37
+ "method": "fixed_signed",
38
+ "n_prompts": 24,
39
+ "n_rows": 24,
40
+ "mean_donor_win": 1.0,
41
+ "donor_win_ci95": [
42
+ 1.0,
43
+ 1.0
44
+ ],
45
+ "mean_donor_token_match": 1.0,
46
+ "donor_token_match_ci95": [
47
+ 1.0,
48
+ 1.0
49
+ ],
50
+ "mean_donor_margin": 0.693146829350202,
51
+ "donor_margin_ci95": [
52
+ 0.6931467705455445,
53
+ 0.6931468868609774
54
+ ],
55
+ "mean_js_to_donor": 1.0254671702073614e-08,
56
+ "js_to_donor_ci95": [
57
+ 7.168476255223462e-09,
58
+ 1.3469075570145597e-08
59
+ ],
60
+ "mean_js_to_target": 0.6931468396048738,
61
+ "js_to_target_ci95": [
62
+ 0.6931467811059991,
63
+ 0.6931468975388017
64
+ ]
65
+ },
66
+ {
67
+ "split": "held_out_category_shift",
68
+ "direction": "off_to_think",
69
+ "method": "opposite_signed",
70
+ "n_prompts": 24,
71
+ "n_rows": 24,
72
+ "mean_donor_win": 0.0,
73
+ "donor_win_ci95": [
74
+ 0.0,
75
+ 0.0
76
+ ],
77
+ "mean_donor_token_match": 0.0,
78
+ "donor_token_match_ci95": [
79
+ 0.0,
80
+ 0.0
81
+ ],
82
+ "mean_donor_margin": -0.664232920150759,
83
+ "donor_margin_ci95": [
84
+ -0.6746430907542258,
85
+ -0.6518893759410646
86
+ ],
87
+ "mean_js_to_donor": 0.6931394897419886,
88
+ "js_to_donor_ci95": [
89
+ 0.6931375610578059,
90
+ 0.6931411489713218
91
+ ],
92
+ "mean_js_to_target": 0.02890656959122954,
93
+ "js_to_target_ci95": [
94
+ 0.018496208935534723,
95
+ 0.04124893838680243
96
+ ]
97
+ },
98
+ {
99
+ "split": "held_out_category_shift",
100
+ "direction": "off_to_think",
101
+ "method": "random_norm_matched",
102
+ "n_prompts": 24,
103
+ "n_rows": 72,
104
+ "mean_donor_win": 0.0,
105
+ "donor_win_ci95": [
106
+ 0.0,
107
+ 0.0
108
+ ],
109
+ "mean_donor_token_match": 0.0,
110
+ "donor_token_match_ci95": [
111
+ 0.0,
112
+ 0.0
113
+ ],
114
+ "mean_donor_margin": -0.6931388083093241,
115
+ "donor_margin_ci95": [
116
+ -0.6931403102663989,
117
+ -0.6931371648004525
118
+ ],
119
+ "mean_js_to_donor": 0.693140514046355,
120
+ "js_to_donor_ci95": [
121
+ 0.6931389605772241,
122
+ 0.6931418610883947
123
+ ],
124
+ "mean_js_to_target": 1.705737030983684e-06,
125
+ "js_to_target_ci95": [
126
+ 1.4077425674506097e-06,
127
+ 2.004925185860033e-06
128
+ ]
129
+ },
130
+ {
131
+ "split": "held_out_category_shift",
132
+ "direction": "off_to_think",
133
+ "method": "zero",
134
+ "n_prompts": 24,
135
+ "n_rows": 24,
136
+ "mean_donor_win": 0.0,
137
+ "donor_win_ci95": [
138
+ 0.0,
139
+ 0.0
140
+ ],
141
+ "mean_donor_token_match": 0.0,
142
+ "donor_token_match_ci95": [
143
+ 0.0,
144
+ 0.0
145
+ ],
146
+ "mean_donor_margin": -0.6931469541997154,
147
+ "donor_margin_ci95": [
148
+ -0.693147033870871,
149
+ -0.6931468717167344
150
+ ],
151
+ "mean_js_to_donor": 0.6931469541997154,
152
+ "js_to_donor_ci95": [
153
+ 0.6931468717167344,
154
+ 0.693147033870871
155
+ ],
156
+ "mean_js_to_target": 0.0,
157
+ "js_to_target_ci95": [
158
+ 0.0,
159
+ 0.0
160
+ ]
161
+ },
162
+ {
163
+ "split": "held_out_category_shift",
164
+ "direction": "think_to_off",
165
+ "method": "fixed_earlier_position",
166
+ "n_prompts": 24,
167
+ "n_rows": 24,
168
+ "mean_donor_win": 0.0,
169
+ "donor_win_ci95": [
170
+ 0.0,
171
+ 0.0
172
+ ],
173
+ "mean_donor_token_match": 0.0,
174
+ "donor_token_match_ci95": [
175
+ 0.0,
176
+ 0.0
177
+ ],
178
+ "mean_donor_margin": -0.6931469071121131,
179
+ "donor_margin_ci95": [
180
+ -0.6931469908741696,
181
+ -0.6931468193850381
182
+ ],
183
+ "mean_js_to_donor": 0.6931469073871609,
184
+ "js_to_donor_ci95": [
185
+ 0.6931468196351035,
186
+ 0.6931469911307749
187
+ ],
188
+ "mean_js_to_target": 2.750478486716539e-10,
189
+ "js_to_target_ci95": [
190
+ 1.8645810940810016e-10,
191
+ 3.7568983247762554e-10
192
+ ]
193
+ },
194
+ {
195
+ "split": "held_out_category_shift",
196
+ "direction": "think_to_off",
197
+ "method": "fixed_signed",
198
+ "n_prompts": 24,
199
+ "n_rows": 24,
200
+ "mean_donor_win": 1.0,
201
+ "donor_win_ci95": [
202
+ 1.0,
203
+ 1.0
204
+ ],
205
+ "mean_donor_token_match": 1.0,
206
+ "donor_token_match_ci95": [
207
+ 1.0,
208
+ 1.0
209
+ ],
210
+ "mean_donor_margin": 0.6931222213675711,
211
+ "donor_margin_ci95": [
212
+ 0.6930897001207836,
213
+ 0.6931425804519925
214
+ ],
215
+ "mean_js_to_donor": 1.421713493347396e-06,
216
+ "js_to_donor_ci95": [
217
+ 2.1004394593525438e-07,
218
+ 3.40794108683691e-06
219
+ ],
220
+ "mean_js_to_target": 0.6931236430810644,
221
+ "js_to_target_ci95": [
222
+ 0.6930930928537035,
223
+ 0.6931427898012532
224
+ ]
225
+ },
226
+ {
227
+ "split": "held_out_category_shift",
228
+ "direction": "think_to_off",
229
+ "method": "opposite_signed",
230
+ "n_prompts": 24,
231
+ "n_rows": 24,
232
+ "mean_donor_win": 0.0,
233
+ "donor_win_ci95": [
234
+ 0.0,
235
+ 0.0
236
+ ],
237
+ "mean_donor_token_match": 0.0,
238
+ "donor_token_match_ci95": [
239
+ 0.0,
240
+ 0.0
241
+ ],
242
+ "mean_donor_margin": -0.6931469113579148,
243
+ "donor_margin_ci95": [
244
+ -0.6931470142128139,
245
+ -0.6931468052194389
246
+ ],
247
+ "mean_js_to_donor": 0.6931469943336044,
248
+ "js_to_donor_ci95": [
249
+ 0.6931468972918383,
250
+ 0.6931470830664278
251
+ ],
252
+ "mean_js_to_target": 8.297568956906983e-08,
253
+ "js_to_target_ci95": [
254
+ 3.570710986457489e-08,
255
+ 1.4725269616774358e-07
256
+ ]
257
+ },
258
+ {
259
+ "split": "held_out_category_shift",
260
+ "direction": "think_to_off",
261
+ "method": "random_norm_matched",
262
+ "n_prompts": 24,
263
+ "n_rows": 72,
264
+ "mean_donor_win": 0.0,
265
+ "donor_win_ci95": [
266
+ 0.0,
267
+ 0.0
268
+ ],
269
+ "mean_donor_token_match": 0.0,
270
+ "donor_token_match_ci95": [
271
+ 0.0,
272
+ 0.0
273
+ ],
274
+ "mean_donor_margin": -0.6931468840069249,
275
+ "donor_margin_ci95": [
276
+ -0.6931470078629456,
277
+ -0.6931467427806919
278
+ ],
279
+ "mean_js_to_donor": 0.6931469150052996,
280
+ "js_to_donor_ci95": [
281
+ 0.6931467778346903,
282
+ 0.6931470351359542
283
+ ],
284
+ "mean_js_to_target": 3.099837463828269e-08,
285
+ "js_to_target_ci95": [
286
+ 2.1110269422350448e-08,
287
+ 4.248772125240107e-08
288
+ ]
289
+ },
290
+ {
291
+ "split": "held_out_category_shift",
292
+ "direction": "think_to_off",
293
+ "method": "zero",
294
+ "n_prompts": 24,
295
+ "n_rows": 24,
296
+ "mean_donor_win": 0.0,
297
+ "donor_win_ci95": [
298
+ 0.0,
299
+ 0.0
300
+ ],
301
+ "mean_donor_token_match": 0.0,
302
+ "donor_token_match_ci95": [
303
+ 0.0,
304
+ 0.0
305
+ ],
306
+ "mean_donor_margin": -0.6931469541997154,
307
+ "donor_margin_ci95": [
308
+ -0.6931470352250019,
309
+ -0.6931468694262328
310
+ ],
311
+ "mean_js_to_donor": 0.6931469541997154,
312
+ "js_to_donor_ci95": [
313
+ 0.6931468694262328,
314
+ 0.6931470352250019
315
+ ],
316
+ "mean_js_to_target": 0.0,
317
+ "js_to_target_ci95": [
318
+ 0.0,
319
+ 0.0
320
+ ]
321
+ },
322
+ {
323
+ "split": "held_out_same_category",
324
+ "direction": "off_to_think",
325
+ "method": "fixed_earlier_position",
326
+ "n_prompts": 12,
327
+ "n_rows": 12,
328
+ "mean_donor_win": 0.0,
329
+ "donor_win_ci95": [
330
+ 0.0,
331
+ 0.0
332
+ ],
333
+ "mean_donor_token_match": 0.0,
334
+ "donor_token_match_ci95": [
335
+ 0.0,
336
+ 0.0
337
+ ],
338
+ "mean_donor_margin": -0.6931470391228415,
339
+ "donor_margin_ci95": [
340
+ -0.6931471065838363,
341
+ -0.6931469649350727
342
+ ],
343
+ "mean_js_to_donor": 0.6931470392105079,
344
+ "js_to_donor_ci95": [
345
+ 0.6931469649959342,
346
+ 0.6931471066609561
347
+ ],
348
+ "mean_js_to_target": 8.766638254942828e-11,
349
+ "js_to_target_ci95": [
350
+ 4.2821808441330736e-11,
351
+ 1.501520934997215e-10
352
+ ]
353
+ },
354
+ {
355
+ "split": "held_out_same_category",
356
+ "direction": "off_to_think",
357
+ "method": "fixed_signed",
358
+ "n_prompts": 12,
359
+ "n_rows": 12,
360
+ "mean_donor_win": 1.0,
361
+ "donor_win_ci95": [
362
+ 1.0,
363
+ 1.0
364
+ ],
365
+ "mean_donor_token_match": 1.0,
366
+ "donor_token_match_ci95": [
367
+ 1.0,
368
+ 1.0
369
+ ],
370
+ "mean_donor_margin": 0.6931469492870609,
371
+ "donor_margin_ci95": [
372
+ 0.6931468728833705,
373
+ 0.6931470241212413
374
+ ],
375
+ "mean_js_to_donor": 3.62655579826069e-09,
376
+ "js_to_donor_ci95": [
377
+ 2.2390902763773736e-09,
378
+ 5.7179291137308715e-09
379
+ ],
380
+ "mean_js_to_target": 0.6931469529136166,
381
+ "js_to_target_ci95": [
382
+ 0.6931468779154742,
383
+ 0.6931470268905451
384
+ ]
385
+ },
386
+ {
387
+ "split": "held_out_same_category",
388
+ "direction": "off_to_think",
389
+ "method": "opposite_signed",
390
+ "n_prompts": 12,
391
+ "n_rows": 12,
392
+ "mean_donor_win": 0.0,
393
+ "donor_win_ci95": [
394
+ 0.0,
395
+ 0.0
396
+ ],
397
+ "mean_donor_token_match": 0.0,
398
+ "donor_token_match_ci95": [
399
+ 0.0,
400
+ 0.0
401
+ ],
402
+ "mean_donor_margin": -0.6826024343614879,
403
+ "donor_margin_ci95": [
404
+ -0.6875430173409902,
405
+ -0.6770453834194984
406
+ ],
407
+ "mean_js_to_donor": 0.6931427215058131,
408
+ "js_to_donor_ci95": [
409
+ 0.6931416928924672,
410
+ 0.6931437315671977
411
+ ],
412
+ "mean_js_to_target": 0.010540287144325094,
413
+ "js_to_target_ci95": [
414
+ 0.005600202010181046,
415
+ 0.016096925441729287
416
+ ]
417
+ },
418
+ {
419
+ "split": "held_out_same_category",
420
+ "direction": "off_to_think",
421
+ "method": "random_norm_matched",
422
+ "n_prompts": 12,
423
+ "n_rows": 36,
424
+ "mean_donor_win": 0.0,
425
+ "donor_win_ci95": [
426
+ 0.0,
427
+ 0.0
428
+ ],
429
+ "mean_donor_token_match": 0.0,
430
+ "donor_token_match_ci95": [
431
+ 0.0,
432
+ 0.0
433
+ ],
434
+ "mean_donor_margin": -0.693142238481503,
435
+ "donor_margin_ci95": [
436
+ -0.6931435988393708,
437
+ -0.6931405860850134
438
+ ],
439
+ "mean_js_to_donor": 0.6931428320020324,
440
+ "js_to_donor_ci95": [
441
+ 0.6931413326098694,
442
+ 0.6931440639377024
443
+ ],
444
+ "mean_js_to_target": 5.935205294278544e-07,
445
+ "js_to_target_ci95": [
446
+ 4.328733924193193e-07,
447
+ 7.655502291639151e-07
448
+ ]
449
+ },
450
+ {
451
+ "split": "held_out_same_category",
452
+ "direction": "off_to_think",
453
+ "method": "zero",
454
+ "n_prompts": 12,
455
+ "n_rows": 12,
456
+ "mean_donor_win": 0.0,
457
+ "donor_win_ci95": [
458
+ 0.0,
459
+ 0.0
460
+ ],
461
+ "mean_donor_token_match": 0.0,
462
+ "donor_token_match_ci95": [
463
+ 0.0,
464
+ 0.0
465
+ ],
466
+ "mean_donor_margin": -0.6931470405588983,
467
+ "donor_margin_ci95": [
468
+ -0.6931471076248473,
469
+ -0.693146964931182
470
+ ],
471
+ "mean_js_to_donor": 0.6931470405588983,
472
+ "js_to_donor_ci95": [
473
+ 0.693146964931182,
474
+ 0.6931471076248473
475
+ ],
476
+ "mean_js_to_target": 0.0,
477
+ "js_to_target_ci95": [
478
+ 0.0,
479
+ 0.0
480
+ ]
481
+ },
482
+ {
483
+ "split": "held_out_same_category",
484
+ "direction": "think_to_off",
485
+ "method": "fixed_earlier_position",
486
+ "n_prompts": 12,
487
+ "n_rows": 12,
488
+ "mean_donor_win": 0.0,
489
+ "donor_win_ci95": [
490
+ 0.0,
491
+ 0.0
492
+ ],
493
+ "mean_donor_token_match": 0.0,
494
+ "donor_token_match_ci95": [
495
+ 0.0,
496
+ 0.0
497
+ ],
498
+ "mean_donor_margin": -0.693147001153224,
499
+ "donor_margin_ci95": [
500
+ -0.693147085743529,
501
+ -0.6931469091064172
502
+ ],
503
+ "mean_js_to_donor": 0.6931470014360479,
504
+ "js_to_donor_ci95": [
505
+ 0.693146909588993,
506
+ 0.6931470859924309
507
+ ],
508
+ "mean_js_to_target": 2.828238285696571e-10,
509
+ "js_to_target_ci95": [
510
+ 1.0045687227564119e-10,
511
+ 4.849645496648741e-10
512
+ ]
513
+ },
514
+ {
515
+ "split": "held_out_same_category",
516
+ "direction": "think_to_off",
517
+ "method": "fixed_signed",
518
+ "n_prompts": 12,
519
+ "n_rows": 12,
520
+ "mean_donor_win": 1.0,
521
+ "donor_win_ci95": [
522
+ 1.0,
523
+ 1.0
524
+ ],
525
+ "mean_donor_token_match": 1.0,
526
+ "donor_token_match_ci95": [
527
+ 1.0,
528
+ 1.0
529
+ ],
530
+ "mean_donor_margin": 0.6931441571407341,
531
+ "donor_margin_ci95": [
532
+ 0.6931426917443885,
533
+ 0.6931453816588616
534
+ ],
535
+ "mean_js_to_donor": 1.3734455141439547e-07,
536
+ "js_to_donor_ci95": [
537
+ 8.224476379777067e-08,
538
+ 2.0484122422803148e-07
539
+ ],
540
+ "mean_js_to_target": 0.6931442944852856,
541
+ "js_to_target_ci95": [
542
+ 0.6931428994009677,
543
+ 0.6931454636560184
544
+ ]
545
+ },
546
+ {
547
+ "split": "held_out_same_category",
548
+ "direction": "think_to_off",
549
+ "method": "opposite_signed",
550
+ "n_prompts": 12,
551
+ "n_rows": 12,
552
+ "mean_donor_win": 0.0,
553
+ "donor_win_ci95": [
554
+ 0.0,
555
+ 0.0
556
+ ],
557
+ "mean_donor_token_match": 0.0,
558
+ "donor_token_match_ci95": [
559
+ 0.0,
560
+ 0.0
561
+ ],
562
+ "mean_donor_margin": -0.6931469080636986,
563
+ "donor_margin_ci95": [
564
+ -0.6931470585812479,
565
+ -0.6931467314567078
566
+ ],
567
+ "mean_js_to_donor": 0.6931469637884285,
568
+ "js_to_donor_ci95": [
569
+ 0.6931468137099002,
570
+ 0.6931470945657948
571
+ ],
572
+ "mean_js_to_target": 5.57247296562856e-08,
573
+ "js_to_target_ci95": [
574
+ 3.1679295320022627e-08,
575
+ 8.397671304444428e-08
576
+ ]
577
+ },
578
+ {
579
+ "split": "held_out_same_category",
580
+ "direction": "think_to_off",
581
+ "method": "random_norm_matched",
582
+ "n_prompts": 12,
583
+ "n_rows": 36,
584
+ "mean_donor_win": 0.0,
585
+ "donor_win_ci95": [
586
+ 0.0,
587
+ 0.0
588
+ ],
589
+ "mean_donor_token_match": 0.0,
590
+ "donor_token_match_ci95": [
591
+ 0.0,
592
+ 0.0
593
+ ],
594
+ "mean_donor_margin": -0.6931466880718892,
595
+ "donor_margin_ci95": [
596
+ -0.6931469659763809,
597
+ -0.6931463988517631
598
+ ],
599
+ "mean_js_to_donor": 0.6931467248435123,
600
+ "js_to_donor_ci95": [
601
+ 0.6931464460246677,
602
+ 0.6931469922693474
603
+ ],
604
+ "mean_js_to_target": 3.677162303641701e-08,
605
+ "js_to_target_ci95": [
606
+ 2.22479523035754e-08,
607
+ 5.118976411056396e-08
608
+ ]
609
+ },
610
+ {
611
+ "split": "held_out_same_category",
612
+ "direction": "think_to_off",
613
+ "method": "zero",
614
+ "n_prompts": 12,
615
+ "n_rows": 12,
616
+ "mean_donor_win": 0.0,
617
+ "donor_win_ci95": [
618
+ 0.0,
619
+ 0.0
620
+ ],
621
+ "mean_donor_token_match": 0.0,
622
+ "donor_token_match_ci95": [
623
+ 0.0,
624
+ 0.0
625
+ ],
626
+ "mean_donor_margin": -0.6931470405588983,
627
+ "donor_margin_ci95": [
628
+ -0.6931471081431074,
629
+ -0.6931469663043396
630
+ ],
631
+ "mean_js_to_donor": 0.6931470405588983,
632
+ "js_to_donor_ci95": [
633
+ 0.6931469663043396,
634
+ 0.6931471081431074
635
+ ],
636
+ "mean_js_to_target": 0.0,
637
+ "js_to_target_ci95": [
638
+ 0.0,
639
+ 0.0
640
+ ]
641
+ }
642
+ ]
qwen3-soft-switch-fixed-controller/audit_rows.json ADDED
The diff for this file is too large to render. See raw diff
 
qwen3-soft-switch-fixed-controller/completion.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "completed_at_utc": "2026-07-28T07:58:50.154853+00:00",
3
+ "selected_alpha": 1.0,
4
+ "row_count": 504,
5
+ "held_out_prompt_count": 36
6
+ }
qwen3-soft-switch-fixed-controller/controller_diagnostics.json ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "fit_count": 12,
3
+ "hidden_size": 2560,
4
+ "mean_vector_norm": 112.5869140625,
5
+ "individual_difference_norm_mean": 115.49103546142578,
6
+ "individual_difference_norm_range": [
7
+ 110.46172332763672,
8
+ 120.38420867919922
9
+ ],
10
+ "individual_cosine_to_mean": [
11
+ {
12
+ "id": "count_strawberries_r",
13
+ "cosine": 0.9632018804550171
14
+ },
15
+ {
16
+ "id": "count_committee_m",
17
+ "cosine": 0.9722326397895813
18
+ },
19
+ {
20
+ "id": "mul_317_248",
21
+ "cosine": 0.9816783666610718
22
+ },
23
+ {
24
+ "id": "mul_126_94",
25
+ "cosine": 0.9842002391815186
26
+ },
27
+ {
28
+ "id": "eq_3_7_25",
29
+ "cosine": 0.9815255403518677
30
+ },
31
+ {
32
+ "id": "eq_5_-4_31",
33
+ "cosine": 0.9807822704315186
34
+ },
35
+ {
36
+ "id": "ages_Ava_Ben_Cara_Dan",
37
+ "cosine": 0.9691709280014038
38
+ },
39
+ {
40
+ "id": "ages_Iris_Jules_Kai_Lena",
41
+ "cosine": 0.9678541421890259
42
+ },
43
+ {
44
+ "id": "coin_6",
45
+ "cosine": 0.9857755303382874
46
+ },
47
+ {
48
+ "id": "coin_7",
49
+ "cosine": 0.984341025352478
50
+ },
51
+ {
52
+ "id": "heads_exact2_3",
53
+ "cosine": 0.9638394713401794
54
+ },
55
+ {
56
+ "id": "heads_exact2_4",
57
+ "cosine": 0.9640707969665527
58
+ }
59
+ ],
60
+ "cosine_to_mean_mean": 0.9748894572257996,
61
+ "cosine_to_mean_range": [
62
+ 0.9632018804550171,
63
+ 0.9857755303382874
64
+ ]
65
+ }
qwen3-soft-switch-fixed-controller/fixed_controller_vector.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:db208aefc092d64e2cad33c4b9870f89e02f7d0854fb37c8e4df3a926859dee6
3
+ size 8025
qwen3-soft-switch-fixed-controller/prompt_splits.json ADDED
@@ -0,0 +1,248 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "fit": [
3
+ {
4
+ "id": "count_strawberries_r",
5
+ "category": "counting",
6
+ "prompt": "How many r letters are in the word strawberries?"
7
+ },
8
+ {
9
+ "id": "count_committee_m",
10
+ "category": "counting",
11
+ "prompt": "How many m letters are in the word committee?"
12
+ },
13
+ {
14
+ "id": "mul_317_248",
15
+ "category": "arithmetic",
16
+ "prompt": "Compute 317 multiplied by 248 and give the exact result."
17
+ },
18
+ {
19
+ "id": "mul_126_94",
20
+ "category": "arithmetic",
21
+ "prompt": "Compute 126 multiplied by 94 and give the exact result."
22
+ },
23
+ {
24
+ "id": "eq_3_7_25",
25
+ "category": "algebra",
26
+ "prompt": "Solve for x: 3x + 7 = 25."
27
+ },
28
+ {
29
+ "id": "eq_5_-4_31",
30
+ "category": "algebra",
31
+ "prompt": "Solve for x: 5x + -4 = 31."
32
+ },
33
+ {
34
+ "id": "ages_Ava_Ben_Cara_Dan",
35
+ "category": "logic_ordering",
36
+ "prompt": "Ava is older than Ben. Ben is older than Cara. Cara is older than Dan. Who is the second youngest?"
37
+ },
38
+ {
39
+ "id": "ages_Iris_Jules_Kai_Lena",
40
+ "category": "logic_ordering",
41
+ "prompt": "Iris is older than Jules. Jules is older than Kai. Kai is older than Lena. Who is the second youngest?"
42
+ },
43
+ {
44
+ "id": "coin_6",
45
+ "category": "dynamic_programming",
46
+ "prompt": "You have coins worth 1, 3, and 4. What is the minimum number of coins needed to make 6?"
47
+ },
48
+ {
49
+ "id": "coin_7",
50
+ "category": "dynamic_programming",
51
+ "prompt": "You have coins worth 1, 3, and 4. What is the minimum number of coins needed to make 7?"
52
+ },
53
+ {
54
+ "id": "heads_exact2_3",
55
+ "category": "probability",
56
+ "prompt": "A fair coin is flipped 3 times. What is the probability of getting exactly two heads?"
57
+ },
58
+ {
59
+ "id": "heads_exact2_4",
60
+ "category": "probability",
61
+ "prompt": "A fair coin is flipped 4 times. What is the probability of getting exactly two heads?"
62
+ }
63
+ ],
64
+ "held_out_same_category": [
65
+ {
66
+ "id": "count_bookkeeper_e",
67
+ "category": "counting",
68
+ "prompt": "How many e letters are in the word bookkeeper?"
69
+ },
70
+ {
71
+ "id": "count_programming_g",
72
+ "category": "counting",
73
+ "prompt": "How many g letters are in the word programming?"
74
+ },
75
+ {
76
+ "id": "mul_124_125",
77
+ "category": "arithmetic",
78
+ "prompt": "Compute 124 multiplied by 125 and give the exact result."
79
+ },
80
+ {
81
+ "id": "mul_808_19",
82
+ "category": "arithmetic",
83
+ "prompt": "Compute 808 multiplied by 19 and give the exact result."
84
+ },
85
+ {
86
+ "id": "eq_4_3_35",
87
+ "category": "algebra",
88
+ "prompt": "Solve for x: 4x + 3 = 35."
89
+ },
90
+ {
91
+ "id": "eq_6_-1_23",
92
+ "category": "algebra",
93
+ "prompt": "Solve for x: 6x + -1 = 23."
94
+ },
95
+ {
96
+ "id": "ages_Uma_Vera_Wade_Xena",
97
+ "category": "logic_ordering",
98
+ "prompt": "Uma is older than Vera. Vera is older than Wade. Wade is older than Xena. Who is the second youngest?"
99
+ },
100
+ {
101
+ "id": "ages_Yara_Zane_Arlo_Bree",
102
+ "category": "logic_ordering",
103
+ "prompt": "Yara is older than Zane. Zane is older than Arlo. Arlo is older than Bree. Who is the second youngest?"
104
+ },
105
+ {
106
+ "id": "coin_10",
107
+ "category": "dynamic_programming",
108
+ "prompt": "You have coins worth 1, 3, and 4. What is the minimum number of coins needed to make 10?"
109
+ },
110
+ {
111
+ "id": "coin_11",
112
+ "category": "dynamic_programming",
113
+ "prompt": "You have coins worth 1, 3, and 4. What is the minimum number of coins needed to make 11?"
114
+ },
115
+ {
116
+ "id": "heads_exact2_7",
117
+ "category": "probability",
118
+ "prompt": "A fair coin is flipped 7 times. What is the probability of getting exactly two heads?"
119
+ },
120
+ {
121
+ "id": "heads_exact2_8",
122
+ "category": "probability",
123
+ "prompt": "A fair coin is flipped 8 times. What is the probability of getting exactly two heads?"
124
+ }
125
+ ],
126
+ "held_out_category_shift": [
127
+ {
128
+ "id": "calendar_monday_march_1",
129
+ "category": "calendar",
130
+ "prompt": "If January 1 is a Monday in a non-leap year, what day of the week is March 1?"
131
+ },
132
+ {
133
+ "id": "calendar_tuesday_april_1",
134
+ "category": "calendar",
135
+ "prompt": "If January 1 is a Tuesday in a non-leap year, what day of the week is April 1?"
136
+ },
137
+ {
138
+ "id": "calendar_wednesday_may_1",
139
+ "category": "calendar",
140
+ "prompt": "If January 1 is a Wednesday in a non-leap year, what day of the week is May 1?"
141
+ },
142
+ {
143
+ "id": "calendar_thursday_june_1",
144
+ "category": "calendar",
145
+ "prompt": "If January 1 is a Thursday in a non-leap year, what day of the week is June 1?"
146
+ },
147
+ {
148
+ "id": "calendar_friday_july_1",
149
+ "category": "calendar",
150
+ "prompt": "If January 1 is a Friday in a non-leap year, what day of the week is July 1?"
151
+ },
152
+ {
153
+ "id": "calendar_saturday_august_1",
154
+ "category": "calendar",
155
+ "prompt": "If January 1 is a Saturday in a non-leap year, what day of the week is August 1?"
156
+ },
157
+ {
158
+ "id": "sort_1",
159
+ "category": "language",
160
+ "prompt": "Sort these words alphabetically: banana, apple, apricot, blueberry."
161
+ },
162
+ {
163
+ "id": "sort_2",
164
+ "category": "language",
165
+ "prompt": "Sort these words alphabetically: zebra, ant, koala, yak."
166
+ },
167
+ {
168
+ "id": "sort_3",
169
+ "category": "language",
170
+ "prompt": "Sort these words alphabetically: violet, amber, cyan, beige."
171
+ },
172
+ {
173
+ "id": "sort_4",
174
+ "category": "language",
175
+ "prompt": "Sort these words alphabetically: delta, alpha, charlie, bravo."
176
+ },
177
+ {
178
+ "id": "sort_5",
179
+ "category": "language",
180
+ "prompt": "Sort these words alphabetically: plum, peach, pear, pineapple."
181
+ },
182
+ {
183
+ "id": "sort_6",
184
+ "category": "language",
185
+ "prompt": "Sort these words alphabetically: maple, cedar, birch, aspen."
186
+ },
187
+ {
188
+ "id": "code_bug_1",
189
+ "category": "coding",
190
+ "prompt": "What is wrong with this Python code?\n\ntotal = sum([1, 2, 3]) / len(1, 2, 3)"
191
+ },
192
+ {
193
+ "id": "code_bug_2",
194
+ "category": "coding",
195
+ "prompt": "What is wrong with this Python code?\n\nvalue = int('7') + '3'"
196
+ },
197
+ {
198
+ "id": "code_bug_3",
199
+ "category": "coding",
200
+ "prompt": "What is wrong with this Python code?\n\nitems.append[4]"
201
+ },
202
+ {
203
+ "id": "code_bug_4",
204
+ "category": "coding",
205
+ "prompt": "What is wrong with this Python code?\n\nfor i in range(3)\n print(i)"
206
+ },
207
+ {
208
+ "id": "code_bug_5",
209
+ "category": "coding",
210
+ "prompt": "What is wrong with this Python code?\n\nif x = 5:\n print(x)"
211
+ },
212
+ {
213
+ "id": "code_bug_6",
214
+ "category": "coding",
215
+ "prompt": "What is wrong with this Python code?\n\nresult = max(4, 9, '11')"
216
+ },
217
+ {
218
+ "id": "trip_bank",
219
+ "category": "planning",
220
+ "prompt": "I need to visit the bank, grocery store, and post office in one trip. The bank closes first, the grocery store closes second, and the post office closes last. What order is safest if I want to avoid closures?"
221
+ },
222
+ {
223
+ "id": "trip_pharmacy",
224
+ "category": "planning",
225
+ "prompt": "I need to visit the pharmacy, hardware store, and bakery in one trip. The pharmacy closes first, the hardware store closes second, and the bakery closes last. What order is safest if I want to avoid closures?"
226
+ },
227
+ {
228
+ "id": "trip_city_hall",
229
+ "category": "planning",
230
+ "prompt": "I need to visit the city hall, library, and cafe in one trip. The city hall closes first, the library closes second, and the cafe closes last. What order is safest if I want to avoid closures?"
231
+ },
232
+ {
233
+ "id": "trip_dentist",
234
+ "category": "planning",
235
+ "prompt": "I need to visit the dentist, laundromat, and market in one trip. The dentist closes first, the laundromat closes second, and the market closes last. What order is safest if I want to avoid closures?"
236
+ },
237
+ {
238
+ "id": "trip_doctor",
239
+ "category": "planning",
240
+ "prompt": "I need to visit the doctor, pet store, and bookstore in one trip. The doctor closes first, the pet store closes second, and the bookstore closes last. What order is safest if I want to avoid closures?"
241
+ },
242
+ {
243
+ "id": "trip_tailor",
244
+ "category": "planning",
245
+ "prompt": "I need to visit the tailor, florist, and museum in one trip. The tailor closes first, the florist closes second, and the museum closes last. What order is safest if I want to avoid closures?"
246
+ }
247
+ ]
248
+ }
qwen3-soft-switch-fixed-controller/run_metadata.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at_utc": "2026-07-28T07:58:03.969587+00:00",
3
+ "script": "/Users/d/coding/reasoning-switch-research-rebuttal/scripts/run_qwen3_soft_switch_fixed_controller_audit.py",
4
+ "script_sha256": "5fb77f9060dffc25ddde323805677d791ad3723f3eb35cab0af5ad0050c39a53",
5
+ "protocol": "/Users/d/coding/reasoning-switch-research-rebuttal/artifacts/rebuttal-20260728/qwen3-soft-switch-fixed-controller-protocol.json",
6
+ "protocol_sha256": "f2de193db33a892caefcf267a47ea5e8baa76a57565a53e6dff7f831991017e9",
7
+ "protocol_commit": "bd81f43519f6197269af1d833467cc28a84b94cc",
8
+ "model_id": "mlx-community/Qwen3-4B-4bit",
9
+ "model_revision": "4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25",
10
+ "platform": "macOS-15.5-arm64-arm-64bit",
11
+ "python": "3.12.13 (main, Mar 10 2026, 18:18:07) [Clang 21.1.4 ]",
12
+ "pid": 32203
13
+ }