claude-autorouter 0.3.7 → 0.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (49) hide show
  1. package/.env.example +6 -3
  2. package/CODE_OF_CONDUCT.md +9 -0
  3. package/CONTRIBUTING.md +57 -0
  4. package/README.md +47 -70
  5. package/SECURITY.md +23 -0
  6. package/SUPPORT.md +18 -0
  7. package/bin/autorouter.mjs +40 -57
  8. package/docs/development.md +48 -2
  9. package/docs/hardware-benchmark.md +29 -0
  10. package/docs/hardware-comparison.md +55 -0
  11. package/docs/hardware-results-16gb.json +4002 -0
  12. package/docs/hardware-results-16gb.md +26 -0
  13. package/docs/hardware-results-64gb.json +4020 -0
  14. package/docs/reference.md +92 -37
  15. package/docs/releasing.md +79 -37
  16. package/docs/router-performance.json +1697 -0
  17. package/docs/router-performance.md +50 -0
  18. package/docs/status-performance.json +363 -0
  19. package/docs/status-performance.md +44 -0
  20. package/docs/subscription-integration.md +27 -0
  21. package/package.json +66 -10
  22. package/src/auto-routing.mjs +184 -24
  23. package/src/bounded-json.mjs +57 -0
  24. package/src/cli-help.mjs +90 -0
  25. package/src/config-command.mjs +158 -0
  26. package/src/config.mjs +53 -28
  27. package/src/contracts.mjs +123 -0
  28. package/src/evaluation-report.mjs +114 -0
  29. package/src/keychain.mjs +58 -0
  30. package/src/local-diagnostic.mjs +191 -0
  31. package/src/model-catalog.mjs +96 -0
  32. package/src/model-request.mjs +6 -7
  33. package/src/ollama-evaluator.mjs +9 -27
  34. package/src/onboarding.mjs +130 -26
  35. package/src/prompt-state.mjs +22 -7
  36. package/src/redaction.mjs +97 -0
  37. package/src/request-validation.mjs +54 -0
  38. package/src/response-observer.mjs +126 -18
  39. package/src/router.mjs +151 -61
  40. package/src/savings.mjs +74 -16
  41. package/src/server.mjs +79 -12
  42. package/src/session-history.mjs +262 -0
  43. package/src/session-log.mjs +9 -58
  44. package/src/status-state.mjs +110 -62
  45. package/src/statusline.mjs +57 -27
  46. package/src/telemetry-event.mjs +200 -0
  47. package/src/token-counter.mjs +3 -1
  48. package/src/turn-state.mjs +132 -0
  49. package/src/user-config.mjs +81 -10
@@ -0,0 +1,4002 @@
1
+ {
2
+ "type": "ollama_routing_evaluation",
3
+ "timestamp": "2026-10-05T09:08:09.288Z",
4
+ "quality_policy": {
5
+ "profile": "compatible",
6
+ "minAgreement": 1,
7
+ "maxUnderRouteRate": 0,
8
+ "requiredTiers": [
9
+ "haiku",
10
+ "sonnet",
11
+ "opus"
12
+ ]
13
+ },
14
+ "hardware": {
15
+ "cpu": "Apple M4",
16
+ "logical_cpus": 10,
17
+ "architecture": "arm64",
18
+ "platform": "darwin",
19
+ "os_release": "25.6.0",
20
+ "node": "v26.10.0",
21
+ "total_memory_bytes": 17179869184,
22
+ "initial_free_memory_bytes": 882540544,
23
+ "initial_load_average": [
24
+ 3.55126953125,
25
+ 3.3505859375,
26
+ 3.47998046875
27
+ ],
28
+ "final_free_memory_bytes": 127926272,
29
+ "final_load_average": [
30
+ 2.2275390625,
31
+ 3.59716796875,
32
+ 3.66943359375
33
+ ]
34
+ },
35
+ "ollama_version": "0.35.0",
36
+ "fixture_sha256": "1ee6ab593e47d87a67e922c0afb0c9db5787d8a27d1a0ff3e24baa0543291fc5",
37
+ "split": "heldout",
38
+ "rounds": 3,
39
+ "stress_rounds": 8,
40
+ "warm_timeout_ms": 30000,
41
+ "cold_timeout_ms": 60000,
42
+ "state_character_budget": 3000,
43
+ "model_confidence": "not used for routing; native entropy confidence is not calibrated accuracy",
44
+ "models": [
45
+ {
46
+ "model": "tev1:4b-q4_K_M",
47
+ "digest": "3509ac7180e86e5fa8efc7b5745e32d55dd9d4e0a86bc9a88aba5323a5d29bc6",
48
+ "download_bytes": 2708817101,
49
+ "protocol": "/v1/systemone",
50
+ "rubric_sha256": "be151cedb4de4b7ef3f7162d751f70ce7d9dd14efc66fae1835f73ffd04027be",
51
+ "context_configuration": "model/server defaults; measured below",
52
+ "cold": {
53
+ "case": "held-h-arithmetic",
54
+ "split": "heldout",
55
+ "expected": "haiku",
56
+ "actual": "haiku",
57
+ "classified_tier": "haiku",
58
+ "source": "ollama",
59
+ "evaluator": "ollama",
60
+ "selected_model": null,
61
+ "confirmed_model": null,
62
+ "wall_ms": 7128.3,
63
+ "state_bytes": 298,
64
+ "metrics": {
65
+ "input_tokens": 667,
66
+ "output_tokens": 1
67
+ }
68
+ },
69
+ "resident_memory": {
70
+ "model_size_bytes": 2912388382,
71
+ "model_vram_bytes": 2912388382,
72
+ "context_length": 2050,
73
+ "parameter_size": "4.2B",
74
+ "quantization": "Q4_K_M",
75
+ "runtime_process_rss_bytes": 3089285120,
76
+ "system_free_bytes": 356401152
77
+ },
78
+ "rows": [
79
+ {
80
+ "round": 1,
81
+ "case": "held-h-link",
82
+ "split": "heldout",
83
+ "expected": "haiku",
84
+ "actual": "haiku",
85
+ "classified_tier": "haiku",
86
+ "source": "ollama",
87
+ "evaluator": "ollama",
88
+ "selected_model": null,
89
+ "confirmed_model": null,
90
+ "wall_ms": 2585.01,
91
+ "state_bytes": 386,
92
+ "metrics": {
93
+ "input_tokens": 673,
94
+ "output_tokens": 1
95
+ }
96
+ },
97
+ {
98
+ "round": 1,
99
+ "case": "held-o-injection",
100
+ "split": "heldout",
101
+ "expected": "opus",
102
+ "actual": "opus",
103
+ "classified_tier": "opus",
104
+ "source": "ollama",
105
+ "evaluator": "ollama",
106
+ "selected_model": null,
107
+ "confirmed_model": null,
108
+ "wall_ms": 2734.33,
109
+ "state_bytes": 694,
110
+ "metrics": {
111
+ "input_tokens": 723,
112
+ "output_tokens": 1
113
+ }
114
+ },
115
+ {
116
+ "round": 1,
117
+ "case": "held-h-marker",
118
+ "split": "heldout",
119
+ "expected": "haiku",
120
+ "actual": "haiku",
121
+ "classified_tier": "haiku",
122
+ "source": "ollama",
123
+ "evaluator": "ollama",
124
+ "selected_model": null,
125
+ "confirmed_model": null,
126
+ "wall_ms": 2458.88,
127
+ "state_bytes": 323,
128
+ "metrics": {
129
+ "input_tokens": 669,
130
+ "output_tokens": 1
131
+ }
132
+ },
133
+ {
134
+ "round": 1,
135
+ "case": "held-tier-influence-mechanical",
136
+ "split": "heldout",
137
+ "expected": "haiku",
138
+ "actual": "haiku",
139
+ "classified_tier": "haiku",
140
+ "source": "ollama",
141
+ "evaluator": "ollama",
142
+ "selected_model": null,
143
+ "confirmed_model": null,
144
+ "wall_ms": 2597.47,
145
+ "state_bytes": 432,
146
+ "metrics": {
147
+ "input_tokens": 695,
148
+ "output_tokens": 1
149
+ }
150
+ },
151
+ {
152
+ "round": 1,
153
+ "case": "held-s-tool-followup",
154
+ "split": "heldout",
155
+ "expected": "sonnet",
156
+ "actual": "sonnet",
157
+ "classified_tier": "sonnet",
158
+ "source": "ollama",
159
+ "evaluator": "ollama",
160
+ "selected_model": null,
161
+ "confirmed_model": null,
162
+ "wall_ms": 2717.86,
163
+ "state_bytes": 508,
164
+ "metrics": {
165
+ "input_tokens": 715,
166
+ "output_tokens": 1
167
+ }
168
+ },
169
+ {
170
+ "round": 1,
171
+ "case": "held-h-reminder",
172
+ "split": "heldout",
173
+ "expected": "haiku",
174
+ "actual": "haiku",
175
+ "classified_tier": "haiku",
176
+ "source": "ollama",
177
+ "evaluator": "ollama",
178
+ "selected_model": null,
179
+ "confirmed_model": null,
180
+ "wall_ms": 2496.95,
181
+ "state_bytes": 294,
182
+ "metrics": {
183
+ "input_tokens": 671,
184
+ "output_tokens": 1
185
+ }
186
+ },
187
+ {
188
+ "round": 1,
189
+ "case": "held-h-injection",
190
+ "split": "heldout",
191
+ "expected": "haiku",
192
+ "actual": "haiku",
193
+ "classified_tier": "haiku",
194
+ "source": "ollama",
195
+ "evaluator": "ollama",
196
+ "selected_model": null,
197
+ "confirmed_model": null,
198
+ "wall_ms": 2601.73,
199
+ "state_bytes": 512,
200
+ "metrics": {
201
+ "input_tokens": 697,
202
+ "output_tokens": 1
203
+ }
204
+ },
205
+ {
206
+ "round": 1,
207
+ "case": "held-o-exactly-once",
208
+ "split": "heldout",
209
+ "expected": "opus",
210
+ "actual": "opus",
211
+ "classified_tier": "opus",
212
+ "source": "ollama",
213
+ "evaluator": "ollama",
214
+ "selected_model": null,
215
+ "confirmed_model": null,
216
+ "wall_ms": 2711.22,
217
+ "state_bytes": 676,
218
+ "metrics": {
219
+ "input_tokens": 715,
220
+ "output_tokens": 1
221
+ }
222
+ },
223
+ {
224
+ "round": 1,
225
+ "case": "held-s-refactor",
226
+ "split": "heldout",
227
+ "expected": "sonnet",
228
+ "actual": "sonnet",
229
+ "classified_tier": "sonnet",
230
+ "source": "ollama",
231
+ "evaluator": "ollama",
232
+ "selected_model": null,
233
+ "confirmed_model": null,
234
+ "wall_ms": 2591.52,
235
+ "state_bytes": 464,
236
+ "metrics": {
237
+ "input_tokens": 673,
238
+ "output_tokens": 1
239
+ }
240
+ },
241
+ {
242
+ "round": 1,
243
+ "case": "held-o-tenant-cache",
244
+ "split": "heldout",
245
+ "expected": "opus",
246
+ "actual": "opus",
247
+ "classified_tier": "opus",
248
+ "source": "ollama",
249
+ "evaluator": "ollama",
250
+ "selected_model": null,
251
+ "confirmed_model": null,
252
+ "wall_ms": 2700.02,
253
+ "state_bytes": 658,
254
+ "metrics": {
255
+ "input_tokens": 707,
256
+ "output_tokens": 1
257
+ }
258
+ },
259
+ {
260
+ "round": 1,
261
+ "case": "held-s-unclear-polish",
262
+ "split": "heldout",
263
+ "expected": "sonnet",
264
+ "actual": "sonnet",
265
+ "classified_tier": "sonnet",
266
+ "source": "ollama",
267
+ "evaluator": "ollama",
268
+ "selected_model": null,
269
+ "confirmed_model": null,
270
+ "wall_ms": 2608.66,
271
+ "state_bytes": 478,
272
+ "metrics": {
273
+ "input_tokens": 679,
274
+ "output_tokens": 1
275
+ }
276
+ },
277
+ {
278
+ "round": 1,
279
+ "case": "held-s-cli",
280
+ "split": "heldout",
281
+ "expected": "sonnet",
282
+ "actual": "sonnet",
283
+ "classified_tier": "sonnet",
284
+ "source": "ollama",
285
+ "evaluator": "ollama",
286
+ "selected_model": null,
287
+ "confirmed_model": null,
288
+ "wall_ms": 2624.05,
289
+ "state_bytes": 480,
290
+ "metrics": {
291
+ "input_tokens": 687,
292
+ "output_tokens": 1
293
+ }
294
+ },
295
+ {
296
+ "round": 1,
297
+ "case": "held-s-validation",
298
+ "split": "heldout",
299
+ "expected": "sonnet",
300
+ "actual": "sonnet",
301
+ "classified_tier": "sonnet",
302
+ "source": "ollama",
303
+ "evaluator": "ollama",
304
+ "selected_model": null,
305
+ "confirmed_model": null,
306
+ "wall_ms": 2610.95,
307
+ "state_bytes": 484,
308
+ "metrics": {
309
+ "input_tokens": 677,
310
+ "output_tokens": 1
311
+ }
312
+ },
313
+ {
314
+ "round": 1,
315
+ "case": "held-s-query",
316
+ "split": "heldout",
317
+ "expected": "sonnet",
318
+ "actual": "sonnet",
319
+ "classified_tier": "sonnet",
320
+ "source": "ollama",
321
+ "evaluator": "ollama",
322
+ "selected_model": null,
323
+ "confirmed_model": null,
324
+ "wall_ms": 2671.49,
325
+ "state_bytes": 522,
326
+ "metrics": {
327
+ "input_tokens": 687,
328
+ "output_tokens": 1
329
+ }
330
+ },
331
+ {
332
+ "round": 1,
333
+ "case": "held-o-crypto",
334
+ "split": "heldout",
335
+ "expected": "opus",
336
+ "actual": "opus",
337
+ "classified_tier": "opus",
338
+ "source": "ollama",
339
+ "evaluator": "ollama",
340
+ "selected_model": null,
341
+ "confirmed_model": null,
342
+ "wall_ms": 2710.45,
343
+ "state_bytes": 654,
344
+ "metrics": {
345
+ "input_tokens": 703,
346
+ "output_tokens": 1
347
+ }
348
+ },
349
+ {
350
+ "round": 1,
351
+ "case": "held-french-race",
352
+ "split": "heldout",
353
+ "expected": "opus",
354
+ "actual": "opus",
355
+ "classified_tier": "opus",
356
+ "source": "ollama",
357
+ "evaluator": "ollama",
358
+ "selected_model": null,
359
+ "confirmed_model": null,
360
+ "wall_ms": 2744.97,
361
+ "state_bytes": 510,
362
+ "metrics": {
363
+ "input_tokens": 701,
364
+ "output_tokens": 1
365
+ }
366
+ },
367
+ {
368
+ "round": 1,
369
+ "case": "held-spanish-feature",
370
+ "split": "heldout",
371
+ "expected": "sonnet",
372
+ "actual": "sonnet",
373
+ "classified_tier": "sonnet",
374
+ "source": "ollama",
375
+ "evaluator": "ollama",
376
+ "selected_model": null,
377
+ "confirmed_model": null,
378
+ "wall_ms": 2771.66,
379
+ "state_bytes": 380,
380
+ "metrics": {
381
+ "input_tokens": 673,
382
+ "output_tokens": 1
383
+ }
384
+ },
385
+ {
386
+ "round": 1,
387
+ "case": "held-h-translation",
388
+ "split": "heldout",
389
+ "expected": "haiku",
390
+ "actual": "haiku",
391
+ "classified_tier": "haiku",
392
+ "source": "ollama",
393
+ "evaluator": "ollama",
394
+ "selected_model": null,
395
+ "confirmed_model": null,
396
+ "wall_ms": 2673.68,
397
+ "state_bytes": 290,
398
+ "metrics": {
399
+ "input_tokens": 649,
400
+ "output_tokens": 1
401
+ }
402
+ },
403
+ {
404
+ "round": 1,
405
+ "case": "held-h-summary",
406
+ "split": "heldout",
407
+ "expected": "haiku",
408
+ "actual": "haiku",
409
+ "classified_tier": "haiku",
410
+ "source": "ollama",
411
+ "evaluator": "ollama",
412
+ "selected_model": null,
413
+ "confirmed_model": null,
414
+ "wall_ms": 2788.07,
415
+ "state_bytes": 406,
416
+ "metrics": {
417
+ "input_tokens": 677,
418
+ "output_tokens": 1
419
+ }
420
+ },
421
+ {
422
+ "round": 1,
423
+ "case": "held-o-compiler",
424
+ "split": "heldout",
425
+ "expected": "opus",
426
+ "actual": "sonnet",
427
+ "classified_tier": "sonnet",
428
+ "source": "ollama",
429
+ "evaluator": "ollama",
430
+ "selected_model": null,
431
+ "confirmed_model": null,
432
+ "wall_ms": 2828.86,
433
+ "state_bytes": 654,
434
+ "metrics": {
435
+ "input_tokens": 699,
436
+ "output_tokens": 1
437
+ }
438
+ },
439
+ {
440
+ "round": 1,
441
+ "case": "held-tier-influence-difficult",
442
+ "split": "heldout",
443
+ "expected": "opus",
444
+ "actual": "opus",
445
+ "classified_tier": "opus",
446
+ "source": "ollama",
447
+ "evaluator": "ollama",
448
+ "selected_model": null,
449
+ "confirmed_model": null,
450
+ "wall_ms": 2953.57,
451
+ "state_bytes": 640,
452
+ "metrics": {
453
+ "input_tokens": 715,
454
+ "output_tokens": 1
455
+ }
456
+ },
457
+ {
458
+ "round": 1,
459
+ "case": "held-o-rolling-format",
460
+ "split": "heldout",
461
+ "expected": "opus",
462
+ "actual": "opus",
463
+ "classified_tier": "opus",
464
+ "source": "ollama",
465
+ "evaluator": "ollama",
466
+ "selected_model": null,
467
+ "confirmed_model": null,
468
+ "wall_ms": 2845.86,
469
+ "state_bytes": 562,
470
+ "metrics": {
471
+ "input_tokens": 697,
472
+ "output_tokens": 1
473
+ }
474
+ },
475
+ {
476
+ "round": 1,
477
+ "case": "held-h-arithmetic",
478
+ "split": "heldout",
479
+ "expected": "haiku",
480
+ "actual": "haiku",
481
+ "classified_tier": "haiku",
482
+ "source": "ollama",
483
+ "evaluator": "ollama",
484
+ "selected_model": null,
485
+ "confirmed_model": null,
486
+ "wall_ms": 216.34,
487
+ "state_bytes": 298,
488
+ "metrics": {
489
+ "input_tokens": 667,
490
+ "output_tokens": 1
491
+ }
492
+ },
493
+ {
494
+ "round": 1,
495
+ "case": "held-h-literal-opus",
496
+ "split": "heldout",
497
+ "expected": "haiku",
498
+ "actual": "opus",
499
+ "classified_tier": "opus",
500
+ "source": "ollama",
501
+ "evaluator": "ollama",
502
+ "selected_model": null,
503
+ "confirmed_model": null,
504
+ "wall_ms": 2675.2,
505
+ "state_bytes": 312,
506
+ "metrics": {
507
+ "input_tokens": 655,
508
+ "output_tokens": 1
509
+ }
510
+ },
511
+ {
512
+ "round": 1,
513
+ "case": "held-o-snapshot",
514
+ "split": "heldout",
515
+ "expected": "opus",
516
+ "actual": "opus",
517
+ "classified_tier": "opus",
518
+ "source": "ollama",
519
+ "evaluator": "ollama",
520
+ "selected_model": null,
521
+ "confirmed_model": null,
522
+ "wall_ms": 2929.63,
523
+ "state_bytes": 764,
524
+ "metrics": {
525
+ "input_tokens": 725,
526
+ "output_tokens": 1
527
+ }
528
+ },
529
+ {
530
+ "round": 1,
531
+ "case": "held-italian-literal",
532
+ "split": "heldout",
533
+ "expected": "haiku",
534
+ "actual": "haiku",
535
+ "classified_tier": "haiku",
536
+ "source": "ollama",
537
+ "evaluator": "ollama",
538
+ "selected_model": null,
539
+ "confirmed_model": null,
540
+ "wall_ms": 2731.49,
541
+ "state_bytes": 252,
542
+ "metrics": {
543
+ "input_tokens": 645,
544
+ "output_tokens": 1
545
+ }
546
+ },
547
+ {
548
+ "round": 1,
549
+ "case": "held-short-difficult-followup",
550
+ "split": "heldout",
551
+ "expected": "opus",
552
+ "actual": "opus",
553
+ "classified_tier": "opus",
554
+ "source": "ollama",
555
+ "evaluator": "ollama",
556
+ "selected_model": null,
557
+ "confirmed_model": null,
558
+ "wall_ms": 2878.73,
559
+ "state_bytes": 483,
560
+ "metrics": {
561
+ "input_tokens": 689,
562
+ "output_tokens": 1
563
+ }
564
+ },
565
+ {
566
+ "round": 1,
567
+ "case": "held-o-fencing-followup",
568
+ "split": "heldout",
569
+ "expected": "opus",
570
+ "actual": "opus",
571
+ "classified_tier": "opus",
572
+ "source": "ollama",
573
+ "evaluator": "ollama",
574
+ "selected_model": null,
575
+ "confirmed_model": null,
576
+ "wall_ms": 3185.31,
577
+ "state_bytes": 780,
578
+ "metrics": {
579
+ "input_tokens": 753,
580
+ "output_tokens": 1
581
+ }
582
+ },
583
+ {
584
+ "round": 1,
585
+ "case": "held-s-null-bug",
586
+ "split": "heldout",
587
+ "expected": "sonnet",
588
+ "actual": "sonnet",
589
+ "classified_tier": "sonnet",
590
+ "source": "ollama",
591
+ "evaluator": "ollama",
592
+ "selected_model": null,
593
+ "confirmed_model": null,
594
+ "wall_ms": 2986.15,
595
+ "state_bytes": 546,
596
+ "metrics": {
597
+ "input_tokens": 691,
598
+ "output_tokens": 1
599
+ }
600
+ },
601
+ {
602
+ "round": 1,
603
+ "case": "held-s-pagination",
604
+ "split": "heldout",
605
+ "expected": "sonnet",
606
+ "actual": "sonnet",
607
+ "classified_tier": "sonnet",
608
+ "source": "ollama",
609
+ "evaluator": "ollama",
610
+ "selected_model": null,
611
+ "confirmed_model": null,
612
+ "wall_ms": 3112.41,
613
+ "state_bytes": 482,
614
+ "metrics": {
615
+ "input_tokens": 679,
616
+ "output_tokens": 1
617
+ }
618
+ },
619
+ {
620
+ "round": 2,
621
+ "case": "held-s-null-bug",
622
+ "split": "heldout",
623
+ "expected": "sonnet",
624
+ "actual": "sonnet",
625
+ "classified_tier": "sonnet",
626
+ "source": "ollama",
627
+ "evaluator": "ollama",
628
+ "selected_model": null,
629
+ "confirmed_model": null,
630
+ "wall_ms": 185.96,
631
+ "state_bytes": 546,
632
+ "metrics": {
633
+ "input_tokens": 691,
634
+ "output_tokens": 1
635
+ }
636
+ },
637
+ {
638
+ "round": 2,
639
+ "case": "held-o-injection",
640
+ "split": "heldout",
641
+ "expected": "opus",
642
+ "actual": "opus",
643
+ "classified_tier": "opus",
644
+ "source": "ollama",
645
+ "evaluator": "ollama",
646
+ "selected_model": null,
647
+ "confirmed_model": null,
648
+ "wall_ms": 951.65,
649
+ "state_bytes": 694,
650
+ "metrics": {
651
+ "input_tokens": 723,
652
+ "output_tokens": 1
653
+ }
654
+ },
655
+ {
656
+ "round": 2,
657
+ "case": "held-o-compiler",
658
+ "split": "heldout",
659
+ "expected": "opus",
660
+ "actual": "sonnet",
661
+ "classified_tier": "sonnet",
662
+ "source": "ollama",
663
+ "evaluator": "ollama",
664
+ "selected_model": null,
665
+ "confirmed_model": null,
666
+ "wall_ms": 424.95,
667
+ "state_bytes": 654,
668
+ "metrics": {
669
+ "input_tokens": 699,
670
+ "output_tokens": 1
671
+ }
672
+ },
673
+ {
674
+ "round": 2,
675
+ "case": "held-h-injection",
676
+ "split": "heldout",
677
+ "expected": "haiku",
678
+ "actual": "haiku",
679
+ "classified_tier": "haiku",
680
+ "source": "ollama",
681
+ "evaluator": "ollama",
682
+ "selected_model": null,
683
+ "confirmed_model": null,
684
+ "wall_ms": 391.31,
685
+ "state_bytes": 512,
686
+ "metrics": {
687
+ "input_tokens": 697,
688
+ "output_tokens": 1
689
+ }
690
+ },
691
+ {
692
+ "round": 2,
693
+ "case": "held-o-rolling-format",
694
+ "split": "heldout",
695
+ "expected": "opus",
696
+ "actual": "opus",
697
+ "classified_tier": "opus",
698
+ "source": "ollama",
699
+ "evaluator": "ollama",
700
+ "selected_model": null,
701
+ "confirmed_model": null,
702
+ "wall_ms": 399.91,
703
+ "state_bytes": 562,
704
+ "metrics": {
705
+ "input_tokens": 697,
706
+ "output_tokens": 1
707
+ }
708
+ },
709
+ {
710
+ "round": 2,
711
+ "case": "held-s-pagination",
712
+ "split": "heldout",
713
+ "expected": "sonnet",
714
+ "actual": "sonnet",
715
+ "classified_tier": "sonnet",
716
+ "source": "ollama",
717
+ "evaluator": "ollama",
718
+ "selected_model": null,
719
+ "confirmed_model": null,
720
+ "wall_ms": 166.2,
721
+ "state_bytes": 482,
722
+ "metrics": {
723
+ "input_tokens": 679,
724
+ "output_tokens": 1
725
+ }
726
+ },
727
+ {
728
+ "round": 2,
729
+ "case": "held-s-refactor",
730
+ "split": "heldout",
731
+ "expected": "sonnet",
732
+ "actual": "sonnet",
733
+ "classified_tier": "sonnet",
734
+ "source": "ollama",
735
+ "evaluator": "ollama",
736
+ "selected_model": null,
737
+ "confirmed_model": null,
738
+ "wall_ms": 213.08,
739
+ "state_bytes": 464,
740
+ "metrics": {
741
+ "input_tokens": 673,
742
+ "output_tokens": 1
743
+ }
744
+ },
745
+ {
746
+ "round": 2,
747
+ "case": "held-o-fencing-followup",
748
+ "split": "heldout",
749
+ "expected": "opus",
750
+ "actual": "opus",
751
+ "classified_tier": "opus",
752
+ "source": "ollama",
753
+ "evaluator": "ollama",
754
+ "selected_model": null,
755
+ "confirmed_model": null,
756
+ "wall_ms": 427.62,
757
+ "state_bytes": 780,
758
+ "metrics": {
759
+ "input_tokens": 753,
760
+ "output_tokens": 1
761
+ }
762
+ },
763
+ {
764
+ "round": 2,
765
+ "case": "held-s-tool-followup",
766
+ "split": "heldout",
767
+ "expected": "sonnet",
768
+ "actual": "sonnet",
769
+ "classified_tier": "sonnet",
770
+ "source": "ollama",
771
+ "evaluator": "ollama",
772
+ "selected_model": null,
773
+ "confirmed_model": null,
774
+ "wall_ms": 383.81,
775
+ "state_bytes": 508,
776
+ "metrics": {
777
+ "input_tokens": 715,
778
+ "output_tokens": 1
779
+ }
780
+ },
781
+ {
782
+ "round": 2,
783
+ "case": "held-spanish-feature",
784
+ "split": "heldout",
785
+ "expected": "sonnet",
786
+ "actual": "sonnet",
787
+ "classified_tier": "sonnet",
788
+ "source": "ollama",
789
+ "evaluator": "ollama",
790
+ "selected_model": null,
791
+ "confirmed_model": null,
792
+ "wall_ms": 350.64,
793
+ "state_bytes": 380,
794
+ "metrics": {
795
+ "input_tokens": 673,
796
+ "output_tokens": 1
797
+ }
798
+ },
799
+ {
800
+ "round": 2,
801
+ "case": "held-h-arithmetic",
802
+ "split": "heldout",
803
+ "expected": "haiku",
804
+ "actual": "haiku",
805
+ "classified_tier": "haiku",
806
+ "source": "ollama",
807
+ "evaluator": "ollama",
808
+ "selected_model": null,
809
+ "confirmed_model": null,
810
+ "wall_ms": 165.83,
811
+ "state_bytes": 298,
812
+ "metrics": {
813
+ "input_tokens": 667,
814
+ "output_tokens": 1
815
+ }
816
+ },
817
+ {
818
+ "round": 2,
819
+ "case": "held-o-tenant-cache",
820
+ "split": "heldout",
821
+ "expected": "opus",
822
+ "actual": "opus",
823
+ "classified_tier": "opus",
824
+ "source": "ollama",
825
+ "evaluator": "ollama",
826
+ "selected_model": null,
827
+ "confirmed_model": null,
828
+ "wall_ms": 167.37,
829
+ "state_bytes": 658,
830
+ "metrics": {
831
+ "input_tokens": 707,
832
+ "output_tokens": 1
833
+ }
834
+ },
835
+ {
836
+ "round": 2,
837
+ "case": "held-h-literal-opus",
838
+ "split": "heldout",
839
+ "expected": "haiku",
840
+ "actual": "opus",
841
+ "classified_tier": "opus",
842
+ "source": "ollama",
843
+ "evaluator": "ollama",
844
+ "selected_model": null,
845
+ "confirmed_model": null,
846
+ "wall_ms": 170.02,
847
+ "state_bytes": 312,
848
+ "metrics": {
849
+ "input_tokens": 655,
850
+ "output_tokens": 1
851
+ }
852
+ },
853
+ {
854
+ "round": 2,
855
+ "case": "held-h-marker",
856
+ "split": "heldout",
857
+ "expected": "haiku",
858
+ "actual": "haiku",
859
+ "classified_tier": "haiku",
860
+ "source": "ollama",
861
+ "evaluator": "ollama",
862
+ "selected_model": null,
863
+ "confirmed_model": null,
864
+ "wall_ms": 793.59,
865
+ "state_bytes": 323,
866
+ "metrics": {
867
+ "input_tokens": 669,
868
+ "output_tokens": 1
869
+ }
870
+ },
871
+ {
872
+ "round": 2,
873
+ "case": "held-tier-influence-mechanical",
874
+ "split": "heldout",
875
+ "expected": "haiku",
876
+ "actual": "haiku",
877
+ "classified_tier": "haiku",
878
+ "source": "ollama",
879
+ "evaluator": "ollama",
880
+ "selected_model": null,
881
+ "confirmed_model": null,
882
+ "wall_ms": 641.22,
883
+ "state_bytes": 432,
884
+ "metrics": {
885
+ "input_tokens": 695,
886
+ "output_tokens": 1
887
+ }
888
+ },
889
+ {
890
+ "round": 2,
891
+ "case": "held-s-unclear-polish",
892
+ "split": "heldout",
893
+ "expected": "sonnet",
894
+ "actual": "sonnet",
895
+ "classified_tier": "sonnet",
896
+ "source": "ollama",
897
+ "evaluator": "ollama",
898
+ "selected_model": null,
899
+ "confirmed_model": null,
900
+ "wall_ms": 384.48,
901
+ "state_bytes": 478,
902
+ "metrics": {
903
+ "input_tokens": 679,
904
+ "output_tokens": 1
905
+ }
906
+ },
907
+ {
908
+ "round": 2,
909
+ "case": "held-o-snapshot",
910
+ "split": "heldout",
911
+ "expected": "opus",
912
+ "actual": "opus",
913
+ "classified_tier": "opus",
914
+ "source": "ollama",
915
+ "evaluator": "ollama",
916
+ "selected_model": null,
917
+ "confirmed_model": null,
918
+ "wall_ms": 192.15,
919
+ "state_bytes": 764,
920
+ "metrics": {
921
+ "input_tokens": 725,
922
+ "output_tokens": 1
923
+ }
924
+ },
925
+ {
926
+ "round": 2,
927
+ "case": "held-s-cli",
928
+ "split": "heldout",
929
+ "expected": "sonnet",
930
+ "actual": "sonnet",
931
+ "classified_tier": "sonnet",
932
+ "source": "ollama",
933
+ "evaluator": "ollama",
934
+ "selected_model": null,
935
+ "confirmed_model": null,
936
+ "wall_ms": 189.09,
937
+ "state_bytes": 480,
938
+ "metrics": {
939
+ "input_tokens": 687,
940
+ "output_tokens": 1
941
+ }
942
+ },
943
+ {
944
+ "round": 2,
945
+ "case": "held-h-summary",
946
+ "split": "heldout",
947
+ "expected": "haiku",
948
+ "actual": "haiku",
949
+ "classified_tier": "haiku",
950
+ "source": "ollama",
951
+ "evaluator": "ollama",
952
+ "selected_model": null,
953
+ "confirmed_model": null,
954
+ "wall_ms": 221.78,
955
+ "state_bytes": 406,
956
+ "metrics": {
957
+ "input_tokens": 677,
958
+ "output_tokens": 1
959
+ }
960
+ },
961
+ {
962
+ "round": 2,
963
+ "case": "held-short-difficult-followup",
964
+ "split": "heldout",
965
+ "expected": "opus",
966
+ "actual": "opus",
967
+ "classified_tier": "opus",
968
+ "source": "ollama",
969
+ "evaluator": "ollama",
970
+ "selected_model": null,
971
+ "confirmed_model": null,
972
+ "wall_ms": 173.28,
973
+ "state_bytes": 483,
974
+ "metrics": {
975
+ "input_tokens": 689,
976
+ "output_tokens": 1
977
+ }
978
+ },
979
+ {
980
+ "round": 2,
981
+ "case": "held-italian-literal",
982
+ "split": "heldout",
983
+ "expected": "haiku",
984
+ "actual": "haiku",
985
+ "classified_tier": "haiku",
986
+ "source": "ollama",
987
+ "evaluator": "ollama",
988
+ "selected_model": null,
989
+ "confirmed_model": null,
990
+ "wall_ms": 183.03,
991
+ "state_bytes": 252,
992
+ "metrics": {
993
+ "input_tokens": 645,
994
+ "output_tokens": 1
995
+ }
996
+ },
997
+ {
998
+ "round": 2,
999
+ "case": "held-s-validation",
1000
+ "split": "heldout",
1001
+ "expected": "sonnet",
1002
+ "actual": "sonnet",
1003
+ "classified_tier": "sonnet",
1004
+ "source": "ollama",
1005
+ "evaluator": "ollama",
1006
+ "selected_model": null,
1007
+ "confirmed_model": null,
1008
+ "wall_ms": 166.28,
1009
+ "state_bytes": 484,
1010
+ "metrics": {
1011
+ "input_tokens": 677,
1012
+ "output_tokens": 1
1013
+ }
1014
+ },
1015
+ {
1016
+ "round": 2,
1017
+ "case": "held-o-crypto",
1018
+ "split": "heldout",
1019
+ "expected": "opus",
1020
+ "actual": "opus",
1021
+ "classified_tier": "opus",
1022
+ "source": "ollama",
1023
+ "evaluator": "ollama",
1024
+ "selected_model": null,
1025
+ "confirmed_model": null,
1026
+ "wall_ms": 191.28,
1027
+ "state_bytes": 654,
1028
+ "metrics": {
1029
+ "input_tokens": 703,
1030
+ "output_tokens": 1
1031
+ }
1032
+ },
1033
+ {
1034
+ "round": 2,
1035
+ "case": "held-h-reminder",
1036
+ "split": "heldout",
1037
+ "expected": "haiku",
1038
+ "actual": "haiku",
1039
+ "classified_tier": "haiku",
1040
+ "source": "ollama",
1041
+ "evaluator": "ollama",
1042
+ "selected_model": null,
1043
+ "confirmed_model": null,
1044
+ "wall_ms": 411.76,
1045
+ "state_bytes": 294,
1046
+ "metrics": {
1047
+ "input_tokens": 671,
1048
+ "output_tokens": 1
1049
+ }
1050
+ },
1051
+ {
1052
+ "round": 2,
1053
+ "case": "held-french-race",
1054
+ "split": "heldout",
1055
+ "expected": "opus",
1056
+ "actual": "opus",
1057
+ "classified_tier": "opus",
1058
+ "source": "ollama",
1059
+ "evaluator": "ollama",
1060
+ "selected_model": null,
1061
+ "confirmed_model": null,
1062
+ "wall_ms": 414.19,
1063
+ "state_bytes": 510,
1064
+ "metrics": {
1065
+ "input_tokens": 701,
1066
+ "output_tokens": 1
1067
+ }
1068
+ },
1069
+ {
1070
+ "round": 2,
1071
+ "case": "held-h-link",
1072
+ "split": "heldout",
1073
+ "expected": "haiku",
1074
+ "actual": "haiku",
1075
+ "classified_tier": "haiku",
1076
+ "source": "ollama",
1077
+ "evaluator": "ollama",
1078
+ "selected_model": null,
1079
+ "confirmed_model": null,
1080
+ "wall_ms": 940.01,
1081
+ "state_bytes": 386,
1082
+ "metrics": {
1083
+ "input_tokens": 673,
1084
+ "output_tokens": 1
1085
+ }
1086
+ },
1087
+ {
1088
+ "round": 2,
1089
+ "case": "held-h-translation",
1090
+ "split": "heldout",
1091
+ "expected": "haiku",
1092
+ "actual": "haiku",
1093
+ "classified_tier": "haiku",
1094
+ "source": "ollama",
1095
+ "evaluator": "ollama",
1096
+ "selected_model": null,
1097
+ "confirmed_model": null,
1098
+ "wall_ms": 965.96,
1099
+ "state_bytes": 290,
1100
+ "metrics": {
1101
+ "input_tokens": 649,
1102
+ "output_tokens": 1
1103
+ }
1104
+ },
1105
+ {
1106
+ "round": 2,
1107
+ "case": "held-o-exactly-once",
1108
+ "split": "heldout",
1109
+ "expected": "opus",
1110
+ "actual": "opus",
1111
+ "classified_tier": "opus",
1112
+ "source": "ollama",
1113
+ "evaluator": "ollama",
1114
+ "selected_model": null,
1115
+ "confirmed_model": null,
1116
+ "wall_ms": 1101.71,
1117
+ "state_bytes": 676,
1118
+ "metrics": {
1119
+ "input_tokens": 715,
1120
+ "output_tokens": 1
1121
+ }
1122
+ },
1123
+ {
1124
+ "round": 2,
1125
+ "case": "held-s-query",
1126
+ "split": "heldout",
1127
+ "expected": "sonnet",
1128
+ "actual": "sonnet",
1129
+ "classified_tier": "sonnet",
1130
+ "source": "ollama",
1131
+ "evaluator": "ollama",
1132
+ "selected_model": null,
1133
+ "confirmed_model": null,
1134
+ "wall_ms": 594.82,
1135
+ "state_bytes": 522,
1136
+ "metrics": {
1137
+ "input_tokens": 687,
1138
+ "output_tokens": 1
1139
+ }
1140
+ },
1141
+ {
1142
+ "round": 2,
1143
+ "case": "held-tier-influence-difficult",
1144
+ "split": "heldout",
1145
+ "expected": "opus",
1146
+ "actual": "opus",
1147
+ "classified_tier": "opus",
1148
+ "source": "ollama",
1149
+ "evaluator": "ollama",
1150
+ "selected_model": null,
1151
+ "confirmed_model": null,
1152
+ "wall_ms": 171.59,
1153
+ "state_bytes": 640,
1154
+ "metrics": {
1155
+ "input_tokens": 715,
1156
+ "output_tokens": 1
1157
+ }
1158
+ },
1159
+ {
1160
+ "round": 3,
1161
+ "case": "held-h-link",
1162
+ "split": "heldout",
1163
+ "expected": "haiku",
1164
+ "actual": "haiku",
1165
+ "classified_tier": "haiku",
1166
+ "source": "ollama",
1167
+ "evaluator": "ollama",
1168
+ "selected_model": null,
1169
+ "confirmed_model": null,
1170
+ "wall_ms": 147.62,
1171
+ "state_bytes": 386,
1172
+ "metrics": {
1173
+ "input_tokens": 673,
1174
+ "output_tokens": 1
1175
+ }
1176
+ },
1177
+ {
1178
+ "round": 3,
1179
+ "case": "held-s-tool-followup",
1180
+ "split": "heldout",
1181
+ "expected": "sonnet",
1182
+ "actual": "sonnet",
1183
+ "classified_tier": "sonnet",
1184
+ "source": "ollama",
1185
+ "evaluator": "ollama",
1186
+ "selected_model": null,
1187
+ "confirmed_model": null,
1188
+ "wall_ms": 156.03,
1189
+ "state_bytes": 508,
1190
+ "metrics": {
1191
+ "input_tokens": 715,
1192
+ "output_tokens": 1
1193
+ }
1194
+ },
1195
+ {
1196
+ "round": 3,
1197
+ "case": "held-s-null-bug",
1198
+ "split": "heldout",
1199
+ "expected": "sonnet",
1200
+ "actual": "sonnet",
1201
+ "classified_tier": "sonnet",
1202
+ "source": "ollama",
1203
+ "evaluator": "ollama",
1204
+ "selected_model": null,
1205
+ "confirmed_model": null,
1206
+ "wall_ms": 168.44,
1207
+ "state_bytes": 546,
1208
+ "metrics": {
1209
+ "input_tokens": 691,
1210
+ "output_tokens": 1
1211
+ }
1212
+ },
1213
+ {
1214
+ "round": 3,
1215
+ "case": "held-h-injection",
1216
+ "split": "heldout",
1217
+ "expected": "haiku",
1218
+ "actual": "haiku",
1219
+ "classified_tier": "haiku",
1220
+ "source": "ollama",
1221
+ "evaluator": "ollama",
1222
+ "selected_model": null,
1223
+ "confirmed_model": null,
1224
+ "wall_ms": 166.51,
1225
+ "state_bytes": 512,
1226
+ "metrics": {
1227
+ "input_tokens": 697,
1228
+ "output_tokens": 1
1229
+ }
1230
+ },
1231
+ {
1232
+ "round": 3,
1233
+ "case": "held-o-compiler",
1234
+ "split": "heldout",
1235
+ "expected": "opus",
1236
+ "actual": "sonnet",
1237
+ "classified_tier": "sonnet",
1238
+ "source": "ollama",
1239
+ "evaluator": "ollama",
1240
+ "selected_model": null,
1241
+ "confirmed_model": null,
1242
+ "wall_ms": 167.25,
1243
+ "state_bytes": 654,
1244
+ "metrics": {
1245
+ "input_tokens": 699,
1246
+ "output_tokens": 1
1247
+ }
1248
+ },
1249
+ {
1250
+ "round": 3,
1251
+ "case": "held-italian-literal",
1252
+ "split": "heldout",
1253
+ "expected": "haiku",
1254
+ "actual": "haiku",
1255
+ "classified_tier": "haiku",
1256
+ "source": "ollama",
1257
+ "evaluator": "ollama",
1258
+ "selected_model": null,
1259
+ "confirmed_model": null,
1260
+ "wall_ms": 163.69,
1261
+ "state_bytes": 252,
1262
+ "metrics": {
1263
+ "input_tokens": 645,
1264
+ "output_tokens": 1
1265
+ }
1266
+ },
1267
+ {
1268
+ "round": 3,
1269
+ "case": "held-h-reminder",
1270
+ "split": "heldout",
1271
+ "expected": "haiku",
1272
+ "actual": "haiku",
1273
+ "classified_tier": "haiku",
1274
+ "source": "ollama",
1275
+ "evaluator": "ollama",
1276
+ "selected_model": null,
1277
+ "confirmed_model": null,
1278
+ "wall_ms": 162.97,
1279
+ "state_bytes": 294,
1280
+ "metrics": {
1281
+ "input_tokens": 671,
1282
+ "output_tokens": 1
1283
+ }
1284
+ },
1285
+ {
1286
+ "round": 3,
1287
+ "case": "held-s-pagination",
1288
+ "split": "heldout",
1289
+ "expected": "sonnet",
1290
+ "actual": "sonnet",
1291
+ "classified_tier": "sonnet",
1292
+ "source": "ollama",
1293
+ "evaluator": "ollama",
1294
+ "selected_model": null,
1295
+ "confirmed_model": null,
1296
+ "wall_ms": 185.91,
1297
+ "state_bytes": 482,
1298
+ "metrics": {
1299
+ "input_tokens": 679,
1300
+ "output_tokens": 1
1301
+ }
1302
+ },
1303
+ {
1304
+ "round": 3,
1305
+ "case": "held-s-validation",
1306
+ "split": "heldout",
1307
+ "expected": "sonnet",
1308
+ "actual": "sonnet",
1309
+ "classified_tier": "sonnet",
1310
+ "source": "ollama",
1311
+ "evaluator": "ollama",
1312
+ "selected_model": null,
1313
+ "confirmed_model": null,
1314
+ "wall_ms": 172.31,
1315
+ "state_bytes": 484,
1316
+ "metrics": {
1317
+ "input_tokens": 677,
1318
+ "output_tokens": 1
1319
+ }
1320
+ },
1321
+ {
1322
+ "round": 3,
1323
+ "case": "held-o-tenant-cache",
1324
+ "split": "heldout",
1325
+ "expected": "opus",
1326
+ "actual": "opus",
1327
+ "classified_tier": "opus",
1328
+ "source": "ollama",
1329
+ "evaluator": "ollama",
1330
+ "selected_model": null,
1331
+ "confirmed_model": null,
1332
+ "wall_ms": 182.33,
1333
+ "state_bytes": 658,
1334
+ "metrics": {
1335
+ "input_tokens": 707,
1336
+ "output_tokens": 1
1337
+ }
1338
+ },
1339
+ {
1340
+ "round": 3,
1341
+ "case": "held-spanish-feature",
1342
+ "split": "heldout",
1343
+ "expected": "sonnet",
1344
+ "actual": "sonnet",
1345
+ "classified_tier": "sonnet",
1346
+ "source": "ollama",
1347
+ "evaluator": "ollama",
1348
+ "selected_model": null,
1349
+ "confirmed_model": null,
1350
+ "wall_ms": 175.6,
1351
+ "state_bytes": 380,
1352
+ "metrics": {
1353
+ "input_tokens": 673,
1354
+ "output_tokens": 1
1355
+ }
1356
+ },
1357
+ {
1358
+ "round": 3,
1359
+ "case": "held-s-unclear-polish",
1360
+ "split": "heldout",
1361
+ "expected": "sonnet",
1362
+ "actual": "sonnet",
1363
+ "classified_tier": "sonnet",
1364
+ "source": "ollama",
1365
+ "evaluator": "ollama",
1366
+ "selected_model": null,
1367
+ "confirmed_model": null,
1368
+ "wall_ms": 174.71,
1369
+ "state_bytes": 478,
1370
+ "metrics": {
1371
+ "input_tokens": 679,
1372
+ "output_tokens": 1
1373
+ }
1374
+ },
1375
+ {
1376
+ "round": 3,
1377
+ "case": "held-h-marker",
1378
+ "split": "heldout",
1379
+ "expected": "haiku",
1380
+ "actual": "haiku",
1381
+ "classified_tier": "haiku",
1382
+ "source": "ollama",
1383
+ "evaluator": "ollama",
1384
+ "selected_model": null,
1385
+ "confirmed_model": null,
1386
+ "wall_ms": 200.69,
1387
+ "state_bytes": 323,
1388
+ "metrics": {
1389
+ "input_tokens": 669,
1390
+ "output_tokens": 1
1391
+ }
1392
+ },
1393
+ {
1394
+ "round": 3,
1395
+ "case": "held-tier-influence-difficult",
1396
+ "split": "heldout",
1397
+ "expected": "opus",
1398
+ "actual": "opus",
1399
+ "classified_tier": "opus",
1400
+ "source": "ollama",
1401
+ "evaluator": "ollama",
1402
+ "selected_model": null,
1403
+ "confirmed_model": null,
1404
+ "wall_ms": 157.28,
1405
+ "state_bytes": 640,
1406
+ "metrics": {
1407
+ "input_tokens": 715,
1408
+ "output_tokens": 1
1409
+ }
1410
+ },
1411
+ {
1412
+ "round": 3,
1413
+ "case": "held-s-cli",
1414
+ "split": "heldout",
1415
+ "expected": "sonnet",
1416
+ "actual": "sonnet",
1417
+ "classified_tier": "sonnet",
1418
+ "source": "ollama",
1419
+ "evaluator": "ollama",
1420
+ "selected_model": null,
1421
+ "confirmed_model": null,
1422
+ "wall_ms": 164.11,
1423
+ "state_bytes": 480,
1424
+ "metrics": {
1425
+ "input_tokens": 687,
1426
+ "output_tokens": 1
1427
+ }
1428
+ },
1429
+ {
1430
+ "round": 3,
1431
+ "case": "held-tier-influence-mechanical",
1432
+ "split": "heldout",
1433
+ "expected": "haiku",
1434
+ "actual": "haiku",
1435
+ "classified_tier": "haiku",
1436
+ "source": "ollama",
1437
+ "evaluator": "ollama",
1438
+ "selected_model": null,
1439
+ "confirmed_model": null,
1440
+ "wall_ms": 167.5,
1441
+ "state_bytes": 432,
1442
+ "metrics": {
1443
+ "input_tokens": 695,
1444
+ "output_tokens": 1
1445
+ }
1446
+ },
1447
+ {
1448
+ "round": 3,
1449
+ "case": "held-o-exactly-once",
1450
+ "split": "heldout",
1451
+ "expected": "opus",
1452
+ "actual": "opus",
1453
+ "classified_tier": "opus",
1454
+ "source": "ollama",
1455
+ "evaluator": "ollama",
1456
+ "selected_model": null,
1457
+ "confirmed_model": null,
1458
+ "wall_ms": 159.71,
1459
+ "state_bytes": 676,
1460
+ "metrics": {
1461
+ "input_tokens": 715,
1462
+ "output_tokens": 1
1463
+ }
1464
+ },
1465
+ {
1466
+ "round": 3,
1467
+ "case": "held-french-race",
1468
+ "split": "heldout",
1469
+ "expected": "opus",
1470
+ "actual": "opus",
1471
+ "classified_tier": "opus",
1472
+ "source": "ollama",
1473
+ "evaluator": "ollama",
1474
+ "selected_model": null,
1475
+ "confirmed_model": null,
1476
+ "wall_ms": 161.08,
1477
+ "state_bytes": 510,
1478
+ "metrics": {
1479
+ "input_tokens": 701,
1480
+ "output_tokens": 1
1481
+ }
1482
+ },
1483
+ {
1484
+ "round": 3,
1485
+ "case": "held-h-summary",
1486
+ "split": "heldout",
1487
+ "expected": "haiku",
1488
+ "actual": "haiku",
1489
+ "classified_tier": "haiku",
1490
+ "source": "ollama",
1491
+ "evaluator": "ollama",
1492
+ "selected_model": null,
1493
+ "confirmed_model": null,
1494
+ "wall_ms": 222.8,
1495
+ "state_bytes": 406,
1496
+ "metrics": {
1497
+ "input_tokens": 677,
1498
+ "output_tokens": 1
1499
+ }
1500
+ },
1501
+ {
1502
+ "round": 3,
1503
+ "case": "held-s-refactor",
1504
+ "split": "heldout",
1505
+ "expected": "sonnet",
1506
+ "actual": "sonnet",
1507
+ "classified_tier": "sonnet",
1508
+ "source": "ollama",
1509
+ "evaluator": "ollama",
1510
+ "selected_model": null,
1511
+ "confirmed_model": null,
1512
+ "wall_ms": 178.21,
1513
+ "state_bytes": 464,
1514
+ "metrics": {
1515
+ "input_tokens": 673,
1516
+ "output_tokens": 1
1517
+ }
1518
+ },
1519
+ {
1520
+ "round": 3,
1521
+ "case": "held-short-difficult-followup",
1522
+ "split": "heldout",
1523
+ "expected": "opus",
1524
+ "actual": "opus",
1525
+ "classified_tier": "opus",
1526
+ "source": "ollama",
1527
+ "evaluator": "ollama",
1528
+ "selected_model": null,
1529
+ "confirmed_model": null,
1530
+ "wall_ms": 244.09,
1531
+ "state_bytes": 483,
1532
+ "metrics": {
1533
+ "input_tokens": 689,
1534
+ "output_tokens": 1
1535
+ }
1536
+ },
1537
+ {
1538
+ "round": 3,
1539
+ "case": "held-h-literal-opus",
1540
+ "split": "heldout",
1541
+ "expected": "haiku",
1542
+ "actual": "opus",
1543
+ "classified_tier": "opus",
1544
+ "source": "ollama",
1545
+ "evaluator": "ollama",
1546
+ "selected_model": null,
1547
+ "confirmed_model": null,
1548
+ "wall_ms": 232.54,
1549
+ "state_bytes": 312,
1550
+ "metrics": {
1551
+ "input_tokens": 655,
1552
+ "output_tokens": 1
1553
+ }
1554
+ },
1555
+ {
1556
+ "round": 3,
1557
+ "case": "held-o-injection",
1558
+ "split": "heldout",
1559
+ "expected": "opus",
1560
+ "actual": "opus",
1561
+ "classified_tier": "opus",
1562
+ "source": "ollama",
1563
+ "evaluator": "ollama",
1564
+ "selected_model": null,
1565
+ "confirmed_model": null,
1566
+ "wall_ms": 165.29,
1567
+ "state_bytes": 694,
1568
+ "metrics": {
1569
+ "input_tokens": 723,
1570
+ "output_tokens": 1
1571
+ }
1572
+ },
1573
+ {
1574
+ "round": 3,
1575
+ "case": "held-o-rolling-format",
1576
+ "split": "heldout",
1577
+ "expected": "opus",
1578
+ "actual": "opus",
1579
+ "classified_tier": "opus",
1580
+ "source": "ollama",
1581
+ "evaluator": "ollama",
1582
+ "selected_model": null,
1583
+ "confirmed_model": null,
1584
+ "wall_ms": 165.63,
1585
+ "state_bytes": 562,
1586
+ "metrics": {
1587
+ "input_tokens": 697,
1588
+ "output_tokens": 1
1589
+ }
1590
+ },
1591
+ {
1592
+ "round": 3,
1593
+ "case": "held-s-query",
1594
+ "split": "heldout",
1595
+ "expected": "sonnet",
1596
+ "actual": "sonnet",
1597
+ "classified_tier": "sonnet",
1598
+ "source": "ollama",
1599
+ "evaluator": "ollama",
1600
+ "selected_model": null,
1601
+ "confirmed_model": null,
1602
+ "wall_ms": 192.55,
1603
+ "state_bytes": 522,
1604
+ "metrics": {
1605
+ "input_tokens": 687,
1606
+ "output_tokens": 1
1607
+ }
1608
+ },
1609
+ {
1610
+ "round": 3,
1611
+ "case": "held-h-arithmetic",
1612
+ "split": "heldout",
1613
+ "expected": "haiku",
1614
+ "actual": "haiku",
1615
+ "classified_tier": "haiku",
1616
+ "source": "ollama",
1617
+ "evaluator": "ollama",
1618
+ "selected_model": null,
1619
+ "confirmed_model": null,
1620
+ "wall_ms": 166.79,
1621
+ "state_bytes": 298,
1622
+ "metrics": {
1623
+ "input_tokens": 667,
1624
+ "output_tokens": 1
1625
+ }
1626
+ },
1627
+ {
1628
+ "round": 3,
1629
+ "case": "held-o-snapshot",
1630
+ "split": "heldout",
1631
+ "expected": "opus",
1632
+ "actual": "opus",
1633
+ "classified_tier": "opus",
1634
+ "source": "ollama",
1635
+ "evaluator": "ollama",
1636
+ "selected_model": null,
1637
+ "confirmed_model": null,
1638
+ "wall_ms": 171.44,
1639
+ "state_bytes": 764,
1640
+ "metrics": {
1641
+ "input_tokens": 725,
1642
+ "output_tokens": 1
1643
+ }
1644
+ },
1645
+ {
1646
+ "round": 3,
1647
+ "case": "held-o-crypto",
1648
+ "split": "heldout",
1649
+ "expected": "opus",
1650
+ "actual": "opus",
1651
+ "classified_tier": "opus",
1652
+ "source": "ollama",
1653
+ "evaluator": "ollama",
1654
+ "selected_model": null,
1655
+ "confirmed_model": null,
1656
+ "wall_ms": 179.27,
1657
+ "state_bytes": 654,
1658
+ "metrics": {
1659
+ "input_tokens": 703,
1660
+ "output_tokens": 1
1661
+ }
1662
+ },
1663
+ {
1664
+ "round": 3,
1665
+ "case": "held-o-fencing-followup",
1666
+ "split": "heldout",
1667
+ "expected": "opus",
1668
+ "actual": "opus",
1669
+ "classified_tier": "opus",
1670
+ "source": "ollama",
1671
+ "evaluator": "ollama",
1672
+ "selected_model": null,
1673
+ "confirmed_model": null,
1674
+ "wall_ms": 191.34,
1675
+ "state_bytes": 780,
1676
+ "metrics": {
1677
+ "input_tokens": 753,
1678
+ "output_tokens": 1
1679
+ }
1680
+ },
1681
+ {
1682
+ "round": 3,
1683
+ "case": "held-h-translation",
1684
+ "split": "heldout",
1685
+ "expected": "haiku",
1686
+ "actual": "haiku",
1687
+ "classified_tier": "haiku",
1688
+ "source": "ollama",
1689
+ "evaluator": "ollama",
1690
+ "selected_model": null,
1691
+ "confirmed_model": null,
1692
+ "wall_ms": 187.48,
1693
+ "state_bytes": 290,
1694
+ "metrics": {
1695
+ "input_tokens": 649,
1696
+ "output_tokens": 1
1697
+ }
1698
+ }
1699
+ ],
1700
+ "summary": {
1701
+ "requests": 90,
1702
+ "valid": 90,
1703
+ "errors": 0,
1704
+ "exact_rubric_agreement": 0.9333333333333333,
1705
+ "under_routes": 3,
1706
+ "over_routes": 3,
1707
+ "opus_under_routes": 3,
1708
+ "wall_p50_ms": 244.09,
1709
+ "wall_p95_ms": 2929.63,
1710
+ "successful_wall_p50_ms": 244.09,
1711
+ "successful_wall_p95_ms": 2929.63,
1712
+ "confusion": {
1713
+ "haiku": {
1714
+ "haiku": 27,
1715
+ "sonnet": 0,
1716
+ "opus": 3,
1717
+ "error": 0
1718
+ },
1719
+ "sonnet": {
1720
+ "haiku": 0,
1721
+ "sonnet": 27,
1722
+ "opus": 0,
1723
+ "error": 0
1724
+ },
1725
+ "opus": {
1726
+ "haiku": 0,
1727
+ "sonnet": 3,
1728
+ "opus": 30,
1729
+ "error": 0
1730
+ }
1731
+ }
1732
+ },
1733
+ "acceptance": {
1734
+ "schema_version": 1,
1735
+ "policy": {
1736
+ "profile": "compatible",
1737
+ "minAgreement": 1,
1738
+ "maxUnderRouteRate": 0,
1739
+ "requiredTiers": [
1740
+ "haiku",
1741
+ "sonnet",
1742
+ "opus"
1743
+ ]
1744
+ },
1745
+ "gates": {
1746
+ "transport": {
1747
+ "passed": null,
1748
+ "reason": "Claude transport not exercised"
1749
+ },
1750
+ "evaluator": {
1751
+ "passed": true,
1752
+ "expected": "ollama",
1753
+ "evaluated_requests": 90,
1754
+ "successful_evaluations": 90,
1755
+ "fallbacks": 0
1756
+ },
1757
+ "rubric": {
1758
+ "passed": false,
1759
+ "agreement": 0.9333333333333333,
1760
+ "under_routes": 3,
1761
+ "under_route_rate": 0.03333333333333333,
1762
+ "min_agreement": 1,
1763
+ "max_under_route_rate": 0
1764
+ },
1765
+ "policy": {
1766
+ "passed": null,
1767
+ "reason": "Classifier-only benchmark; routing policy not exercised"
1768
+ },
1769
+ "coverage": {
1770
+ "passed": true,
1771
+ "required": [
1772
+ "haiku",
1773
+ "sonnet",
1774
+ "opus"
1775
+ ],
1776
+ "observed": [
1777
+ "haiku",
1778
+ "opus",
1779
+ "sonnet"
1780
+ ],
1781
+ "missing": []
1782
+ },
1783
+ "task": {
1784
+ "passed": null,
1785
+ "reason": "No Claude task completion measured"
1786
+ }
1787
+ },
1788
+ "passed": false,
1789
+ "requests": 90
1790
+ },
1791
+ "final_resident_memory": {
1792
+ "model_size_bytes": 2912388382,
1793
+ "model_vram_bytes": 2912388382,
1794
+ "context_length": 2050,
1795
+ "parameter_size": "4.2B",
1796
+ "quantization": "Q4_K_M",
1797
+ "runtime_process_rss_bytes": 4774903808,
1798
+ "system_free_bytes": 185729024
1799
+ },
1800
+ "max_budget_stress": {
1801
+ "rows": [
1802
+ {
1803
+ "case": "full-budget-0",
1804
+ "split": "performance_stress",
1805
+ "expected": "haiku",
1806
+ "actual": "haiku",
1807
+ "classified_tier": "haiku",
1808
+ "source": "ollama",
1809
+ "evaluator": "ollama",
1810
+ "selected_model": null,
1811
+ "confirmed_model": null,
1812
+ "wall_ms": 6694.33,
1813
+ "state_bytes": 3000,
1814
+ "metrics": {
1815
+ "input_tokens": 1737,
1816
+ "output_tokens": 1
1817
+ }
1818
+ },
1819
+ {
1820
+ "case": "full-budget-1",
1821
+ "split": "performance_stress",
1822
+ "expected": "haiku",
1823
+ "actual": "haiku",
1824
+ "classified_tier": "haiku",
1825
+ "source": "ollama",
1826
+ "evaluator": "ollama",
1827
+ "selected_model": null,
1828
+ "confirmed_model": null,
1829
+ "wall_ms": 6724.69,
1830
+ "state_bytes": 3000,
1831
+ "metrics": {
1832
+ "input_tokens": 1671,
1833
+ "output_tokens": 1
1834
+ }
1835
+ },
1836
+ {
1837
+ "case": "full-budget-2",
1838
+ "split": "performance_stress",
1839
+ "expected": "haiku",
1840
+ "actual": "haiku",
1841
+ "classified_tier": "haiku",
1842
+ "source": "ollama",
1843
+ "evaluator": "ollama",
1844
+ "selected_model": null,
1845
+ "confirmed_model": null,
1846
+ "wall_ms": 6766.26,
1847
+ "state_bytes": 3000,
1848
+ "metrics": {
1849
+ "input_tokens": 1693,
1850
+ "output_tokens": 1
1851
+ }
1852
+ },
1853
+ {
1854
+ "case": "full-budget-3",
1855
+ "split": "performance_stress",
1856
+ "expected": "haiku",
1857
+ "actual": "haiku",
1858
+ "classified_tier": "haiku",
1859
+ "source": "ollama",
1860
+ "evaluator": "ollama",
1861
+ "selected_model": null,
1862
+ "confirmed_model": null,
1863
+ "wall_ms": 7002.82,
1864
+ "state_bytes": 3000,
1865
+ "metrics": {
1866
+ "input_tokens": 1671,
1867
+ "output_tokens": 1
1868
+ }
1869
+ },
1870
+ {
1871
+ "case": "full-budget-4",
1872
+ "split": "performance_stress",
1873
+ "expected": "haiku",
1874
+ "actual": "haiku",
1875
+ "classified_tier": "haiku",
1876
+ "source": "ollama",
1877
+ "evaluator": "ollama",
1878
+ "selected_model": null,
1879
+ "confirmed_model": null,
1880
+ "wall_ms": 7045.82,
1881
+ "state_bytes": 3000,
1882
+ "metrics": {
1883
+ "input_tokens": 1693,
1884
+ "output_tokens": 1
1885
+ }
1886
+ },
1887
+ {
1888
+ "case": "full-budget-5",
1889
+ "split": "performance_stress",
1890
+ "expected": "haiku",
1891
+ "actual": "haiku",
1892
+ "classified_tier": "haiku",
1893
+ "source": "ollama",
1894
+ "evaluator": "ollama",
1895
+ "selected_model": null,
1896
+ "confirmed_model": null,
1897
+ "wall_ms": 7534.74,
1898
+ "state_bytes": 3000,
1899
+ "metrics": {
1900
+ "input_tokens": 1737,
1901
+ "output_tokens": 1
1902
+ }
1903
+ },
1904
+ {
1905
+ "case": "full-budget-6",
1906
+ "split": "performance_stress",
1907
+ "expected": "haiku",
1908
+ "actual": "haiku",
1909
+ "classified_tier": "haiku",
1910
+ "source": "ollama",
1911
+ "evaluator": "ollama",
1912
+ "selected_model": null,
1913
+ "confirmed_model": null,
1914
+ "wall_ms": 7854.15,
1915
+ "state_bytes": 3000,
1916
+ "metrics": {
1917
+ "input_tokens": 1759,
1918
+ "output_tokens": 1
1919
+ }
1920
+ },
1921
+ {
1922
+ "case": "full-budget-7",
1923
+ "split": "performance_stress",
1924
+ "expected": "haiku",
1925
+ "actual": "haiku",
1926
+ "classified_tier": "haiku",
1927
+ "source": "ollama",
1928
+ "evaluator": "ollama",
1929
+ "selected_model": null,
1930
+ "confirmed_model": null,
1931
+ "wall_ms": 7589.76,
1932
+ "state_bytes": 3000,
1933
+ "metrics": {
1934
+ "input_tokens": 1737,
1935
+ "output_tokens": 1
1936
+ }
1937
+ }
1938
+ ],
1939
+ "summary": {
1940
+ "requests": 8,
1941
+ "valid": 8,
1942
+ "errors": 0,
1943
+ "exact_rubric_agreement": 1,
1944
+ "under_routes": 0,
1945
+ "over_routes": 0,
1946
+ "opus_under_routes": 0,
1947
+ "wall_p50_ms": 7002.82,
1948
+ "wall_p95_ms": 7854.15,
1949
+ "successful_wall_p50_ms": 7002.82,
1950
+ "successful_wall_p95_ms": 7854.15,
1951
+ "confusion": {
1952
+ "haiku": {
1953
+ "haiku": 8,
1954
+ "sonnet": 0,
1955
+ "opus": 0,
1956
+ "error": 0
1957
+ },
1958
+ "sonnet": {
1959
+ "haiku": 0,
1960
+ "sonnet": 0,
1961
+ "opus": 0,
1962
+ "error": 0
1963
+ },
1964
+ "opus": {
1965
+ "haiku": 0,
1966
+ "sonnet": 0,
1967
+ "opus": 0,
1968
+ "error": 0
1969
+ }
1970
+ }
1971
+ },
1972
+ "acceptance": {
1973
+ "schema_version": 1,
1974
+ "policy": {
1975
+ "profile": "compatible",
1976
+ "minAgreement": 1,
1977
+ "maxUnderRouteRate": 0,
1978
+ "requiredTiers": [
1979
+ "haiku"
1980
+ ]
1981
+ },
1982
+ "gates": {
1983
+ "transport": {
1984
+ "passed": null,
1985
+ "reason": "Claude transport not exercised"
1986
+ },
1987
+ "evaluator": {
1988
+ "passed": true,
1989
+ "expected": "ollama",
1990
+ "evaluated_requests": 8,
1991
+ "successful_evaluations": 8,
1992
+ "fallbacks": 0
1993
+ },
1994
+ "rubric": {
1995
+ "passed": true,
1996
+ "agreement": 1,
1997
+ "under_routes": 0,
1998
+ "under_route_rate": 0,
1999
+ "min_agreement": 1,
2000
+ "max_under_route_rate": 0
2001
+ },
2002
+ "policy": {
2003
+ "passed": null,
2004
+ "reason": "Classifier-only benchmark; routing policy not exercised"
2005
+ },
2006
+ "coverage": {
2007
+ "passed": true,
2008
+ "required": [
2009
+ "haiku"
2010
+ ],
2011
+ "observed": [
2012
+ "haiku"
2013
+ ],
2014
+ "missing": []
2015
+ },
2016
+ "task": {
2017
+ "passed": null,
2018
+ "reason": "No Claude task completion measured"
2019
+ }
2020
+ },
2021
+ "passed": true,
2022
+ "requests": 8
2023
+ }
2024
+ }
2025
+ },
2026
+ {
2027
+ "model": "nimble:9b-q4_K_M",
2028
+ "digest": "3776806da5587387a996e28e75d5d07fbebe7410d47879e1f71782f36c896ce3",
2029
+ "download_bytes": 5629109675,
2030
+ "protocol": "/v1/systemone",
2031
+ "rubric_sha256": "be151cedb4de4b7ef3f7162d751f70ce7d9dd14efc66fae1835f73ffd04027be",
2032
+ "context_configuration": "model/server defaults; measured below",
2033
+ "cold": {
2034
+ "case": "held-h-arithmetic",
2035
+ "split": "heldout",
2036
+ "expected": "haiku",
2037
+ "actual": "haiku",
2038
+ "classified_tier": "haiku",
2039
+ "source": "ollama",
2040
+ "evaluator": "ollama",
2041
+ "selected_model": null,
2042
+ "confirmed_model": null,
2043
+ "wall_ms": 15926.94,
2044
+ "state_bytes": 298,
2045
+ "metrics": {
2046
+ "input_tokens": 705,
2047
+ "output_tokens": 1
2048
+ }
2049
+ },
2050
+ "resident_memory": {
2051
+ "model_size_bytes": 6166423796,
2052
+ "model_vram_bytes": 6166423796,
2053
+ "context_length": 8194,
2054
+ "parameter_size": "9.0B",
2055
+ "quantization": "Q4_K_M",
2056
+ "runtime_process_rss_bytes": 6072107008,
2057
+ "system_free_bytes": 165101568
2058
+ },
2059
+ "rows": [
2060
+ {
2061
+ "round": 1,
2062
+ "case": "held-h-link",
2063
+ "split": "heldout",
2064
+ "expected": "haiku",
2065
+ "actual": "haiku",
2066
+ "classified_tier": "haiku",
2067
+ "source": "ollama",
2068
+ "evaluator": "ollama",
2069
+ "selected_model": null,
2070
+ "confirmed_model": null,
2071
+ "wall_ms": 7191.98,
2072
+ "state_bytes": 386,
2073
+ "metrics": {
2074
+ "input_tokens": 711,
2075
+ "output_tokens": 1
2076
+ }
2077
+ },
2078
+ {
2079
+ "round": 1,
2080
+ "case": "held-o-injection",
2081
+ "split": "heldout",
2082
+ "expected": "opus",
2083
+ "actual": "haiku",
2084
+ "classified_tier": "haiku",
2085
+ "source": "ollama",
2086
+ "evaluator": "ollama",
2087
+ "selected_model": null,
2088
+ "confirmed_model": null,
2089
+ "wall_ms": 5664,
2090
+ "state_bytes": 694,
2091
+ "metrics": {
2092
+ "input_tokens": 761,
2093
+ "output_tokens": 1
2094
+ }
2095
+ },
2096
+ {
2097
+ "round": 1,
2098
+ "case": "held-h-marker",
2099
+ "split": "heldout",
2100
+ "expected": "haiku",
2101
+ "actual": "haiku",
2102
+ "classified_tier": "haiku",
2103
+ "source": "ollama",
2104
+ "evaluator": "ollama",
2105
+ "selected_model": null,
2106
+ "confirmed_model": null,
2107
+ "wall_ms": 5518.35,
2108
+ "state_bytes": 323,
2109
+ "metrics": {
2110
+ "input_tokens": 707,
2111
+ "output_tokens": 1
2112
+ }
2113
+ },
2114
+ {
2115
+ "round": 1,
2116
+ "case": "held-tier-influence-mechanical",
2117
+ "split": "heldout",
2118
+ "expected": "haiku",
2119
+ "actual": "haiku",
2120
+ "classified_tier": "haiku",
2121
+ "source": "ollama",
2122
+ "evaluator": "ollama",
2123
+ "selected_model": null,
2124
+ "confirmed_model": null,
2125
+ "wall_ms": 5754.87,
2126
+ "state_bytes": 432,
2127
+ "metrics": {
2128
+ "input_tokens": 733,
2129
+ "output_tokens": 1
2130
+ }
2131
+ },
2132
+ {
2133
+ "round": 1,
2134
+ "case": "held-s-tool-followup",
2135
+ "split": "heldout",
2136
+ "expected": "sonnet",
2137
+ "actual": "sonnet",
2138
+ "classified_tier": "sonnet",
2139
+ "source": "ollama",
2140
+ "evaluator": "ollama",
2141
+ "selected_model": null,
2142
+ "confirmed_model": null,
2143
+ "wall_ms": 6105.83,
2144
+ "state_bytes": 508,
2145
+ "metrics": {
2146
+ "input_tokens": 753,
2147
+ "output_tokens": 1
2148
+ }
2149
+ },
2150
+ {
2151
+ "round": 1,
2152
+ "case": "held-h-reminder",
2153
+ "split": "heldout",
2154
+ "expected": "haiku",
2155
+ "actual": "haiku",
2156
+ "classified_tier": "haiku",
2157
+ "source": "ollama",
2158
+ "evaluator": "ollama",
2159
+ "selected_model": null,
2160
+ "confirmed_model": null,
2161
+ "wall_ms": 6299.93,
2162
+ "state_bytes": 294,
2163
+ "metrics": {
2164
+ "input_tokens": 709,
2165
+ "output_tokens": 1
2166
+ }
2167
+ },
2168
+ {
2169
+ "round": 1,
2170
+ "case": "held-h-injection",
2171
+ "split": "heldout",
2172
+ "expected": "haiku",
2173
+ "actual": "haiku",
2174
+ "classified_tier": "haiku",
2175
+ "source": "ollama",
2176
+ "evaluator": "ollama",
2177
+ "selected_model": null,
2178
+ "confirmed_model": null,
2179
+ "wall_ms": 6575.67,
2180
+ "state_bytes": 512,
2181
+ "metrics": {
2182
+ "input_tokens": 735,
2183
+ "output_tokens": 1
2184
+ }
2185
+ },
2186
+ {
2187
+ "round": 1,
2188
+ "case": "held-o-exactly-once",
2189
+ "split": "heldout",
2190
+ "expected": "opus",
2191
+ "actual": null,
2192
+ "classified_tier": null,
2193
+ "source": "error",
2194
+ "evaluator": "ollama",
2195
+ "selected_model": null,
2196
+ "confirmed_model": null,
2197
+ "wall_ms": 56589.66,
2198
+ "state_bytes": 676,
2199
+ "error": "timeout"
2200
+ },
2201
+ {
2202
+ "round": 1,
2203
+ "case": "held-s-refactor",
2204
+ "split": "heldout",
2205
+ "expected": "sonnet",
2206
+ "actual": null,
2207
+ "classified_tier": null,
2208
+ "source": "error",
2209
+ "evaluator": "ollama",
2210
+ "selected_model": null,
2211
+ "confirmed_model": null,
2212
+ "wall_ms": 969208.59,
2213
+ "state_bytes": 464,
2214
+ "error": "timeout"
2215
+ },
2216
+ {
2217
+ "round": 1,
2218
+ "case": "held-o-tenant-cache",
2219
+ "split": "heldout",
2220
+ "expected": "opus",
2221
+ "actual": "opus",
2222
+ "classified_tier": "opus",
2223
+ "source": "ollama",
2224
+ "evaluator": "ollama",
2225
+ "selected_model": null,
2226
+ "confirmed_model": null,
2227
+ "wall_ms": 13644.66,
2228
+ "state_bytes": 658,
2229
+ "metrics": {
2230
+ "input_tokens": 745,
2231
+ "output_tokens": 1
2232
+ }
2233
+ },
2234
+ {
2235
+ "round": 1,
2236
+ "case": "held-s-unclear-polish",
2237
+ "split": "heldout",
2238
+ "expected": "sonnet",
2239
+ "actual": "sonnet",
2240
+ "classified_tier": "sonnet",
2241
+ "source": "ollama",
2242
+ "evaluator": "ollama",
2243
+ "selected_model": null,
2244
+ "confirmed_model": null,
2245
+ "wall_ms": 13881.57,
2246
+ "state_bytes": 478,
2247
+ "metrics": {
2248
+ "input_tokens": 717,
2249
+ "output_tokens": 1
2250
+ }
2251
+ },
2252
+ {
2253
+ "round": 1,
2254
+ "case": "held-s-cli",
2255
+ "split": "heldout",
2256
+ "expected": "sonnet",
2257
+ "actual": "sonnet",
2258
+ "classified_tier": "sonnet",
2259
+ "source": "ollama",
2260
+ "evaluator": "ollama",
2261
+ "selected_model": null,
2262
+ "confirmed_model": null,
2263
+ "wall_ms": 12440.65,
2264
+ "state_bytes": 480,
2265
+ "metrics": {
2266
+ "input_tokens": 725,
2267
+ "output_tokens": 1
2268
+ }
2269
+ },
2270
+ {
2271
+ "round": 1,
2272
+ "case": "held-s-validation",
2273
+ "split": "heldout",
2274
+ "expected": "sonnet",
2275
+ "actual": "sonnet",
2276
+ "classified_tier": "sonnet",
2277
+ "source": "ollama",
2278
+ "evaluator": "ollama",
2279
+ "selected_model": null,
2280
+ "confirmed_model": null,
2281
+ "wall_ms": 11249.27,
2282
+ "state_bytes": 484,
2283
+ "metrics": {
2284
+ "input_tokens": 715,
2285
+ "output_tokens": 1
2286
+ }
2287
+ },
2288
+ {
2289
+ "round": 1,
2290
+ "case": "held-s-query",
2291
+ "split": "heldout",
2292
+ "expected": "sonnet",
2293
+ "actual": "sonnet",
2294
+ "classified_tier": "sonnet",
2295
+ "source": "ollama",
2296
+ "evaluator": "ollama",
2297
+ "selected_model": null,
2298
+ "confirmed_model": null,
2299
+ "wall_ms": 10910.16,
2300
+ "state_bytes": 522,
2301
+ "metrics": {
2302
+ "input_tokens": 725,
2303
+ "output_tokens": 1
2304
+ }
2305
+ },
2306
+ {
2307
+ "round": 1,
2308
+ "case": "held-o-crypto",
2309
+ "split": "heldout",
2310
+ "expected": "opus",
2311
+ "actual": "opus",
2312
+ "classified_tier": "opus",
2313
+ "source": "ollama",
2314
+ "evaluator": "ollama",
2315
+ "selected_model": null,
2316
+ "confirmed_model": null,
2317
+ "wall_ms": 11604.32,
2318
+ "state_bytes": 654,
2319
+ "metrics": {
2320
+ "input_tokens": 741,
2321
+ "output_tokens": 1
2322
+ }
2323
+ },
2324
+ {
2325
+ "round": 1,
2326
+ "case": "held-french-race",
2327
+ "split": "heldout",
2328
+ "expected": "opus",
2329
+ "actual": "opus",
2330
+ "classified_tier": "opus",
2331
+ "source": "ollama",
2332
+ "evaluator": "ollama",
2333
+ "selected_model": null,
2334
+ "confirmed_model": null,
2335
+ "wall_ms": 10658.91,
2336
+ "state_bytes": 510,
2337
+ "metrics": {
2338
+ "input_tokens": 739,
2339
+ "output_tokens": 1
2340
+ }
2341
+ },
2342
+ {
2343
+ "round": 1,
2344
+ "case": "held-spanish-feature",
2345
+ "split": "heldout",
2346
+ "expected": "sonnet",
2347
+ "actual": "sonnet",
2348
+ "classified_tier": "sonnet",
2349
+ "source": "ollama",
2350
+ "evaluator": "ollama",
2351
+ "selected_model": null,
2352
+ "confirmed_model": null,
2353
+ "wall_ms": 10153.48,
2354
+ "state_bytes": 380,
2355
+ "metrics": {
2356
+ "input_tokens": 711,
2357
+ "output_tokens": 1
2358
+ }
2359
+ },
2360
+ {
2361
+ "round": 1,
2362
+ "case": "held-h-translation",
2363
+ "split": "heldout",
2364
+ "expected": "haiku",
2365
+ "actual": "haiku",
2366
+ "classified_tier": "haiku",
2367
+ "source": "ollama",
2368
+ "evaluator": "ollama",
2369
+ "selected_model": null,
2370
+ "confirmed_model": null,
2371
+ "wall_ms": 10721.07,
2372
+ "state_bytes": 290,
2373
+ "metrics": {
2374
+ "input_tokens": 687,
2375
+ "output_tokens": 1
2376
+ }
2377
+ },
2378
+ {
2379
+ "round": 1,
2380
+ "case": "held-h-summary",
2381
+ "split": "heldout",
2382
+ "expected": "haiku",
2383
+ "actual": "haiku",
2384
+ "classified_tier": "haiku",
2385
+ "source": "ollama",
2386
+ "evaluator": "ollama",
2387
+ "selected_model": null,
2388
+ "confirmed_model": null,
2389
+ "wall_ms": 10636.97,
2390
+ "state_bytes": 406,
2391
+ "metrics": {
2392
+ "input_tokens": 715,
2393
+ "output_tokens": 1
2394
+ }
2395
+ },
2396
+ {
2397
+ "round": 1,
2398
+ "case": "held-o-compiler",
2399
+ "split": "heldout",
2400
+ "expected": "opus",
2401
+ "actual": null,
2402
+ "classified_tier": null,
2403
+ "source": "error",
2404
+ "evaluator": "ollama",
2405
+ "selected_model": null,
2406
+ "confirmed_model": null,
2407
+ "wall_ms": 173623.07,
2408
+ "state_bytes": 654,
2409
+ "error": "timeout"
2410
+ },
2411
+ {
2412
+ "round": 1,
2413
+ "case": "held-tier-influence-difficult",
2414
+ "split": "heldout",
2415
+ "expected": "opus",
2416
+ "actual": "opus",
2417
+ "classified_tier": "opus",
2418
+ "source": "ollama",
2419
+ "evaluator": "ollama",
2420
+ "selected_model": null,
2421
+ "confirmed_model": null,
2422
+ "wall_ms": 7404.93,
2423
+ "state_bytes": 640,
2424
+ "metrics": {
2425
+ "input_tokens": 753,
2426
+ "output_tokens": 1
2427
+ }
2428
+ },
2429
+ {
2430
+ "round": 1,
2431
+ "case": "held-o-rolling-format",
2432
+ "split": "heldout",
2433
+ "expected": "opus",
2434
+ "actual": "opus",
2435
+ "classified_tier": "opus",
2436
+ "source": "ollama",
2437
+ "evaluator": "ollama",
2438
+ "selected_model": null,
2439
+ "confirmed_model": null,
2440
+ "wall_ms": 4998.11,
2441
+ "state_bytes": 562,
2442
+ "metrics": {
2443
+ "input_tokens": 735,
2444
+ "output_tokens": 1
2445
+ }
2446
+ },
2447
+ {
2448
+ "round": 1,
2449
+ "case": "held-h-arithmetic",
2450
+ "split": "heldout",
2451
+ "expected": "haiku",
2452
+ "actual": "haiku",
2453
+ "classified_tier": "haiku",
2454
+ "source": "ollama",
2455
+ "evaluator": "ollama",
2456
+ "selected_model": null,
2457
+ "confirmed_model": null,
2458
+ "wall_ms": 1257.05,
2459
+ "state_bytes": 298,
2460
+ "metrics": {
2461
+ "input_tokens": 705,
2462
+ "output_tokens": 1
2463
+ }
2464
+ },
2465
+ {
2466
+ "round": 1,
2467
+ "case": "held-h-literal-opus",
2468
+ "split": "heldout",
2469
+ "expected": "haiku",
2470
+ "actual": "haiku",
2471
+ "classified_tier": "haiku",
2472
+ "source": "ollama",
2473
+ "evaluator": "ollama",
2474
+ "selected_model": null,
2475
+ "confirmed_model": null,
2476
+ "wall_ms": 4707.2,
2477
+ "state_bytes": 312,
2478
+ "metrics": {
2479
+ "input_tokens": 693,
2480
+ "output_tokens": 1
2481
+ }
2482
+ },
2483
+ {
2484
+ "round": 1,
2485
+ "case": "held-o-snapshot",
2486
+ "split": "heldout",
2487
+ "expected": "opus",
2488
+ "actual": "opus",
2489
+ "classified_tier": "opus",
2490
+ "source": "ollama",
2491
+ "evaluator": "ollama",
2492
+ "selected_model": null,
2493
+ "confirmed_model": null,
2494
+ "wall_ms": 5212.16,
2495
+ "state_bytes": 764,
2496
+ "metrics": {
2497
+ "input_tokens": 763,
2498
+ "output_tokens": 1
2499
+ }
2500
+ },
2501
+ {
2502
+ "round": 1,
2503
+ "case": "held-italian-literal",
2504
+ "split": "heldout",
2505
+ "expected": "haiku",
2506
+ "actual": "haiku",
2507
+ "classified_tier": "haiku",
2508
+ "source": "ollama",
2509
+ "evaluator": "ollama",
2510
+ "selected_model": null,
2511
+ "confirmed_model": null,
2512
+ "wall_ms": 4690.57,
2513
+ "state_bytes": 252,
2514
+ "metrics": {
2515
+ "input_tokens": 683,
2516
+ "output_tokens": 1
2517
+ }
2518
+ },
2519
+ {
2520
+ "round": 1,
2521
+ "case": "held-short-difficult-followup",
2522
+ "split": "heldout",
2523
+ "expected": "opus",
2524
+ "actual": "opus",
2525
+ "classified_tier": "opus",
2526
+ "source": "ollama",
2527
+ "evaluator": "ollama",
2528
+ "selected_model": null,
2529
+ "confirmed_model": null,
2530
+ "wall_ms": 4942,
2531
+ "state_bytes": 483,
2532
+ "metrics": {
2533
+ "input_tokens": 727,
2534
+ "output_tokens": 1
2535
+ }
2536
+ },
2537
+ {
2538
+ "round": 1,
2539
+ "case": "held-o-fencing-followup",
2540
+ "split": "heldout",
2541
+ "expected": "opus",
2542
+ "actual": "opus",
2543
+ "classified_tier": "opus",
2544
+ "source": "ollama",
2545
+ "evaluator": "ollama",
2546
+ "selected_model": null,
2547
+ "confirmed_model": null,
2548
+ "wall_ms": 5351.05,
2549
+ "state_bytes": 780,
2550
+ "metrics": {
2551
+ "input_tokens": 791,
2552
+ "output_tokens": 1
2553
+ }
2554
+ },
2555
+ {
2556
+ "round": 1,
2557
+ "case": "held-s-null-bug",
2558
+ "split": "heldout",
2559
+ "expected": "sonnet",
2560
+ "actual": "sonnet",
2561
+ "classified_tier": "sonnet",
2562
+ "source": "ollama",
2563
+ "evaluator": "ollama",
2564
+ "selected_model": null,
2565
+ "confirmed_model": null,
2566
+ "wall_ms": 4944.76,
2567
+ "state_bytes": 546,
2568
+ "metrics": {
2569
+ "input_tokens": 729,
2570
+ "output_tokens": 1
2571
+ }
2572
+ },
2573
+ {
2574
+ "round": 1,
2575
+ "case": "held-s-pagination",
2576
+ "split": "heldout",
2577
+ "expected": "sonnet",
2578
+ "actual": "sonnet",
2579
+ "classified_tier": "sonnet",
2580
+ "source": "ollama",
2581
+ "evaluator": "ollama",
2582
+ "selected_model": null,
2583
+ "confirmed_model": null,
2584
+ "wall_ms": 5004.53,
2585
+ "state_bytes": 482,
2586
+ "metrics": {
2587
+ "input_tokens": 717,
2588
+ "output_tokens": 1
2589
+ }
2590
+ },
2591
+ {
2592
+ "round": 2,
2593
+ "case": "held-s-null-bug",
2594
+ "split": "heldout",
2595
+ "expected": "sonnet",
2596
+ "actual": "sonnet",
2597
+ "classified_tier": "sonnet",
2598
+ "source": "ollama",
2599
+ "evaluator": "ollama",
2600
+ "selected_model": null,
2601
+ "confirmed_model": null,
2602
+ "wall_ms": 283.58,
2603
+ "state_bytes": 546,
2604
+ "metrics": {
2605
+ "input_tokens": 729,
2606
+ "output_tokens": 1
2607
+ }
2608
+ },
2609
+ {
2610
+ "round": 2,
2611
+ "case": "held-o-injection",
2612
+ "split": "heldout",
2613
+ "expected": "opus",
2614
+ "actual": "haiku",
2615
+ "classified_tier": "haiku",
2616
+ "source": "ollama",
2617
+ "evaluator": "ollama",
2618
+ "selected_model": null,
2619
+ "confirmed_model": null,
2620
+ "wall_ms": 1350.31,
2621
+ "state_bytes": 694,
2622
+ "metrics": {
2623
+ "input_tokens": 761,
2624
+ "output_tokens": 1
2625
+ }
2626
+ },
2627
+ {
2628
+ "round": 2,
2629
+ "case": "held-o-compiler",
2630
+ "split": "heldout",
2631
+ "expected": "opus",
2632
+ "actual": "opus",
2633
+ "classified_tier": "opus",
2634
+ "source": "ollama",
2635
+ "evaluator": "ollama",
2636
+ "selected_model": null,
2637
+ "confirmed_model": null,
2638
+ "wall_ms": 416.71,
2639
+ "state_bytes": 654,
2640
+ "metrics": {
2641
+ "input_tokens": 737,
2642
+ "output_tokens": 1
2643
+ }
2644
+ },
2645
+ {
2646
+ "round": 2,
2647
+ "case": "held-h-injection",
2648
+ "split": "heldout",
2649
+ "expected": "haiku",
2650
+ "actual": "haiku",
2651
+ "classified_tier": "haiku",
2652
+ "source": "ollama",
2653
+ "evaluator": "ollama",
2654
+ "selected_model": null,
2655
+ "confirmed_model": null,
2656
+ "wall_ms": 1072.63,
2657
+ "state_bytes": 512,
2658
+ "metrics": {
2659
+ "input_tokens": 735,
2660
+ "output_tokens": 1
2661
+ }
2662
+ },
2663
+ {
2664
+ "round": 2,
2665
+ "case": "held-o-rolling-format",
2666
+ "split": "heldout",
2667
+ "expected": "opus",
2668
+ "actual": "opus",
2669
+ "classified_tier": "opus",
2670
+ "source": "ollama",
2671
+ "evaluator": "ollama",
2672
+ "selected_model": null,
2673
+ "confirmed_model": null,
2674
+ "wall_ms": 278.46,
2675
+ "state_bytes": 562,
2676
+ "metrics": {
2677
+ "input_tokens": 735,
2678
+ "output_tokens": 1
2679
+ }
2680
+ },
2681
+ {
2682
+ "round": 2,
2683
+ "case": "held-s-pagination",
2684
+ "split": "heldout",
2685
+ "expected": "sonnet",
2686
+ "actual": "sonnet",
2687
+ "classified_tier": "sonnet",
2688
+ "source": "ollama",
2689
+ "evaluator": "ollama",
2690
+ "selected_model": null,
2691
+ "confirmed_model": null,
2692
+ "wall_ms": 246.01,
2693
+ "state_bytes": 482,
2694
+ "metrics": {
2695
+ "input_tokens": 717,
2696
+ "output_tokens": 1
2697
+ }
2698
+ },
2699
+ {
2700
+ "round": 2,
2701
+ "case": "held-s-refactor",
2702
+ "split": "heldout",
2703
+ "expected": "sonnet",
2704
+ "actual": "sonnet",
2705
+ "classified_tier": "sonnet",
2706
+ "source": "ollama",
2707
+ "evaluator": "ollama",
2708
+ "selected_model": null,
2709
+ "confirmed_model": null,
2710
+ "wall_ms": 931.79,
2711
+ "state_bytes": 464,
2712
+ "metrics": {
2713
+ "input_tokens": 711,
2714
+ "output_tokens": 1
2715
+ }
2716
+ },
2717
+ {
2718
+ "round": 2,
2719
+ "case": "held-o-fencing-followup",
2720
+ "split": "heldout",
2721
+ "expected": "opus",
2722
+ "actual": "opus",
2723
+ "classified_tier": "opus",
2724
+ "source": "ollama",
2725
+ "evaluator": "ollama",
2726
+ "selected_model": null,
2727
+ "confirmed_model": null,
2728
+ "wall_ms": 271.43,
2729
+ "state_bytes": 780,
2730
+ "metrics": {
2731
+ "input_tokens": 791,
2732
+ "output_tokens": 1
2733
+ }
2734
+ },
2735
+ {
2736
+ "round": 2,
2737
+ "case": "held-s-tool-followup",
2738
+ "split": "heldout",
2739
+ "expected": "sonnet",
2740
+ "actual": "sonnet",
2741
+ "classified_tier": "sonnet",
2742
+ "source": "ollama",
2743
+ "evaluator": "ollama",
2744
+ "selected_model": null,
2745
+ "confirmed_model": null,
2746
+ "wall_ms": 1060.89,
2747
+ "state_bytes": 508,
2748
+ "metrics": {
2749
+ "input_tokens": 753,
2750
+ "output_tokens": 1
2751
+ }
2752
+ },
2753
+ {
2754
+ "round": 2,
2755
+ "case": "held-spanish-feature",
2756
+ "split": "heldout",
2757
+ "expected": "sonnet",
2758
+ "actual": "sonnet",
2759
+ "classified_tier": "sonnet",
2760
+ "source": "ollama",
2761
+ "evaluator": "ollama",
2762
+ "selected_model": null,
2763
+ "confirmed_model": null,
2764
+ "wall_ms": 381.21,
2765
+ "state_bytes": 380,
2766
+ "metrics": {
2767
+ "input_tokens": 711,
2768
+ "output_tokens": 1
2769
+ }
2770
+ },
2771
+ {
2772
+ "round": 2,
2773
+ "case": "held-h-arithmetic",
2774
+ "split": "heldout",
2775
+ "expected": "haiku",
2776
+ "actual": "haiku",
2777
+ "classified_tier": "haiku",
2778
+ "source": "ollama",
2779
+ "evaluator": "ollama",
2780
+ "selected_model": null,
2781
+ "confirmed_model": null,
2782
+ "wall_ms": 276.52,
2783
+ "state_bytes": 298,
2784
+ "metrics": {
2785
+ "input_tokens": 705,
2786
+ "output_tokens": 1
2787
+ }
2788
+ },
2789
+ {
2790
+ "round": 2,
2791
+ "case": "held-o-tenant-cache",
2792
+ "split": "heldout",
2793
+ "expected": "opus",
2794
+ "actual": "opus",
2795
+ "classified_tier": "opus",
2796
+ "source": "ollama",
2797
+ "evaluator": "ollama",
2798
+ "selected_model": null,
2799
+ "confirmed_model": null,
2800
+ "wall_ms": 2209.12,
2801
+ "state_bytes": 658,
2802
+ "metrics": {
2803
+ "input_tokens": 745,
2804
+ "output_tokens": 1
2805
+ }
2806
+ },
2807
+ {
2808
+ "round": 2,
2809
+ "case": "held-h-literal-opus",
2810
+ "split": "heldout",
2811
+ "expected": "haiku",
2812
+ "actual": "haiku",
2813
+ "classified_tier": "haiku",
2814
+ "source": "ollama",
2815
+ "evaluator": "ollama",
2816
+ "selected_model": null,
2817
+ "confirmed_model": null,
2818
+ "wall_ms": 343.71,
2819
+ "state_bytes": 312,
2820
+ "metrics": {
2821
+ "input_tokens": 693,
2822
+ "output_tokens": 1
2823
+ }
2824
+ },
2825
+ {
2826
+ "round": 2,
2827
+ "case": "held-h-marker",
2828
+ "split": "heldout",
2829
+ "expected": "haiku",
2830
+ "actual": "haiku",
2831
+ "classified_tier": "haiku",
2832
+ "source": "ollama",
2833
+ "evaluator": "ollama",
2834
+ "selected_model": null,
2835
+ "confirmed_model": null,
2836
+ "wall_ms": 2038.34,
2837
+ "state_bytes": 323,
2838
+ "metrics": {
2839
+ "input_tokens": 707,
2840
+ "output_tokens": 1
2841
+ }
2842
+ },
2843
+ {
2844
+ "round": 2,
2845
+ "case": "held-tier-influence-mechanical",
2846
+ "split": "heldout",
2847
+ "expected": "haiku",
2848
+ "actual": "haiku",
2849
+ "classified_tier": "haiku",
2850
+ "source": "ollama",
2851
+ "evaluator": "ollama",
2852
+ "selected_model": null,
2853
+ "confirmed_model": null,
2854
+ "wall_ms": 2522.32,
2855
+ "state_bytes": 432,
2856
+ "metrics": {
2857
+ "input_tokens": 733,
2858
+ "output_tokens": 1
2859
+ }
2860
+ },
2861
+ {
2862
+ "round": 2,
2863
+ "case": "held-s-unclear-polish",
2864
+ "split": "heldout",
2865
+ "expected": "sonnet",
2866
+ "actual": "sonnet",
2867
+ "classified_tier": "sonnet",
2868
+ "source": "ollama",
2869
+ "evaluator": "ollama",
2870
+ "selected_model": null,
2871
+ "confirmed_model": null,
2872
+ "wall_ms": 4392.7,
2873
+ "state_bytes": 478,
2874
+ "metrics": {
2875
+ "input_tokens": 717,
2876
+ "output_tokens": 1
2877
+ }
2878
+ },
2879
+ {
2880
+ "round": 2,
2881
+ "case": "held-o-snapshot",
2882
+ "split": "heldout",
2883
+ "expected": "opus",
2884
+ "actual": "opus",
2885
+ "classified_tier": "opus",
2886
+ "source": "ollama",
2887
+ "evaluator": "ollama",
2888
+ "selected_model": null,
2889
+ "confirmed_model": null,
2890
+ "wall_ms": 263.43,
2891
+ "state_bytes": 764,
2892
+ "metrics": {
2893
+ "input_tokens": 763,
2894
+ "output_tokens": 1
2895
+ }
2896
+ },
2897
+ {
2898
+ "round": 2,
2899
+ "case": "held-s-cli",
2900
+ "split": "heldout",
2901
+ "expected": "sonnet",
2902
+ "actual": "sonnet",
2903
+ "classified_tier": "sonnet",
2904
+ "source": "ollama",
2905
+ "evaluator": "ollama",
2906
+ "selected_model": null,
2907
+ "confirmed_model": null,
2908
+ "wall_ms": 1721.85,
2909
+ "state_bytes": 480,
2910
+ "metrics": {
2911
+ "input_tokens": 725,
2912
+ "output_tokens": 1
2913
+ }
2914
+ },
2915
+ {
2916
+ "round": 2,
2917
+ "case": "held-h-summary",
2918
+ "split": "heldout",
2919
+ "expected": "haiku",
2920
+ "actual": "haiku",
2921
+ "classified_tier": "haiku",
2922
+ "source": "ollama",
2923
+ "evaluator": "ollama",
2924
+ "selected_model": null,
2925
+ "confirmed_model": null,
2926
+ "wall_ms": 1362.47,
2927
+ "state_bytes": 406,
2928
+ "metrics": {
2929
+ "input_tokens": 715,
2930
+ "output_tokens": 1
2931
+ }
2932
+ },
2933
+ {
2934
+ "round": 2,
2935
+ "case": "held-short-difficult-followup",
2936
+ "split": "heldout",
2937
+ "expected": "opus",
2938
+ "actual": "opus",
2939
+ "classified_tier": "opus",
2940
+ "source": "ollama",
2941
+ "evaluator": "ollama",
2942
+ "selected_model": null,
2943
+ "confirmed_model": null,
2944
+ "wall_ms": 250.6,
2945
+ "state_bytes": 483,
2946
+ "metrics": {
2947
+ "input_tokens": 727,
2948
+ "output_tokens": 1
2949
+ }
2950
+ },
2951
+ {
2952
+ "round": 2,
2953
+ "case": "held-italian-literal",
2954
+ "split": "heldout",
2955
+ "expected": "haiku",
2956
+ "actual": "haiku",
2957
+ "classified_tier": "haiku",
2958
+ "source": "ollama",
2959
+ "evaluator": "ollama",
2960
+ "selected_model": null,
2961
+ "confirmed_model": null,
2962
+ "wall_ms": 254.08,
2963
+ "state_bytes": 252,
2964
+ "metrics": {
2965
+ "input_tokens": 683,
2966
+ "output_tokens": 1
2967
+ }
2968
+ },
2969
+ {
2970
+ "round": 2,
2971
+ "case": "held-s-validation",
2972
+ "split": "heldout",
2973
+ "expected": "sonnet",
2974
+ "actual": "sonnet",
2975
+ "classified_tier": "sonnet",
2976
+ "source": "ollama",
2977
+ "evaluator": "ollama",
2978
+ "selected_model": null,
2979
+ "confirmed_model": null,
2980
+ "wall_ms": 1381.37,
2981
+ "state_bytes": 484,
2982
+ "metrics": {
2983
+ "input_tokens": 715,
2984
+ "output_tokens": 1
2985
+ }
2986
+ },
2987
+ {
2988
+ "round": 2,
2989
+ "case": "held-o-crypto",
2990
+ "split": "heldout",
2991
+ "expected": "opus",
2992
+ "actual": "opus",
2993
+ "classified_tier": "opus",
2994
+ "source": "ollama",
2995
+ "evaluator": "ollama",
2996
+ "selected_model": null,
2997
+ "confirmed_model": null,
2998
+ "wall_ms": 3289.41,
2999
+ "state_bytes": 654,
3000
+ "metrics": {
3001
+ "input_tokens": 741,
3002
+ "output_tokens": 1
3003
+ }
3004
+ },
3005
+ {
3006
+ "round": 2,
3007
+ "case": "held-h-reminder",
3008
+ "split": "heldout",
3009
+ "expected": "haiku",
3010
+ "actual": "haiku",
3011
+ "classified_tier": "haiku",
3012
+ "source": "ollama",
3013
+ "evaluator": "ollama",
3014
+ "selected_model": null,
3015
+ "confirmed_model": null,
3016
+ "wall_ms": 1788.8,
3017
+ "state_bytes": 294,
3018
+ "metrics": {
3019
+ "input_tokens": 709,
3020
+ "output_tokens": 1
3021
+ }
3022
+ },
3023
+ {
3024
+ "round": 2,
3025
+ "case": "held-french-race",
3026
+ "split": "heldout",
3027
+ "expected": "opus",
3028
+ "actual": "opus",
3029
+ "classified_tier": "opus",
3030
+ "source": "ollama",
3031
+ "evaluator": "ollama",
3032
+ "selected_model": null,
3033
+ "confirmed_model": null,
3034
+ "wall_ms": 1055.79,
3035
+ "state_bytes": 510,
3036
+ "metrics": {
3037
+ "input_tokens": 739,
3038
+ "output_tokens": 1
3039
+ }
3040
+ },
3041
+ {
3042
+ "round": 2,
3043
+ "case": "held-h-link",
3044
+ "split": "heldout",
3045
+ "expected": "haiku",
3046
+ "actual": "haiku",
3047
+ "classified_tier": "haiku",
3048
+ "source": "ollama",
3049
+ "evaluator": "ollama",
3050
+ "selected_model": null,
3051
+ "confirmed_model": null,
3052
+ "wall_ms": 1249.52,
3053
+ "state_bytes": 386,
3054
+ "metrics": {
3055
+ "input_tokens": 711,
3056
+ "output_tokens": 1
3057
+ }
3058
+ },
3059
+ {
3060
+ "round": 2,
3061
+ "case": "held-h-translation",
3062
+ "split": "heldout",
3063
+ "expected": "haiku",
3064
+ "actual": "haiku",
3065
+ "classified_tier": "haiku",
3066
+ "source": "ollama",
3067
+ "evaluator": "ollama",
3068
+ "selected_model": null,
3069
+ "confirmed_model": null,
3070
+ "wall_ms": 1253.44,
3071
+ "state_bytes": 290,
3072
+ "metrics": {
3073
+ "input_tokens": 687,
3074
+ "output_tokens": 1
3075
+ }
3076
+ },
3077
+ {
3078
+ "round": 2,
3079
+ "case": "held-o-exactly-once",
3080
+ "split": "heldout",
3081
+ "expected": "opus",
3082
+ "actual": "opus",
3083
+ "classified_tier": "opus",
3084
+ "source": "ollama",
3085
+ "evaluator": "ollama",
3086
+ "selected_model": null,
3087
+ "confirmed_model": null,
3088
+ "wall_ms": 1144.69,
3089
+ "state_bytes": 676,
3090
+ "metrics": {
3091
+ "input_tokens": 753,
3092
+ "output_tokens": 1
3093
+ }
3094
+ },
3095
+ {
3096
+ "round": 2,
3097
+ "case": "held-s-query",
3098
+ "split": "heldout",
3099
+ "expected": "sonnet",
3100
+ "actual": "sonnet",
3101
+ "classified_tier": "sonnet",
3102
+ "source": "ollama",
3103
+ "evaluator": "ollama",
3104
+ "selected_model": null,
3105
+ "confirmed_model": null,
3106
+ "wall_ms": 995.03,
3107
+ "state_bytes": 522,
3108
+ "metrics": {
3109
+ "input_tokens": 725,
3110
+ "output_tokens": 1
3111
+ }
3112
+ },
3113
+ {
3114
+ "round": 2,
3115
+ "case": "held-tier-influence-difficult",
3116
+ "split": "heldout",
3117
+ "expected": "opus",
3118
+ "actual": "opus",
3119
+ "classified_tier": "opus",
3120
+ "source": "ollama",
3121
+ "evaluator": "ollama",
3122
+ "selected_model": null,
3123
+ "confirmed_model": null,
3124
+ "wall_ms": 1253.79,
3125
+ "state_bytes": 640,
3126
+ "metrics": {
3127
+ "input_tokens": 753,
3128
+ "output_tokens": 1
3129
+ }
3130
+ },
3131
+ {
3132
+ "round": 3,
3133
+ "case": "held-h-link",
3134
+ "split": "heldout",
3135
+ "expected": "haiku",
3136
+ "actual": "haiku",
3137
+ "classified_tier": "haiku",
3138
+ "source": "ollama",
3139
+ "evaluator": "ollama",
3140
+ "selected_model": null,
3141
+ "confirmed_model": null,
3142
+ "wall_ms": 244.7,
3143
+ "state_bytes": 386,
3144
+ "metrics": {
3145
+ "input_tokens": 711,
3146
+ "output_tokens": 1
3147
+ }
3148
+ },
3149
+ {
3150
+ "round": 3,
3151
+ "case": "held-s-tool-followup",
3152
+ "split": "heldout",
3153
+ "expected": "sonnet",
3154
+ "actual": "sonnet",
3155
+ "classified_tier": "sonnet",
3156
+ "source": "ollama",
3157
+ "evaluator": "ollama",
3158
+ "selected_model": null,
3159
+ "confirmed_model": null,
3160
+ "wall_ms": 414.84,
3161
+ "state_bytes": 508,
3162
+ "metrics": {
3163
+ "input_tokens": 753,
3164
+ "output_tokens": 1
3165
+ }
3166
+ },
3167
+ {
3168
+ "round": 3,
3169
+ "case": "held-s-null-bug",
3170
+ "split": "heldout",
3171
+ "expected": "sonnet",
3172
+ "actual": "sonnet",
3173
+ "classified_tier": "sonnet",
3174
+ "source": "ollama",
3175
+ "evaluator": "ollama",
3176
+ "selected_model": null,
3177
+ "confirmed_model": null,
3178
+ "wall_ms": 991.56,
3179
+ "state_bytes": 546,
3180
+ "metrics": {
3181
+ "input_tokens": 729,
3182
+ "output_tokens": 1
3183
+ }
3184
+ },
3185
+ {
3186
+ "round": 3,
3187
+ "case": "held-h-injection",
3188
+ "split": "heldout",
3189
+ "expected": "haiku",
3190
+ "actual": "haiku",
3191
+ "classified_tier": "haiku",
3192
+ "source": "ollama",
3193
+ "evaluator": "ollama",
3194
+ "selected_model": null,
3195
+ "confirmed_model": null,
3196
+ "wall_ms": 285.68,
3197
+ "state_bytes": 512,
3198
+ "metrics": {
3199
+ "input_tokens": 735,
3200
+ "output_tokens": 1
3201
+ }
3202
+ },
3203
+ {
3204
+ "round": 3,
3205
+ "case": "held-o-compiler",
3206
+ "split": "heldout",
3207
+ "expected": "opus",
3208
+ "actual": "opus",
3209
+ "classified_tier": "opus",
3210
+ "source": "ollama",
3211
+ "evaluator": "ollama",
3212
+ "selected_model": null,
3213
+ "confirmed_model": null,
3214
+ "wall_ms": 440.5,
3215
+ "state_bytes": 654,
3216
+ "metrics": {
3217
+ "input_tokens": 737,
3218
+ "output_tokens": 1
3219
+ }
3220
+ },
3221
+ {
3222
+ "round": 3,
3223
+ "case": "held-italian-literal",
3224
+ "split": "heldout",
3225
+ "expected": "haiku",
3226
+ "actual": "haiku",
3227
+ "classified_tier": "haiku",
3228
+ "source": "ollama",
3229
+ "evaluator": "ollama",
3230
+ "selected_model": null,
3231
+ "confirmed_model": null,
3232
+ "wall_ms": 245.05,
3233
+ "state_bytes": 252,
3234
+ "metrics": {
3235
+ "input_tokens": 683,
3236
+ "output_tokens": 1
3237
+ }
3238
+ },
3239
+ {
3240
+ "round": 3,
3241
+ "case": "held-h-reminder",
3242
+ "split": "heldout",
3243
+ "expected": "haiku",
3244
+ "actual": "haiku",
3245
+ "classified_tier": "haiku",
3246
+ "source": "ollama",
3247
+ "evaluator": "ollama",
3248
+ "selected_model": null,
3249
+ "confirmed_model": null,
3250
+ "wall_ms": 260.19,
3251
+ "state_bytes": 294,
3252
+ "metrics": {
3253
+ "input_tokens": 709,
3254
+ "output_tokens": 1
3255
+ }
3256
+ },
3257
+ {
3258
+ "round": 3,
3259
+ "case": "held-s-pagination",
3260
+ "split": "heldout",
3261
+ "expected": "sonnet",
3262
+ "actual": "sonnet",
3263
+ "classified_tier": "sonnet",
3264
+ "source": "ollama",
3265
+ "evaluator": "ollama",
3266
+ "selected_model": null,
3267
+ "confirmed_model": null,
3268
+ "wall_ms": 249.07,
3269
+ "state_bytes": 482,
3270
+ "metrics": {
3271
+ "input_tokens": 717,
3272
+ "output_tokens": 1
3273
+ }
3274
+ },
3275
+ {
3276
+ "round": 3,
3277
+ "case": "held-s-validation",
3278
+ "split": "heldout",
3279
+ "expected": "sonnet",
3280
+ "actual": "sonnet",
3281
+ "classified_tier": "sonnet",
3282
+ "source": "ollama",
3283
+ "evaluator": "ollama",
3284
+ "selected_model": null,
3285
+ "confirmed_model": null,
3286
+ "wall_ms": 264.6,
3287
+ "state_bytes": 484,
3288
+ "metrics": {
3289
+ "input_tokens": 715,
3290
+ "output_tokens": 1
3291
+ }
3292
+ },
3293
+ {
3294
+ "round": 3,
3295
+ "case": "held-o-tenant-cache",
3296
+ "split": "heldout",
3297
+ "expected": "opus",
3298
+ "actual": "opus",
3299
+ "classified_tier": "opus",
3300
+ "source": "ollama",
3301
+ "evaluator": "ollama",
3302
+ "selected_model": null,
3303
+ "confirmed_model": null,
3304
+ "wall_ms": 247.38,
3305
+ "state_bytes": 658,
3306
+ "metrics": {
3307
+ "input_tokens": 745,
3308
+ "output_tokens": 1
3309
+ }
3310
+ },
3311
+ {
3312
+ "round": 3,
3313
+ "case": "held-spanish-feature",
3314
+ "split": "heldout",
3315
+ "expected": "sonnet",
3316
+ "actual": "sonnet",
3317
+ "classified_tier": "sonnet",
3318
+ "source": "ollama",
3319
+ "evaluator": "ollama",
3320
+ "selected_model": null,
3321
+ "confirmed_model": null,
3322
+ "wall_ms": 266.72,
3323
+ "state_bytes": 380,
3324
+ "metrics": {
3325
+ "input_tokens": 711,
3326
+ "output_tokens": 1
3327
+ }
3328
+ },
3329
+ {
3330
+ "round": 3,
3331
+ "case": "held-s-unclear-polish",
3332
+ "split": "heldout",
3333
+ "expected": "sonnet",
3334
+ "actual": "sonnet",
3335
+ "classified_tier": "sonnet",
3336
+ "source": "ollama",
3337
+ "evaluator": "ollama",
3338
+ "selected_model": null,
3339
+ "confirmed_model": null,
3340
+ "wall_ms": 265.34,
3341
+ "state_bytes": 478,
3342
+ "metrics": {
3343
+ "input_tokens": 717,
3344
+ "output_tokens": 1
3345
+ }
3346
+ },
3347
+ {
3348
+ "round": 3,
3349
+ "case": "held-h-marker",
3350
+ "split": "heldout",
3351
+ "expected": "haiku",
3352
+ "actual": "haiku",
3353
+ "classified_tier": "haiku",
3354
+ "source": "ollama",
3355
+ "evaluator": "ollama",
3356
+ "selected_model": null,
3357
+ "confirmed_model": null,
3358
+ "wall_ms": 274.65,
3359
+ "state_bytes": 323,
3360
+ "metrics": {
3361
+ "input_tokens": 707,
3362
+ "output_tokens": 1
3363
+ }
3364
+ },
3365
+ {
3366
+ "round": 3,
3367
+ "case": "held-tier-influence-difficult",
3368
+ "split": "heldout",
3369
+ "expected": "opus",
3370
+ "actual": "opus",
3371
+ "classified_tier": "opus",
3372
+ "source": "ollama",
3373
+ "evaluator": "ollama",
3374
+ "selected_model": null,
3375
+ "confirmed_model": null,
3376
+ "wall_ms": 244.84,
3377
+ "state_bytes": 640,
3378
+ "metrics": {
3379
+ "input_tokens": 753,
3380
+ "output_tokens": 1
3381
+ }
3382
+ },
3383
+ {
3384
+ "round": 3,
3385
+ "case": "held-s-cli",
3386
+ "split": "heldout",
3387
+ "expected": "sonnet",
3388
+ "actual": "sonnet",
3389
+ "classified_tier": "sonnet",
3390
+ "source": "ollama",
3391
+ "evaluator": "ollama",
3392
+ "selected_model": null,
3393
+ "confirmed_model": null,
3394
+ "wall_ms": 245.98,
3395
+ "state_bytes": 480,
3396
+ "metrics": {
3397
+ "input_tokens": 725,
3398
+ "output_tokens": 1
3399
+ }
3400
+ },
3401
+ {
3402
+ "round": 3,
3403
+ "case": "held-tier-influence-mechanical",
3404
+ "split": "heldout",
3405
+ "expected": "haiku",
3406
+ "actual": "haiku",
3407
+ "classified_tier": "haiku",
3408
+ "source": "ollama",
3409
+ "evaluator": "ollama",
3410
+ "selected_model": null,
3411
+ "confirmed_model": null,
3412
+ "wall_ms": 249.53,
3413
+ "state_bytes": 432,
3414
+ "metrics": {
3415
+ "input_tokens": 733,
3416
+ "output_tokens": 1
3417
+ }
3418
+ },
3419
+ {
3420
+ "round": 3,
3421
+ "case": "held-o-exactly-once",
3422
+ "split": "heldout",
3423
+ "expected": "opus",
3424
+ "actual": "opus",
3425
+ "classified_tier": "opus",
3426
+ "source": "ollama",
3427
+ "evaluator": "ollama",
3428
+ "selected_model": null,
3429
+ "confirmed_model": null,
3430
+ "wall_ms": 255.21,
3431
+ "state_bytes": 676,
3432
+ "metrics": {
3433
+ "input_tokens": 753,
3434
+ "output_tokens": 1
3435
+ }
3436
+ },
3437
+ {
3438
+ "round": 3,
3439
+ "case": "held-french-race",
3440
+ "split": "heldout",
3441
+ "expected": "opus",
3442
+ "actual": "opus",
3443
+ "classified_tier": "opus",
3444
+ "source": "ollama",
3445
+ "evaluator": "ollama",
3446
+ "selected_model": null,
3447
+ "confirmed_model": null,
3448
+ "wall_ms": 245.88,
3449
+ "state_bytes": 510,
3450
+ "metrics": {
3451
+ "input_tokens": 739,
3452
+ "output_tokens": 1
3453
+ }
3454
+ },
3455
+ {
3456
+ "round": 3,
3457
+ "case": "held-h-summary",
3458
+ "split": "heldout",
3459
+ "expected": "haiku",
3460
+ "actual": "haiku",
3461
+ "classified_tier": "haiku",
3462
+ "source": "ollama",
3463
+ "evaluator": "ollama",
3464
+ "selected_model": null,
3465
+ "confirmed_model": null,
3466
+ "wall_ms": 247.64,
3467
+ "state_bytes": 406,
3468
+ "metrics": {
3469
+ "input_tokens": 715,
3470
+ "output_tokens": 1
3471
+ }
3472
+ },
3473
+ {
3474
+ "round": 3,
3475
+ "case": "held-s-refactor",
3476
+ "split": "heldout",
3477
+ "expected": "sonnet",
3478
+ "actual": "sonnet",
3479
+ "classified_tier": "sonnet",
3480
+ "source": "ollama",
3481
+ "evaluator": "ollama",
3482
+ "selected_model": null,
3483
+ "confirmed_model": null,
3484
+ "wall_ms": 244.6,
3485
+ "state_bytes": 464,
3486
+ "metrics": {
3487
+ "input_tokens": 711,
3488
+ "output_tokens": 1
3489
+ }
3490
+ },
3491
+ {
3492
+ "round": 3,
3493
+ "case": "held-short-difficult-followup",
3494
+ "split": "heldout",
3495
+ "expected": "opus",
3496
+ "actual": "opus",
3497
+ "classified_tier": "opus",
3498
+ "source": "ollama",
3499
+ "evaluator": "ollama",
3500
+ "selected_model": null,
3501
+ "confirmed_model": null,
3502
+ "wall_ms": 322.91,
3503
+ "state_bytes": 483,
3504
+ "metrics": {
3505
+ "input_tokens": 727,
3506
+ "output_tokens": 1
3507
+ }
3508
+ },
3509
+ {
3510
+ "round": 3,
3511
+ "case": "held-h-literal-opus",
3512
+ "split": "heldout",
3513
+ "expected": "haiku",
3514
+ "actual": "haiku",
3515
+ "classified_tier": "haiku",
3516
+ "source": "ollama",
3517
+ "evaluator": "ollama",
3518
+ "selected_model": null,
3519
+ "confirmed_model": null,
3520
+ "wall_ms": 303.6,
3521
+ "state_bytes": 312,
3522
+ "metrics": {
3523
+ "input_tokens": 693,
3524
+ "output_tokens": 1
3525
+ }
3526
+ },
3527
+ {
3528
+ "round": 3,
3529
+ "case": "held-o-injection",
3530
+ "split": "heldout",
3531
+ "expected": "opus",
3532
+ "actual": "haiku",
3533
+ "classified_tier": "haiku",
3534
+ "source": "ollama",
3535
+ "evaluator": "ollama",
3536
+ "selected_model": null,
3537
+ "confirmed_model": null,
3538
+ "wall_ms": 971.02,
3539
+ "state_bytes": 694,
3540
+ "metrics": {
3541
+ "input_tokens": 761,
3542
+ "output_tokens": 1
3543
+ }
3544
+ },
3545
+ {
3546
+ "round": 3,
3547
+ "case": "held-o-rolling-format",
3548
+ "split": "heldout",
3549
+ "expected": "opus",
3550
+ "actual": "opus",
3551
+ "classified_tier": "opus",
3552
+ "source": "ollama",
3553
+ "evaluator": "ollama",
3554
+ "selected_model": null,
3555
+ "confirmed_model": null,
3556
+ "wall_ms": 1985.69,
3557
+ "state_bytes": 562,
3558
+ "metrics": {
3559
+ "input_tokens": 735,
3560
+ "output_tokens": 1
3561
+ }
3562
+ },
3563
+ {
3564
+ "round": 3,
3565
+ "case": "held-s-query",
3566
+ "split": "heldout",
3567
+ "expected": "sonnet",
3568
+ "actual": "sonnet",
3569
+ "classified_tier": "sonnet",
3570
+ "source": "ollama",
3571
+ "evaluator": "ollama",
3572
+ "selected_model": null,
3573
+ "confirmed_model": null,
3574
+ "wall_ms": 254.54,
3575
+ "state_bytes": 522,
3576
+ "metrics": {
3577
+ "input_tokens": 725,
3578
+ "output_tokens": 1
3579
+ }
3580
+ },
3581
+ {
3582
+ "round": 3,
3583
+ "case": "held-h-arithmetic",
3584
+ "split": "heldout",
3585
+ "expected": "haiku",
3586
+ "actual": "haiku",
3587
+ "classified_tier": "haiku",
3588
+ "source": "ollama",
3589
+ "evaluator": "ollama",
3590
+ "selected_model": null,
3591
+ "confirmed_model": null,
3592
+ "wall_ms": 1430.07,
3593
+ "state_bytes": 298,
3594
+ "metrics": {
3595
+ "input_tokens": 705,
3596
+ "output_tokens": 1
3597
+ }
3598
+ },
3599
+ {
3600
+ "round": 3,
3601
+ "case": "held-o-snapshot",
3602
+ "split": "heldout",
3603
+ "expected": "opus",
3604
+ "actual": "opus",
3605
+ "classified_tier": "opus",
3606
+ "source": "ollama",
3607
+ "evaluator": "ollama",
3608
+ "selected_model": null,
3609
+ "confirmed_model": null,
3610
+ "wall_ms": 1148.13,
3611
+ "state_bytes": 764,
3612
+ "metrics": {
3613
+ "input_tokens": 763,
3614
+ "output_tokens": 1
3615
+ }
3616
+ },
3617
+ {
3618
+ "round": 3,
3619
+ "case": "held-o-crypto",
3620
+ "split": "heldout",
3621
+ "expected": "opus",
3622
+ "actual": "opus",
3623
+ "classified_tier": "opus",
3624
+ "source": "ollama",
3625
+ "evaluator": "ollama",
3626
+ "selected_model": null,
3627
+ "confirmed_model": null,
3628
+ "wall_ms": 968.42,
3629
+ "state_bytes": 654,
3630
+ "metrics": {
3631
+ "input_tokens": 741,
3632
+ "output_tokens": 1
3633
+ }
3634
+ },
3635
+ {
3636
+ "round": 3,
3637
+ "case": "held-o-fencing-followup",
3638
+ "split": "heldout",
3639
+ "expected": "opus",
3640
+ "actual": "opus",
3641
+ "classified_tier": "opus",
3642
+ "source": "ollama",
3643
+ "evaluator": "ollama",
3644
+ "selected_model": null,
3645
+ "confirmed_model": null,
3646
+ "wall_ms": 2166,
3647
+ "state_bytes": 780,
3648
+ "metrics": {
3649
+ "input_tokens": 791,
3650
+ "output_tokens": 1
3651
+ }
3652
+ },
3653
+ {
3654
+ "round": 3,
3655
+ "case": "held-h-translation",
3656
+ "split": "heldout",
3657
+ "expected": "haiku",
3658
+ "actual": "haiku",
3659
+ "classified_tier": "haiku",
3660
+ "source": "ollama",
3661
+ "evaluator": "ollama",
3662
+ "selected_model": null,
3663
+ "confirmed_model": null,
3664
+ "wall_ms": 508.66,
3665
+ "state_bytes": 290,
3666
+ "metrics": {
3667
+ "input_tokens": 687,
3668
+ "output_tokens": 1
3669
+ }
3670
+ }
3671
+ ],
3672
+ "summary": {
3673
+ "requests": 90,
3674
+ "valid": 87,
3675
+ "errors": 3,
3676
+ "exact_rubric_agreement": 0.9333333333333333,
3677
+ "under_routes": 3,
3678
+ "over_routes": 0,
3679
+ "opus_under_routes": 3,
3680
+ "wall_p50_ms": 1249.52,
3681
+ "wall_p95_ms": 13644.66,
3682
+ "successful_wall_p50_ms": 1148.13,
3683
+ "successful_wall_p95_ms": 11249.27,
3684
+ "confusion": {
3685
+ "haiku": {
3686
+ "haiku": 30,
3687
+ "sonnet": 0,
3688
+ "opus": 0,
3689
+ "error": 0
3690
+ },
3691
+ "sonnet": {
3692
+ "haiku": 0,
3693
+ "sonnet": 26,
3694
+ "opus": 0,
3695
+ "error": 1
3696
+ },
3697
+ "opus": {
3698
+ "haiku": 3,
3699
+ "sonnet": 0,
3700
+ "opus": 28,
3701
+ "error": 2
3702
+ }
3703
+ }
3704
+ },
3705
+ "acceptance": {
3706
+ "schema_version": 1,
3707
+ "policy": {
3708
+ "profile": "compatible",
3709
+ "minAgreement": 1,
3710
+ "maxUnderRouteRate": 0,
3711
+ "requiredTiers": [
3712
+ "haiku",
3713
+ "sonnet",
3714
+ "opus"
3715
+ ]
3716
+ },
3717
+ "gates": {
3718
+ "transport": {
3719
+ "passed": null,
3720
+ "reason": "Claude transport not exercised"
3721
+ },
3722
+ "evaluator": {
3723
+ "passed": false,
3724
+ "expected": "ollama",
3725
+ "evaluated_requests": 90,
3726
+ "successful_evaluations": 87,
3727
+ "fallbacks": 0
3728
+ },
3729
+ "rubric": {
3730
+ "passed": false,
3731
+ "agreement": 0.9333333333333333,
3732
+ "under_routes": 3,
3733
+ "under_route_rate": 0.03333333333333333,
3734
+ "min_agreement": 1,
3735
+ "max_under_route_rate": 0
3736
+ },
3737
+ "policy": {
3738
+ "passed": null,
3739
+ "reason": "Classifier-only benchmark; routing policy not exercised"
3740
+ },
3741
+ "coverage": {
3742
+ "passed": true,
3743
+ "required": [
3744
+ "haiku",
3745
+ "sonnet",
3746
+ "opus"
3747
+ ],
3748
+ "observed": [
3749
+ "haiku",
3750
+ "opus",
3751
+ "sonnet"
3752
+ ],
3753
+ "missing": []
3754
+ },
3755
+ "task": {
3756
+ "passed": null,
3757
+ "reason": "No Claude task completion measured"
3758
+ }
3759
+ },
3760
+ "passed": false,
3761
+ "requests": 90
3762
+ },
3763
+ "final_resident_memory": {
3764
+ "model_size_bytes": 6166423796,
3765
+ "model_vram_bytes": 6166423796,
3766
+ "context_length": 8194,
3767
+ "parameter_size": "9.0B",
3768
+ "quantization": "Q4_K_M",
3769
+ "runtime_process_rss_bytes": 6397952000,
3770
+ "system_free_bytes": 136183808
3771
+ },
3772
+ "max_budget_stress": {
3773
+ "rows": [
3774
+ {
3775
+ "case": "full-budget-0",
3776
+ "split": "performance_stress",
3777
+ "expected": "haiku",
3778
+ "actual": "haiku",
3779
+ "classified_tier": "haiku",
3780
+ "source": "ollama",
3781
+ "evaluator": "ollama",
3782
+ "selected_model": null,
3783
+ "confirmed_model": null,
3784
+ "wall_ms": 11740.65,
3785
+ "state_bytes": 3000,
3786
+ "metrics": {
3787
+ "input_tokens": 1775,
3788
+ "output_tokens": 1
3789
+ }
3790
+ },
3791
+ {
3792
+ "case": "full-budget-1",
3793
+ "split": "performance_stress",
3794
+ "expected": "haiku",
3795
+ "actual": "haiku",
3796
+ "classified_tier": "haiku",
3797
+ "source": "ollama",
3798
+ "evaluator": "ollama",
3799
+ "selected_model": null,
3800
+ "confirmed_model": null,
3801
+ "wall_ms": 11585.43,
3802
+ "state_bytes": 3000,
3803
+ "metrics": {
3804
+ "input_tokens": 1709,
3805
+ "output_tokens": 1
3806
+ }
3807
+ },
3808
+ {
3809
+ "case": "full-budget-2",
3810
+ "split": "performance_stress",
3811
+ "expected": "haiku",
3812
+ "actual": "haiku",
3813
+ "classified_tier": "haiku",
3814
+ "source": "ollama",
3815
+ "evaluator": "ollama",
3816
+ "selected_model": null,
3817
+ "confirmed_model": null,
3818
+ "wall_ms": 12362.81,
3819
+ "state_bytes": 3000,
3820
+ "metrics": {
3821
+ "input_tokens": 1731,
3822
+ "output_tokens": 1
3823
+ }
3824
+ },
3825
+ {
3826
+ "case": "full-budget-3",
3827
+ "split": "performance_stress",
3828
+ "expected": "haiku",
3829
+ "actual": "haiku",
3830
+ "classified_tier": "haiku",
3831
+ "source": "ollama",
3832
+ "evaluator": "ollama",
3833
+ "selected_model": null,
3834
+ "confirmed_model": null,
3835
+ "wall_ms": 12368.57,
3836
+ "state_bytes": 3000,
3837
+ "metrics": {
3838
+ "input_tokens": 1709,
3839
+ "output_tokens": 1
3840
+ }
3841
+ },
3842
+ {
3843
+ "case": "full-budget-4",
3844
+ "split": "performance_stress",
3845
+ "expected": "haiku",
3846
+ "actual": "haiku",
3847
+ "classified_tier": "haiku",
3848
+ "source": "ollama",
3849
+ "evaluator": "ollama",
3850
+ "selected_model": null,
3851
+ "confirmed_model": null,
3852
+ "wall_ms": 12803.94,
3853
+ "state_bytes": 3000,
3854
+ "metrics": {
3855
+ "input_tokens": 1731,
3856
+ "output_tokens": 1
3857
+ }
3858
+ },
3859
+ {
3860
+ "case": "full-budget-5",
3861
+ "split": "performance_stress",
3862
+ "expected": "haiku",
3863
+ "actual": "haiku",
3864
+ "classified_tier": "haiku",
3865
+ "source": "ollama",
3866
+ "evaluator": "ollama",
3867
+ "selected_model": null,
3868
+ "confirmed_model": null,
3869
+ "wall_ms": 13062.09,
3870
+ "state_bytes": 3000,
3871
+ "metrics": {
3872
+ "input_tokens": 1775,
3873
+ "output_tokens": 1
3874
+ }
3875
+ },
3876
+ {
3877
+ "case": "full-budget-6",
3878
+ "split": "performance_stress",
3879
+ "expected": "haiku",
3880
+ "actual": "haiku",
3881
+ "classified_tier": "haiku",
3882
+ "source": "ollama",
3883
+ "evaluator": "ollama",
3884
+ "selected_model": null,
3885
+ "confirmed_model": null,
3886
+ "wall_ms": 13925.43,
3887
+ "state_bytes": 3000,
3888
+ "metrics": {
3889
+ "input_tokens": 1797,
3890
+ "output_tokens": 1
3891
+ }
3892
+ },
3893
+ {
3894
+ "case": "full-budget-7",
3895
+ "split": "performance_stress",
3896
+ "expected": "haiku",
3897
+ "actual": "haiku",
3898
+ "classified_tier": "haiku",
3899
+ "source": "ollama",
3900
+ "evaluator": "ollama",
3901
+ "selected_model": null,
3902
+ "confirmed_model": null,
3903
+ "wall_ms": 13944.87,
3904
+ "state_bytes": 3000,
3905
+ "metrics": {
3906
+ "input_tokens": 1775,
3907
+ "output_tokens": 1
3908
+ }
3909
+ }
3910
+ ],
3911
+ "summary": {
3912
+ "requests": 8,
3913
+ "valid": 8,
3914
+ "errors": 0,
3915
+ "exact_rubric_agreement": 1,
3916
+ "under_routes": 0,
3917
+ "over_routes": 0,
3918
+ "opus_under_routes": 0,
3919
+ "wall_p50_ms": 12368.57,
3920
+ "wall_p95_ms": 13944.87,
3921
+ "successful_wall_p50_ms": 12368.57,
3922
+ "successful_wall_p95_ms": 13944.87,
3923
+ "confusion": {
3924
+ "haiku": {
3925
+ "haiku": 8,
3926
+ "sonnet": 0,
3927
+ "opus": 0,
3928
+ "error": 0
3929
+ },
3930
+ "sonnet": {
3931
+ "haiku": 0,
3932
+ "sonnet": 0,
3933
+ "opus": 0,
3934
+ "error": 0
3935
+ },
3936
+ "opus": {
3937
+ "haiku": 0,
3938
+ "sonnet": 0,
3939
+ "opus": 0,
3940
+ "error": 0
3941
+ }
3942
+ }
3943
+ },
3944
+ "acceptance": {
3945
+ "schema_version": 1,
3946
+ "policy": {
3947
+ "profile": "compatible",
3948
+ "minAgreement": 1,
3949
+ "maxUnderRouteRate": 0,
3950
+ "requiredTiers": [
3951
+ "haiku"
3952
+ ]
3953
+ },
3954
+ "gates": {
3955
+ "transport": {
3956
+ "passed": null,
3957
+ "reason": "Claude transport not exercised"
3958
+ },
3959
+ "evaluator": {
3960
+ "passed": true,
3961
+ "expected": "ollama",
3962
+ "evaluated_requests": 8,
3963
+ "successful_evaluations": 8,
3964
+ "fallbacks": 0
3965
+ },
3966
+ "rubric": {
3967
+ "passed": true,
3968
+ "agreement": 1,
3969
+ "under_routes": 0,
3970
+ "under_route_rate": 0,
3971
+ "min_agreement": 1,
3972
+ "max_under_route_rate": 0
3973
+ },
3974
+ "policy": {
3975
+ "passed": null,
3976
+ "reason": "Classifier-only benchmark; routing policy not exercised"
3977
+ },
3978
+ "coverage": {
3979
+ "passed": true,
3980
+ "required": [
3981
+ "haiku"
3982
+ ],
3983
+ "observed": [
3984
+ "haiku"
3985
+ ],
3986
+ "missing": []
3987
+ },
3988
+ "task": {
3989
+ "passed": null,
3990
+ "reason": "No Claude task completion measured"
3991
+ }
3992
+ },
3993
+ "passed": true,
3994
+ "requests": 8
3995
+ }
3996
+ }
3997
+ }
3998
+ ],
3999
+ "complete": true,
4000
+ "passed": false,
4001
+ "raw_report_sha256": "f640f828520269c2b6bd90bf8c1676ffa962f2851c82a02ba25816cf776a783b"
4002
+ }