@velum-labs/routekit-eval-setup 1.0.0

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 (56) hide show
  1. package/LICENSE +201 -0
  2. package/README.md +23 -0
  3. package/dist/effect-api.d.ts +12 -0
  4. package/dist/effect-api.js +9 -0
  5. package/dist/errors.d.ts +91 -0
  6. package/dist/errors.js +46 -0
  7. package/dist/host-metadata.d.ts +32 -0
  8. package/dist/host-metadata.js +46 -0
  9. package/dist/index.d.ts +20 -0
  10. package/dist/index.js +13 -0
  11. package/dist/inspection.d.ts +24 -0
  12. package/dist/inspection.js +261 -0
  13. package/dist/model-selection.d.ts +6 -0
  14. package/dist/model-selection.js +37 -0
  15. package/dist/ori-authoring.d.ts +16 -0
  16. package/dist/ori-authoring.js +17 -0
  17. package/dist/ori-result.d.ts +45 -0
  18. package/dist/ori-result.js +1 -0
  19. package/dist/project-artifacts.d.ts +31 -0
  20. package/dist/project-artifacts.js +353 -0
  21. package/dist/project-authoring.d.ts +68 -0
  22. package/dist/project-authoring.js +431 -0
  23. package/dist/project-contracts.d.ts +1197 -0
  24. package/dist/project-contracts.js +396 -0
  25. package/dist/project-store.d.ts +13 -0
  26. package/dist/project-store.js +53 -0
  27. package/dist/project-workflow.d.ts +33 -0
  28. package/dist/project-workflow.js +904 -0
  29. package/dist/questions.d.ts +7 -0
  30. package/dist/questions.js +67 -0
  31. package/dist/runner.d.ts +8 -0
  32. package/dist/runner.js +16 -0
  33. package/dist/service.d.ts +24 -0
  34. package/dist/service.js +279 -0
  35. package/dist/state-store.d.ts +21 -0
  36. package/dist/state-store.js +86 -0
  37. package/dist/test/inspection.test.d.ts +1 -0
  38. package/dist/test/inspection.test.js +68 -0
  39. package/dist/test/model-selection.test.d.ts +1 -0
  40. package/dist/test/model-selection.test.js +15 -0
  41. package/dist/test/project-authoring.test.d.ts +1 -0
  42. package/dist/test/project-authoring.test.js +67 -0
  43. package/dist/test/project-workflow.test.d.ts +1 -0
  44. package/dist/test/project-workflow.test.js +516 -0
  45. package/dist/test/questions.test.d.ts +1 -0
  46. package/dist/test/questions.test.js +48 -0
  47. package/dist/test/skill.test.d.ts +1 -0
  48. package/dist/test/skill.test.js +31 -0
  49. package/dist/test/state-store.test.d.ts +1 -0
  50. package/dist/test/state-store.test.js +30 -0
  51. package/dist/test/workflow.test.d.ts +1 -0
  52. package/dist/test/workflow.test.js +167 -0
  53. package/dist/types.d.ts +77 -0
  54. package/dist/types.js +1 -0
  55. package/package.json +52 -0
  56. package/skills/setup-eval-routing/SKILL.md +149 -0
@@ -0,0 +1,1197 @@
1
+ import { type EvalComparisonResult as EvalComparisonResultType } from "@velum-labs/routekit-eval-contracts";
2
+ import { Schema } from "effect";
3
+ export declare const EVAL_PROJECT_VERSION: 1;
4
+ export declare const EvalProjectQuestion: Schema.Struct<{
5
+ readonly id: Schema.Literals<readonly ["workload-description", "candidate-models", "classifier-model", "author-model", "judge-model", "routing-objective", "maximum-unknown-weight", "routing-constraints"]>;
6
+ readonly prompt: Schema.String;
7
+ readonly options: Schema.$Array<Schema.String>;
8
+ }>;
9
+ export type EvalProjectQuestion = typeof EvalProjectQuestion.Type;
10
+ export declare const EvalProjectSetupProgress: Schema.Union<readonly [Schema.TaggedStruct<"WorkloadDescriptionRequired", {}>, Schema.TaggedStruct<"CandidateModelsRequired", {
11
+ readonly workloadDescription: Schema.String;
12
+ }>, Schema.TaggedStruct<"ClassifierModelRequired", {
13
+ readonly workloadDescription: Schema.String;
14
+ readonly candidateModels: Schema.$Array<Schema.String>;
15
+ }>, Schema.TaggedStruct<"AuthorModelRequired", {
16
+ readonly workloadDescription: Schema.String;
17
+ readonly candidateModels: Schema.$Array<Schema.String>;
18
+ readonly classifierModel: Schema.String;
19
+ }>, Schema.TaggedStruct<"JudgeModelRequired", {
20
+ readonly workloadDescription: Schema.String;
21
+ readonly candidateModels: Schema.$Array<Schema.String>;
22
+ readonly classifierModel: Schema.String;
23
+ readonly authorModel: Schema.String;
24
+ }>, Schema.TaggedStruct<"RoutingObjectiveRequired", {
25
+ readonly workloadDescription: Schema.String;
26
+ readonly candidateModels: Schema.$Array<Schema.String>;
27
+ readonly classifierModel: Schema.String;
28
+ readonly authorModel: Schema.String;
29
+ readonly judgeModel: Schema.String;
30
+ }>, Schema.TaggedStruct<"MaximumUnknownWeightRequired", {
31
+ readonly workloadDescription: Schema.String;
32
+ readonly candidateModels: Schema.$Array<Schema.String>;
33
+ readonly classifierModel: Schema.String;
34
+ readonly authorModel: Schema.String;
35
+ readonly judgeModel: Schema.String;
36
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
37
+ readonly kind: Schema.Literal<"highest-quality">;
38
+ }>, Schema.Struct<{
39
+ readonly kind: Schema.Literal<"lowest-cost">;
40
+ readonly minimumQuality: Schema.Finite;
41
+ }>, Schema.Struct<{
42
+ readonly kind: Schema.Literal<"lowest-latency">;
43
+ readonly minimumQuality: Schema.Finite;
44
+ }>, Schema.Struct<{
45
+ readonly kind: Schema.Literal<"balanced">;
46
+ readonly minimumQuality: Schema.Finite;
47
+ readonly weights: Schema.Struct<{
48
+ readonly quality: Schema.Finite;
49
+ readonly cost: Schema.Finite;
50
+ readonly latency: Schema.Finite;
51
+ }>;
52
+ }>, Schema.Struct<{
53
+ readonly kind: Schema.Literal<"pareto">;
54
+ readonly minimumQuality: Schema.Finite;
55
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
56
+ }>]>;
57
+ }>, Schema.TaggedStruct<"RoutingConstraintsRequired", {
58
+ readonly workloadDescription: Schema.String;
59
+ readonly candidateModels: Schema.$Array<Schema.String>;
60
+ readonly classifierModel: Schema.String;
61
+ readonly authorModel: Schema.String;
62
+ readonly judgeModel: Schema.String;
63
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
64
+ readonly kind: Schema.Literal<"highest-quality">;
65
+ }>, Schema.Struct<{
66
+ readonly kind: Schema.Literal<"lowest-cost">;
67
+ readonly minimumQuality: Schema.Finite;
68
+ }>, Schema.Struct<{
69
+ readonly kind: Schema.Literal<"lowest-latency">;
70
+ readonly minimumQuality: Schema.Finite;
71
+ }>, Schema.Struct<{
72
+ readonly kind: Schema.Literal<"balanced">;
73
+ readonly minimumQuality: Schema.Finite;
74
+ readonly weights: Schema.Struct<{
75
+ readonly quality: Schema.Finite;
76
+ readonly cost: Schema.Finite;
77
+ readonly latency: Schema.Finite;
78
+ }>;
79
+ }>, Schema.Struct<{
80
+ readonly kind: Schema.Literal<"pareto">;
81
+ readonly minimumQuality: Schema.Finite;
82
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
83
+ }>]>;
84
+ readonly maximumUnknownWeight: Schema.Finite;
85
+ }>]>;
86
+ export type EvalProjectSetupProgress = typeof EvalProjectSetupProgress.Type;
87
+ export declare const EvalProjectConfiguration: Schema.Struct<{
88
+ readonly workloadDescription: Schema.String;
89
+ readonly candidateModels: Schema.$Array<Schema.String>;
90
+ readonly classifierModel: Schema.String;
91
+ readonly authorModel: Schema.String;
92
+ readonly judgeModel: Schema.String;
93
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
94
+ readonly kind: Schema.Literal<"highest-quality">;
95
+ }>, Schema.Struct<{
96
+ readonly kind: Schema.Literal<"lowest-cost">;
97
+ readonly minimumQuality: Schema.Finite;
98
+ }>, Schema.Struct<{
99
+ readonly kind: Schema.Literal<"lowest-latency">;
100
+ readonly minimumQuality: Schema.Finite;
101
+ }>, Schema.Struct<{
102
+ readonly kind: Schema.Literal<"balanced">;
103
+ readonly minimumQuality: Schema.Finite;
104
+ readonly weights: Schema.Struct<{
105
+ readonly quality: Schema.Finite;
106
+ readonly cost: Schema.Finite;
107
+ readonly latency: Schema.Finite;
108
+ }>;
109
+ }>, Schema.Struct<{
110
+ readonly kind: Schema.Literal<"pareto">;
111
+ readonly minimumQuality: Schema.Finite;
112
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
113
+ }>]>;
114
+ readonly maximumUnknownWeight: Schema.Finite;
115
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
116
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
117
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
118
+ }>>;
119
+ }>;
120
+ export type EvalProjectConfiguration = typeof EvalProjectConfiguration.Type;
121
+ export declare const EvalProjectState: Schema.Union<readonly [Schema.Struct<{
122
+ readonly stage: Schema.Literal<"setup-required">;
123
+ readonly progress: Schema.Union<readonly [Schema.TaggedStruct<"WorkloadDescriptionRequired", {}>, Schema.TaggedStruct<"CandidateModelsRequired", {
124
+ readonly workloadDescription: Schema.String;
125
+ }>, Schema.TaggedStruct<"ClassifierModelRequired", {
126
+ readonly workloadDescription: Schema.String;
127
+ readonly candidateModels: Schema.$Array<Schema.String>;
128
+ }>, Schema.TaggedStruct<"AuthorModelRequired", {
129
+ readonly workloadDescription: Schema.String;
130
+ readonly candidateModels: Schema.$Array<Schema.String>;
131
+ readonly classifierModel: Schema.String;
132
+ }>, Schema.TaggedStruct<"JudgeModelRequired", {
133
+ readonly workloadDescription: Schema.String;
134
+ readonly candidateModels: Schema.$Array<Schema.String>;
135
+ readonly classifierModel: Schema.String;
136
+ readonly authorModel: Schema.String;
137
+ }>, Schema.TaggedStruct<"RoutingObjectiveRequired", {
138
+ readonly workloadDescription: Schema.String;
139
+ readonly candidateModels: Schema.$Array<Schema.String>;
140
+ readonly classifierModel: Schema.String;
141
+ readonly authorModel: Schema.String;
142
+ readonly judgeModel: Schema.String;
143
+ }>, Schema.TaggedStruct<"MaximumUnknownWeightRequired", {
144
+ readonly workloadDescription: Schema.String;
145
+ readonly candidateModels: Schema.$Array<Schema.String>;
146
+ readonly classifierModel: Schema.String;
147
+ readonly authorModel: Schema.String;
148
+ readonly judgeModel: Schema.String;
149
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
150
+ readonly kind: Schema.Literal<"highest-quality">;
151
+ }>, Schema.Struct<{
152
+ readonly kind: Schema.Literal<"lowest-cost">;
153
+ readonly minimumQuality: Schema.Finite;
154
+ }>, Schema.Struct<{
155
+ readonly kind: Schema.Literal<"lowest-latency">;
156
+ readonly minimumQuality: Schema.Finite;
157
+ }>, Schema.Struct<{
158
+ readonly kind: Schema.Literal<"balanced">;
159
+ readonly minimumQuality: Schema.Finite;
160
+ readonly weights: Schema.Struct<{
161
+ readonly quality: Schema.Finite;
162
+ readonly cost: Schema.Finite;
163
+ readonly latency: Schema.Finite;
164
+ }>;
165
+ }>, Schema.Struct<{
166
+ readonly kind: Schema.Literal<"pareto">;
167
+ readonly minimumQuality: Schema.Finite;
168
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
169
+ }>]>;
170
+ }>, Schema.TaggedStruct<"RoutingConstraintsRequired", {
171
+ readonly workloadDescription: Schema.String;
172
+ readonly candidateModels: Schema.$Array<Schema.String>;
173
+ readonly classifierModel: Schema.String;
174
+ readonly authorModel: Schema.String;
175
+ readonly judgeModel: Schema.String;
176
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
177
+ readonly kind: Schema.Literal<"highest-quality">;
178
+ }>, Schema.Struct<{
179
+ readonly kind: Schema.Literal<"lowest-cost">;
180
+ readonly minimumQuality: Schema.Finite;
181
+ }>, Schema.Struct<{
182
+ readonly kind: Schema.Literal<"lowest-latency">;
183
+ readonly minimumQuality: Schema.Finite;
184
+ }>, Schema.Struct<{
185
+ readonly kind: Schema.Literal<"balanced">;
186
+ readonly minimumQuality: Schema.Finite;
187
+ readonly weights: Schema.Struct<{
188
+ readonly quality: Schema.Finite;
189
+ readonly cost: Schema.Finite;
190
+ readonly latency: Schema.Finite;
191
+ }>;
192
+ }>, Schema.Struct<{
193
+ readonly kind: Schema.Literal<"pareto">;
194
+ readonly minimumQuality: Schema.Finite;
195
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
196
+ }>]>;
197
+ readonly maximumUnknownWeight: Schema.Finite;
198
+ }>]>;
199
+ readonly version: Schema.Literal<1>;
200
+ readonly projectId: Schema.String;
201
+ readonly revision: Schema.Finite;
202
+ readonly createdAt: Schema.String;
203
+ readonly updatedAt: Schema.String;
204
+ readonly sourceInventory: Schema.$Array<Schema.String>;
205
+ }>, Schema.Struct<{
206
+ readonly stage: Schema.Literal<"dimensions-review">;
207
+ readonly configuration: Schema.Struct<{
208
+ readonly workloadDescription: Schema.String;
209
+ readonly candidateModels: Schema.$Array<Schema.String>;
210
+ readonly classifierModel: Schema.String;
211
+ readonly authorModel: Schema.String;
212
+ readonly judgeModel: Schema.String;
213
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
214
+ readonly kind: Schema.Literal<"highest-quality">;
215
+ }>, Schema.Struct<{
216
+ readonly kind: Schema.Literal<"lowest-cost">;
217
+ readonly minimumQuality: Schema.Finite;
218
+ }>, Schema.Struct<{
219
+ readonly kind: Schema.Literal<"lowest-latency">;
220
+ readonly minimumQuality: Schema.Finite;
221
+ }>, Schema.Struct<{
222
+ readonly kind: Schema.Literal<"balanced">;
223
+ readonly minimumQuality: Schema.Finite;
224
+ readonly weights: Schema.Struct<{
225
+ readonly quality: Schema.Finite;
226
+ readonly cost: Schema.Finite;
227
+ readonly latency: Schema.Finite;
228
+ }>;
229
+ }>, Schema.Struct<{
230
+ readonly kind: Schema.Literal<"pareto">;
231
+ readonly minimumQuality: Schema.Finite;
232
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
233
+ }>]>;
234
+ readonly maximumUnknownWeight: Schema.Finite;
235
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
236
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
237
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
238
+ }>>;
239
+ }>;
240
+ readonly version: Schema.Literal<1>;
241
+ readonly projectId: Schema.String;
242
+ readonly revision: Schema.Finite;
243
+ readonly createdAt: Schema.String;
244
+ readonly updatedAt: Schema.String;
245
+ readonly sourceInventory: Schema.$Array<Schema.String>;
246
+ }>, Schema.Struct<{
247
+ readonly stage: Schema.Literal<"evaluations-review">;
248
+ readonly configuration: Schema.Struct<{
249
+ readonly workloadDescription: Schema.String;
250
+ readonly candidateModels: Schema.$Array<Schema.String>;
251
+ readonly classifierModel: Schema.String;
252
+ readonly authorModel: Schema.String;
253
+ readonly judgeModel: Schema.String;
254
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
255
+ readonly kind: Schema.Literal<"highest-quality">;
256
+ }>, Schema.Struct<{
257
+ readonly kind: Schema.Literal<"lowest-cost">;
258
+ readonly minimumQuality: Schema.Finite;
259
+ }>, Schema.Struct<{
260
+ readonly kind: Schema.Literal<"lowest-latency">;
261
+ readonly minimumQuality: Schema.Finite;
262
+ }>, Schema.Struct<{
263
+ readonly kind: Schema.Literal<"balanced">;
264
+ readonly minimumQuality: Schema.Finite;
265
+ readonly weights: Schema.Struct<{
266
+ readonly quality: Schema.Finite;
267
+ readonly cost: Schema.Finite;
268
+ readonly latency: Schema.Finite;
269
+ }>;
270
+ }>, Schema.Struct<{
271
+ readonly kind: Schema.Literal<"pareto">;
272
+ readonly minimumQuality: Schema.Finite;
273
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
274
+ }>]>;
275
+ readonly maximumUnknownWeight: Schema.Finite;
276
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
277
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
278
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
279
+ }>>;
280
+ }>;
281
+ readonly basisDigest: Schema.String;
282
+ readonly version: Schema.Literal<1>;
283
+ readonly projectId: Schema.String;
284
+ readonly revision: Schema.Finite;
285
+ readonly createdAt: Schema.String;
286
+ readonly updatedAt: Schema.String;
287
+ readonly sourceInventory: Schema.$Array<Schema.String>;
288
+ }>, Schema.Struct<{
289
+ readonly stage: Schema.Literal<"ready">;
290
+ readonly configuration: Schema.Struct<{
291
+ readonly workloadDescription: Schema.String;
292
+ readonly candidateModels: Schema.$Array<Schema.String>;
293
+ readonly classifierModel: Schema.String;
294
+ readonly authorModel: Schema.String;
295
+ readonly judgeModel: Schema.String;
296
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
297
+ readonly kind: Schema.Literal<"highest-quality">;
298
+ }>, Schema.Struct<{
299
+ readonly kind: Schema.Literal<"lowest-cost">;
300
+ readonly minimumQuality: Schema.Finite;
301
+ }>, Schema.Struct<{
302
+ readonly kind: Schema.Literal<"lowest-latency">;
303
+ readonly minimumQuality: Schema.Finite;
304
+ }>, Schema.Struct<{
305
+ readonly kind: Schema.Literal<"balanced">;
306
+ readonly minimumQuality: Schema.Finite;
307
+ readonly weights: Schema.Struct<{
308
+ readonly quality: Schema.Finite;
309
+ readonly cost: Schema.Finite;
310
+ readonly latency: Schema.Finite;
311
+ }>;
312
+ }>, Schema.Struct<{
313
+ readonly kind: Schema.Literal<"pareto">;
314
+ readonly minimumQuality: Schema.Finite;
315
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
316
+ }>]>;
317
+ readonly maximumUnknownWeight: Schema.Finite;
318
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
319
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
320
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
321
+ }>>;
322
+ }>;
323
+ readonly basisDigest: Schema.String;
324
+ readonly evaluationDigest: Schema.String;
325
+ readonly version: Schema.Literal<1>;
326
+ readonly projectId: Schema.String;
327
+ readonly revision: Schema.Finite;
328
+ readonly createdAt: Schema.String;
329
+ readonly updatedAt: Schema.String;
330
+ readonly sourceInventory: Schema.$Array<Schema.String>;
331
+ }>, Schema.Struct<{
332
+ readonly stage: Schema.Literal<"running">;
333
+ readonly configuration: Schema.Struct<{
334
+ readonly workloadDescription: Schema.String;
335
+ readonly candidateModels: Schema.$Array<Schema.String>;
336
+ readonly classifierModel: Schema.String;
337
+ readonly authorModel: Schema.String;
338
+ readonly judgeModel: Schema.String;
339
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
340
+ readonly kind: Schema.Literal<"highest-quality">;
341
+ }>, Schema.Struct<{
342
+ readonly kind: Schema.Literal<"lowest-cost">;
343
+ readonly minimumQuality: Schema.Finite;
344
+ }>, Schema.Struct<{
345
+ readonly kind: Schema.Literal<"lowest-latency">;
346
+ readonly minimumQuality: Schema.Finite;
347
+ }>, Schema.Struct<{
348
+ readonly kind: Schema.Literal<"balanced">;
349
+ readonly minimumQuality: Schema.Finite;
350
+ readonly weights: Schema.Struct<{
351
+ readonly quality: Schema.Finite;
352
+ readonly cost: Schema.Finite;
353
+ readonly latency: Schema.Finite;
354
+ }>;
355
+ }>, Schema.Struct<{
356
+ readonly kind: Schema.Literal<"pareto">;
357
+ readonly minimumQuality: Schema.Finite;
358
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
359
+ }>]>;
360
+ readonly maximumUnknownWeight: Schema.Finite;
361
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
362
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
363
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
364
+ }>>;
365
+ }>;
366
+ readonly basisDigest: Schema.String;
367
+ readonly evaluationDigest: Schema.String;
368
+ readonly planId: Schema.String;
369
+ readonly runId: Schema.String;
370
+ readonly version: Schema.Literal<1>;
371
+ readonly projectId: Schema.String;
372
+ readonly revision: Schema.Finite;
373
+ readonly createdAt: Schema.String;
374
+ readonly updatedAt: Schema.String;
375
+ readonly sourceInventory: Schema.$Array<Schema.String>;
376
+ }>, Schema.Struct<{
377
+ readonly stage: Schema.Literal<"qualified">;
378
+ readonly configuration: Schema.Struct<{
379
+ readonly workloadDescription: Schema.String;
380
+ readonly candidateModels: Schema.$Array<Schema.String>;
381
+ readonly classifierModel: Schema.String;
382
+ readonly authorModel: Schema.String;
383
+ readonly judgeModel: Schema.String;
384
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
385
+ readonly kind: Schema.Literal<"highest-quality">;
386
+ }>, Schema.Struct<{
387
+ readonly kind: Schema.Literal<"lowest-cost">;
388
+ readonly minimumQuality: Schema.Finite;
389
+ }>, Schema.Struct<{
390
+ readonly kind: Schema.Literal<"lowest-latency">;
391
+ readonly minimumQuality: Schema.Finite;
392
+ }>, Schema.Struct<{
393
+ readonly kind: Schema.Literal<"balanced">;
394
+ readonly minimumQuality: Schema.Finite;
395
+ readonly weights: Schema.Struct<{
396
+ readonly quality: Schema.Finite;
397
+ readonly cost: Schema.Finite;
398
+ readonly latency: Schema.Finite;
399
+ }>;
400
+ }>, Schema.Struct<{
401
+ readonly kind: Schema.Literal<"pareto">;
402
+ readonly minimumQuality: Schema.Finite;
403
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
404
+ }>]>;
405
+ readonly maximumUnknownWeight: Schema.Finite;
406
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
407
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
408
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
409
+ }>>;
410
+ }>;
411
+ readonly basisDigest: Schema.String;
412
+ readonly evaluationDigest: Schema.String;
413
+ readonly runId: Schema.String;
414
+ readonly reportPath: Schema.String;
415
+ readonly version: Schema.Literal<1>;
416
+ readonly projectId: Schema.String;
417
+ readonly revision: Schema.Finite;
418
+ readonly createdAt: Schema.String;
419
+ readonly updatedAt: Schema.String;
420
+ readonly sourceInventory: Schema.$Array<Schema.String>;
421
+ }>, Schema.Struct<{
422
+ readonly stage: Schema.Literal<"activated">;
423
+ readonly configuration: Schema.Struct<{
424
+ readonly workloadDescription: Schema.String;
425
+ readonly candidateModels: Schema.$Array<Schema.String>;
426
+ readonly classifierModel: Schema.String;
427
+ readonly authorModel: Schema.String;
428
+ readonly judgeModel: Schema.String;
429
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
430
+ readonly kind: Schema.Literal<"highest-quality">;
431
+ }>, Schema.Struct<{
432
+ readonly kind: Schema.Literal<"lowest-cost">;
433
+ readonly minimumQuality: Schema.Finite;
434
+ }>, Schema.Struct<{
435
+ readonly kind: Schema.Literal<"lowest-latency">;
436
+ readonly minimumQuality: Schema.Finite;
437
+ }>, Schema.Struct<{
438
+ readonly kind: Schema.Literal<"balanced">;
439
+ readonly minimumQuality: Schema.Finite;
440
+ readonly weights: Schema.Struct<{
441
+ readonly quality: Schema.Finite;
442
+ readonly cost: Schema.Finite;
443
+ readonly latency: Schema.Finite;
444
+ }>;
445
+ }>, Schema.Struct<{
446
+ readonly kind: Schema.Literal<"pareto">;
447
+ readonly minimumQuality: Schema.Finite;
448
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
449
+ }>]>;
450
+ readonly maximumUnknownWeight: Schema.Finite;
451
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
452
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
453
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
454
+ }>>;
455
+ }>;
456
+ readonly basisDigest: Schema.String;
457
+ readonly evaluationDigest: Schema.String;
458
+ readonly runId: Schema.String;
459
+ readonly reportPath: Schema.String;
460
+ readonly evidenceDigest: Schema.String;
461
+ readonly targetIdentity: Schema.String;
462
+ readonly version: Schema.Literal<1>;
463
+ readonly projectId: Schema.String;
464
+ readonly revision: Schema.Finite;
465
+ readonly createdAt: Schema.String;
466
+ readonly updatedAt: Schema.String;
467
+ readonly sourceInventory: Schema.$Array<Schema.String>;
468
+ }>]>;
469
+ export type EvalProjectState = typeof EvalProjectState.Type;
470
+ export declare const EvalDimensionCase: Schema.Struct<{
471
+ readonly id: Schema.String;
472
+ readonly prompt: Schema.String;
473
+ readonly context: Schema.optionalKey<Schema.String>;
474
+ readonly rubric: Schema.String;
475
+ }>;
476
+ export type EvalDimensionCase = typeof EvalDimensionCase.Type;
477
+ export declare const EvalDimensionSuite: Schema.Struct<{
478
+ readonly version: Schema.Literal<1>;
479
+ readonly dimensionId: Schema.String;
480
+ readonly maximumOutputTokens: Schema.Finite;
481
+ readonly cases: Schema.$Array<Schema.Struct<{
482
+ readonly id: Schema.String;
483
+ readonly prompt: Schema.String;
484
+ readonly context: Schema.optionalKey<Schema.String>;
485
+ readonly rubric: Schema.String;
486
+ }>>;
487
+ }>;
488
+ export type EvalDimensionSuite = typeof EvalDimensionSuite.Type;
489
+ export declare const EvalDecompositionBenchmarkCase: Schema.Struct<{
490
+ readonly id: Schema.String;
491
+ readonly request: Schema.String;
492
+ readonly expected: Schema.Struct<{
493
+ readonly weights: Schema.$Array<Schema.Struct<{
494
+ readonly dimensionId: Schema.String;
495
+ readonly weight: Schema.Finite;
496
+ }>>;
497
+ readonly unknownWeight: Schema.Finite;
498
+ }>;
499
+ }>;
500
+ export type EvalDecompositionBenchmarkCase = typeof EvalDecompositionBenchmarkCase.Type;
501
+ export declare const EvalDecompositionBenchmark: Schema.Struct<{
502
+ readonly maximumVectorL1Error: Schema.Finite;
503
+ readonly cases: Schema.$Array<Schema.Struct<{
504
+ readonly id: Schema.String;
505
+ readonly request: Schema.String;
506
+ readonly expected: Schema.Struct<{
507
+ readonly weights: Schema.$Array<Schema.Struct<{
508
+ readonly dimensionId: Schema.String;
509
+ readonly weight: Schema.Finite;
510
+ }>>;
511
+ readonly unknownWeight: Schema.Finite;
512
+ }>;
513
+ }>>;
514
+ }>;
515
+ export type EvalDecompositionBenchmark = typeof EvalDecompositionBenchmark.Type;
516
+ export declare const EvalCompositionCase: Schema.Struct<{
517
+ readonly id: Schema.String;
518
+ readonly prompt: Schema.String;
519
+ readonly context: Schema.optionalKey<Schema.String>;
520
+ readonly rubric: Schema.String;
521
+ readonly decomposition: Schema.Struct<{
522
+ readonly weights: Schema.$Array<Schema.Struct<{
523
+ readonly dimensionId: Schema.String;
524
+ readonly weight: Schema.Finite;
525
+ }>>;
526
+ readonly unknownWeight: Schema.Finite;
527
+ }>;
528
+ readonly requirements: Schema.Struct<{
529
+ readonly endpoint: Schema.Literals<readonly ["chat", "responses", "anthropic"]>;
530
+ readonly requiresTools: Schema.Boolean;
531
+ readonly requiresVision: Schema.Boolean;
532
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
533
+ readonly maxOutputTokens: Schema.optionalKey<Schema.Finite>;
534
+ }>;
535
+ }>;
536
+ export type EvalCompositionCase = typeof EvalCompositionCase.Type;
537
+ export declare const EvalCompositionSuite: Schema.Struct<{
538
+ readonly maximumOutputTokens: Schema.Finite;
539
+ readonly minimumWinnerScoreGap: Schema.Finite;
540
+ readonly minimumWinnerAgreement: Schema.Finite;
541
+ readonly cases: Schema.$Array<Schema.Struct<{
542
+ readonly id: Schema.String;
543
+ readonly prompt: Schema.String;
544
+ readonly context: Schema.optionalKey<Schema.String>;
545
+ readonly rubric: Schema.String;
546
+ readonly decomposition: Schema.Struct<{
547
+ readonly weights: Schema.$Array<Schema.Struct<{
548
+ readonly dimensionId: Schema.String;
549
+ readonly weight: Schema.Finite;
550
+ }>>;
551
+ readonly unknownWeight: Schema.Finite;
552
+ }>;
553
+ readonly requirements: Schema.Struct<{
554
+ readonly endpoint: Schema.Literals<readonly ["chat", "responses", "anthropic"]>;
555
+ readonly requiresTools: Schema.Boolean;
556
+ readonly requiresVision: Schema.Boolean;
557
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
558
+ readonly maxOutputTokens: Schema.optionalKey<Schema.Finite>;
559
+ }>;
560
+ }>>;
561
+ }>;
562
+ export type EvalCompositionSuite = typeof EvalCompositionSuite.Type;
563
+ export declare const EvalEvaluationProposal: Schema.Struct<{
564
+ readonly version: Schema.Literal<1>;
565
+ readonly evaluationDigest: Schema.String;
566
+ readonly basisDigest: Schema.String;
567
+ readonly candidateModels: Schema.$Array<Schema.String>;
568
+ readonly judgeModel: Schema.String;
569
+ readonly suites: Schema.$Array<Schema.Struct<{
570
+ readonly version: Schema.Literal<1>;
571
+ readonly dimensionId: Schema.String;
572
+ readonly maximumOutputTokens: Schema.Finite;
573
+ readonly cases: Schema.$Array<Schema.Struct<{
574
+ readonly id: Schema.String;
575
+ readonly prompt: Schema.String;
576
+ readonly context: Schema.optionalKey<Schema.String>;
577
+ readonly rubric: Schema.String;
578
+ }>>;
579
+ }>>;
580
+ readonly decompositionBenchmark: Schema.Struct<{
581
+ readonly maximumVectorL1Error: Schema.Finite;
582
+ readonly cases: Schema.$Array<Schema.Struct<{
583
+ readonly id: Schema.String;
584
+ readonly request: Schema.String;
585
+ readonly expected: Schema.Struct<{
586
+ readonly weights: Schema.$Array<Schema.Struct<{
587
+ readonly dimensionId: Schema.String;
588
+ readonly weight: Schema.Finite;
589
+ }>>;
590
+ readonly unknownWeight: Schema.Finite;
591
+ }>;
592
+ }>>;
593
+ }>;
594
+ readonly compositionSuite: Schema.Struct<{
595
+ readonly maximumOutputTokens: Schema.Finite;
596
+ readonly minimumWinnerScoreGap: Schema.Finite;
597
+ readonly minimumWinnerAgreement: Schema.Finite;
598
+ readonly cases: Schema.$Array<Schema.Struct<{
599
+ readonly id: Schema.String;
600
+ readonly prompt: Schema.String;
601
+ readonly context: Schema.optionalKey<Schema.String>;
602
+ readonly rubric: Schema.String;
603
+ readonly decomposition: Schema.Struct<{
604
+ readonly weights: Schema.$Array<Schema.Struct<{
605
+ readonly dimensionId: Schema.String;
606
+ readonly weight: Schema.Finite;
607
+ }>>;
608
+ readonly unknownWeight: Schema.Finite;
609
+ }>;
610
+ readonly requirements: Schema.Struct<{
611
+ readonly endpoint: Schema.Literals<readonly ["chat", "responses", "anthropic"]>;
612
+ readonly requiresTools: Schema.Boolean;
613
+ readonly requiresVision: Schema.Boolean;
614
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
615
+ readonly maxOutputTokens: Schema.optionalKey<Schema.Finite>;
616
+ }>;
617
+ }>>;
618
+ }>;
619
+ }>;
620
+ export type EvalEvaluationProposal = typeof EvalEvaluationProposal.Type;
621
+ export declare const EvalArtifactApproval: Schema.Struct<{
622
+ readonly version: Schema.Literal<1>;
623
+ readonly kind: Schema.Literals<readonly ["routing-basis", "evaluations"]>;
624
+ readonly digest: Schema.String;
625
+ readonly approvedAt: Schema.String;
626
+ }>;
627
+ export type EvalArtifactApproval = typeof EvalArtifactApproval.Type;
628
+ export declare const EvalPlanScope: Schema.Literals<readonly ["pilot", "full"]>;
629
+ export type EvalPlanScope = typeof EvalPlanScope.Type;
630
+ export declare const EvalExecutionPlan: Schema.Struct<{
631
+ readonly version: Schema.Literal<1>;
632
+ readonly planId: Schema.String;
633
+ readonly projectId: Schema.String;
634
+ readonly projectRevision: Schema.Finite;
635
+ readonly createdAt: Schema.String;
636
+ readonly scope: Schema.Literals<readonly ["pilot", "full"]>;
637
+ readonly basisDigest: Schema.String;
638
+ readonly evaluationDigest: Schema.String;
639
+ readonly candidateModels: Schema.$Array<Schema.String>;
640
+ readonly classifierModel: Schema.String;
641
+ readonly authorModel: Schema.String;
642
+ readonly judgeModel: Schema.String;
643
+ readonly selectedCaseIds: Schema.$Array<Schema.Struct<{
644
+ readonly dimensionId: Schema.String;
645
+ readonly caseIds: Schema.$Array<Schema.String>;
646
+ }>>;
647
+ readonly selectedDecompositionCaseIds: Schema.$Array<Schema.String>;
648
+ readonly selectedCompositionCaseIds: Schema.$Array<Schema.String>;
649
+ readonly maximumOutputTokens: Schema.Finite;
650
+ readonly expectedDimensionCandidateCalls: Schema.Finite;
651
+ readonly expectedDimensionJudgeCalls: Schema.Finite;
652
+ readonly expectedClassifierCalls: Schema.Finite;
653
+ readonly expectedCompositionCandidateCalls: Schema.Finite;
654
+ readonly expectedCompositionJudgeCalls: Schema.Finite;
655
+ readonly expectedCandidateCalls: Schema.Finite;
656
+ readonly expectedJudgeCalls: Schema.Finite;
657
+ readonly expectedCallCount: Schema.Finite;
658
+ }>;
659
+ export type EvalExecutionPlan = typeof EvalExecutionPlan.Type;
660
+ export declare const EvalRunTarget: Schema.Struct<{
661
+ readonly kind: Schema.Literals<readonly ["configured", "external"]>;
662
+ readonly identity: Schema.String;
663
+ readonly publishAllowed: Schema.Boolean;
664
+ }>;
665
+ export type EvalRunTarget = typeof EvalRunTarget.Type;
666
+ export declare const EvalRunCleanup: Schema.Struct<{
667
+ readonly sessionOpened: Schema.Boolean;
668
+ readonly sessionClosed: Schema.Boolean;
669
+ readonly detail: Schema.optionalKey<Schema.String>;
670
+ }>;
671
+ export type EvalRunCleanup = typeof EvalRunCleanup.Type;
672
+ export declare const EvalRunLedger: Schema.Struct<{
673
+ readonly expectedCalls: Schema.Finite;
674
+ readonly observedCalls: Schema.Finite;
675
+ readonly observedCandidateRows: Schema.Finite;
676
+ readonly knownInputTokens: Schema.Finite;
677
+ readonly knownOutputTokens: Schema.Finite;
678
+ readonly unknownTokenMeasurements: Schema.Finite;
679
+ readonly knownPricedSubtotalUsd: Schema.Finite;
680
+ readonly unpricedCalls: Schema.Finite;
681
+ }>;
682
+ export type EvalRunLedger = typeof EvalRunLedger.Type;
683
+ export declare const EvalClassifierObservation: Schema.Struct<{
684
+ readonly caseId: Schema.String;
685
+ readonly weights: Schema.$Array<Schema.Struct<{
686
+ readonly dimensionId: Schema.String;
687
+ readonly weight: Schema.Finite;
688
+ }>>;
689
+ readonly unknownWeight: Schema.Finite;
690
+ readonly vectorL1Error: Schema.Finite;
691
+ readonly passed: Schema.Boolean;
692
+ readonly classifierCallId: Schema.optionalKey<Schema.String>;
693
+ readonly measurement: Schema.Struct<{
694
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
695
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
696
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
697
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
698
+ }>;
699
+ }>;
700
+ export type EvalClassifierObservation = typeof EvalClassifierObservation.Type;
701
+ export declare const EvalCompositionCaseResult: Schema.Struct<{
702
+ readonly caseId: Schema.String;
703
+ readonly predictedWinner: Schema.String;
704
+ readonly observedWinner: Schema.String;
705
+ readonly predictedScoreGap: Schema.Finite;
706
+ readonly observedScoreGap: Schema.Finite;
707
+ readonly passed: Schema.Boolean;
708
+ }>;
709
+ export type EvalCompositionCaseResult = typeof EvalCompositionCaseResult.Type;
710
+ export declare const EvalRunQualification: Schema.Struct<{
711
+ readonly decomposition: Schema.Struct<{
712
+ readonly expectedCases: Schema.Finite;
713
+ readonly passedCases: Schema.Finite;
714
+ readonly maximumObservedL1Error: Schema.Finite;
715
+ readonly observations: Schema.$Array<Schema.Struct<{
716
+ readonly caseId: Schema.String;
717
+ readonly weights: Schema.$Array<Schema.Struct<{
718
+ readonly dimensionId: Schema.String;
719
+ readonly weight: Schema.Finite;
720
+ }>>;
721
+ readonly unknownWeight: Schema.Finite;
722
+ readonly vectorL1Error: Schema.Finite;
723
+ readonly passed: Schema.Boolean;
724
+ readonly classifierCallId: Schema.optionalKey<Schema.String>;
725
+ readonly measurement: Schema.Struct<{
726
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
727
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
728
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
729
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
730
+ }>;
731
+ }>>;
732
+ }>;
733
+ readonly composition: Schema.Struct<{
734
+ readonly expectedCases: Schema.Finite;
735
+ readonly comparableCases: Schema.Finite;
736
+ readonly agreeingCases: Schema.Finite;
737
+ readonly winnerAgreement: Schema.Finite;
738
+ readonly cases: Schema.$Array<Schema.Struct<{
739
+ readonly caseId: Schema.String;
740
+ readonly predictedWinner: Schema.String;
741
+ readonly observedWinner: Schema.String;
742
+ readonly predictedScoreGap: Schema.Finite;
743
+ readonly observedScoreGap: Schema.Finite;
744
+ readonly passed: Schema.Boolean;
745
+ }>>;
746
+ }>;
747
+ }>;
748
+ export type EvalRunQualification = typeof EvalRunQualification.Type;
749
+ export declare function summarizeEvalRunLedger(comparisons: readonly EvalComparisonResultType[], expectedCalls: number, classifierObservations?: readonly EvalClassifierObservation[]): EvalRunLedger;
750
+ /**
751
+ * Sanitized durable result of one immutable execution plan.
752
+ *
753
+ * Prompts, model responses, credentials, headers, and raw child output are
754
+ * deliberately absent from this contract. Dollar totals are known-priced
755
+ * subtotals only; callers must render cost as unknown when
756
+ * `unpricedCalls` is non-zero.
757
+ */
758
+ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
759
+ readonly status: Schema.Literal<"passed">;
760
+ readonly activation: Schema.Struct<{
761
+ readonly version: Schema.Literal<2>;
762
+ readonly generatedAt: Schema.String;
763
+ readonly basisDigest: Schema.String;
764
+ readonly evidenceDigest: Schema.String;
765
+ readonly classifierModel: Schema.String;
766
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
767
+ readonly kind: Schema.Literal<"highest-quality">;
768
+ }>, Schema.Struct<{
769
+ readonly kind: Schema.Literal<"lowest-cost">;
770
+ readonly minimumQuality: Schema.Finite;
771
+ }>, Schema.Struct<{
772
+ readonly kind: Schema.Literal<"lowest-latency">;
773
+ readonly minimumQuality: Schema.Finite;
774
+ }>, Schema.Struct<{
775
+ readonly kind: Schema.Literal<"balanced">;
776
+ readonly minimumQuality: Schema.Finite;
777
+ readonly weights: Schema.Struct<{
778
+ readonly quality: Schema.Finite;
779
+ readonly cost: Schema.Finite;
780
+ readonly latency: Schema.Finite;
781
+ }>;
782
+ }>, Schema.Struct<{
783
+ readonly kind: Schema.Literal<"pareto">;
784
+ readonly minimumQuality: Schema.Finite;
785
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
786
+ }>]>;
787
+ readonly maximumUnknownWeight: Schema.Finite;
788
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
789
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
790
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
791
+ }>>;
792
+ readonly dimensions: Schema.$Array<Schema.Struct<{
793
+ readonly id: Schema.String;
794
+ readonly description: Schema.String;
795
+ readonly includes: Schema.$Array<Schema.String>;
796
+ readonly excludes: Schema.$Array<Schema.String>;
797
+ }>>;
798
+ readonly candidateModels: Schema.$Array<Schema.String>;
799
+ readonly evidence: Schema.$Array<Schema.Struct<{
800
+ readonly model: Schema.String;
801
+ readonly dimensionId: Schema.String;
802
+ readonly suiteDigest: Schema.String;
803
+ readonly evidenceDigest: Schema.String;
804
+ readonly quality: Schema.Struct<{
805
+ readonly passRate: Schema.Finite;
806
+ readonly lowerConfidenceBound: Schema.Finite;
807
+ readonly sampleCount: Schema.Finite;
808
+ }>;
809
+ readonly failureRate: Schema.Finite;
810
+ readonly averageJudgeScore: Schema.optionalKey<Schema.Finite>;
811
+ readonly p95DurationMs: Schema.optionalKey<Schema.Finite>;
812
+ readonly averageCostUsd: Schema.optionalKey<Schema.Finite>;
813
+ readonly unpricedCalls: Schema.Finite;
814
+ }>>;
815
+ }>;
816
+ readonly version: Schema.Literal<1>;
817
+ readonly runId: Schema.String;
818
+ readonly planId: Schema.String;
819
+ readonly projectId: Schema.String;
820
+ readonly startedAt: Schema.String;
821
+ readonly finishedAt: Schema.String;
822
+ readonly basisDigest: Schema.String;
823
+ readonly evaluationDigest: Schema.String;
824
+ readonly target: Schema.Struct<{
825
+ readonly kind: Schema.Literals<readonly ["configured", "external"]>;
826
+ readonly identity: Schema.String;
827
+ readonly publishAllowed: Schema.Boolean;
828
+ }>;
829
+ readonly cleanup: Schema.Struct<{
830
+ readonly sessionOpened: Schema.Boolean;
831
+ readonly sessionClosed: Schema.Boolean;
832
+ readonly detail: Schema.optionalKey<Schema.String>;
833
+ }>;
834
+ readonly comparisons: Schema.$Array<Schema.Struct<{
835
+ readonly version: Schema.Literal<1>;
836
+ readonly comparisonId: Schema.String;
837
+ readonly profileId: Schema.String;
838
+ readonly suiteDigest: Schema.String;
839
+ readonly judgeModel: Schema.String;
840
+ readonly startedAt: Schema.String;
841
+ readonly finishedAt: Schema.String;
842
+ readonly models: Schema.$Array<Schema.Struct<{
843
+ readonly model: Schema.String;
844
+ readonly cases: Schema.$Array<Schema.Struct<{
845
+ readonly caseId: Schema.String;
846
+ readonly outcome: Schema.Literals<readonly ["passed", "failed", "unknown", "cutoff"]>;
847
+ readonly measurement: Schema.Struct<{
848
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
849
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
850
+ readonly judgeScore: Schema.optionalKey<Schema.Finite>;
851
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
852
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
853
+ }>;
854
+ readonly error: Schema.optionalKey<Schema.String>;
855
+ }>>;
856
+ }>>;
857
+ readonly calls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
858
+ readonly role: Schema.Literals<readonly ["candidate", "judge"]>;
859
+ readonly model: Schema.String;
860
+ readonly caseId: Schema.String;
861
+ readonly measurement: Schema.Struct<{
862
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
863
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
864
+ readonly judgeScore: Schema.optionalKey<Schema.Finite>;
865
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
866
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
867
+ }>;
868
+ }>>>;
869
+ }>>;
870
+ readonly qualification: Schema.optionalKey<Schema.Struct<{
871
+ readonly decomposition: Schema.Struct<{
872
+ readonly expectedCases: Schema.Finite;
873
+ readonly passedCases: Schema.Finite;
874
+ readonly maximumObservedL1Error: Schema.Finite;
875
+ readonly observations: Schema.$Array<Schema.Struct<{
876
+ readonly caseId: Schema.String;
877
+ readonly weights: Schema.$Array<Schema.Struct<{
878
+ readonly dimensionId: Schema.String;
879
+ readonly weight: Schema.Finite;
880
+ }>>;
881
+ readonly unknownWeight: Schema.Finite;
882
+ readonly vectorL1Error: Schema.Finite;
883
+ readonly passed: Schema.Boolean;
884
+ readonly classifierCallId: Schema.optionalKey<Schema.String>;
885
+ readonly measurement: Schema.Struct<{
886
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
887
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
888
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
889
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
890
+ }>;
891
+ }>>;
892
+ }>;
893
+ readonly composition: Schema.Struct<{
894
+ readonly expectedCases: Schema.Finite;
895
+ readonly comparableCases: Schema.Finite;
896
+ readonly agreeingCases: Schema.Finite;
897
+ readonly winnerAgreement: Schema.Finite;
898
+ readonly cases: Schema.$Array<Schema.Struct<{
899
+ readonly caseId: Schema.String;
900
+ readonly predictedWinner: Schema.String;
901
+ readonly observedWinner: Schema.String;
902
+ readonly predictedScoreGap: Schema.Finite;
903
+ readonly observedScoreGap: Schema.Finite;
904
+ readonly passed: Schema.Boolean;
905
+ }>>;
906
+ }>;
907
+ }>>;
908
+ readonly ledger: Schema.Struct<{
909
+ readonly expectedCalls: Schema.Finite;
910
+ readonly observedCalls: Schema.Finite;
911
+ readonly observedCandidateRows: Schema.Finite;
912
+ readonly knownInputTokens: Schema.Finite;
913
+ readonly knownOutputTokens: Schema.Finite;
914
+ readonly unknownTokenMeasurements: Schema.Finite;
915
+ readonly knownPricedSubtotalUsd: Schema.Finite;
916
+ readonly unpricedCalls: Schema.Finite;
917
+ }>;
918
+ }>, Schema.Struct<{
919
+ readonly status: Schema.Literal<"completed">;
920
+ readonly activation: Schema.Struct<{
921
+ readonly version: Schema.Literal<2>;
922
+ readonly generatedAt: Schema.String;
923
+ readonly basisDigest: Schema.String;
924
+ readonly evidenceDigest: Schema.String;
925
+ readonly classifierModel: Schema.String;
926
+ readonly objective: Schema.Union<readonly [Schema.Struct<{
927
+ readonly kind: Schema.Literal<"highest-quality">;
928
+ }>, Schema.Struct<{
929
+ readonly kind: Schema.Literal<"lowest-cost">;
930
+ readonly minimumQuality: Schema.Finite;
931
+ }>, Schema.Struct<{
932
+ readonly kind: Schema.Literal<"lowest-latency">;
933
+ readonly minimumQuality: Schema.Finite;
934
+ }>, Schema.Struct<{
935
+ readonly kind: Schema.Literal<"balanced">;
936
+ readonly minimumQuality: Schema.Finite;
937
+ readonly weights: Schema.Struct<{
938
+ readonly quality: Schema.Finite;
939
+ readonly cost: Schema.Finite;
940
+ readonly latency: Schema.Finite;
941
+ }>;
942
+ }>, Schema.Struct<{
943
+ readonly kind: Schema.Literal<"pareto">;
944
+ readonly minimumQuality: Schema.Finite;
945
+ readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
946
+ }>]>;
947
+ readonly maximumUnknownWeight: Schema.Finite;
948
+ readonly constraints: Schema.optionalKey<Schema.Struct<{
949
+ readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
950
+ readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
951
+ }>>;
952
+ readonly dimensions: Schema.$Array<Schema.Struct<{
953
+ readonly id: Schema.String;
954
+ readonly description: Schema.String;
955
+ readonly includes: Schema.$Array<Schema.String>;
956
+ readonly excludes: Schema.$Array<Schema.String>;
957
+ }>>;
958
+ readonly candidateModels: Schema.$Array<Schema.String>;
959
+ readonly evidence: Schema.$Array<Schema.Struct<{
960
+ readonly model: Schema.String;
961
+ readonly dimensionId: Schema.String;
962
+ readonly suiteDigest: Schema.String;
963
+ readonly evidenceDigest: Schema.String;
964
+ readonly quality: Schema.Struct<{
965
+ readonly passRate: Schema.Finite;
966
+ readonly lowerConfidenceBound: Schema.Finite;
967
+ readonly sampleCount: Schema.Finite;
968
+ }>;
969
+ readonly failureRate: Schema.Finite;
970
+ readonly averageJudgeScore: Schema.optionalKey<Schema.Finite>;
971
+ readonly p95DurationMs: Schema.optionalKey<Schema.Finite>;
972
+ readonly averageCostUsd: Schema.optionalKey<Schema.Finite>;
973
+ readonly unpricedCalls: Schema.Finite;
974
+ }>>;
975
+ }>;
976
+ readonly version: Schema.Literal<1>;
977
+ readonly runId: Schema.String;
978
+ readonly planId: Schema.String;
979
+ readonly projectId: Schema.String;
980
+ readonly startedAt: Schema.String;
981
+ readonly finishedAt: Schema.String;
982
+ readonly basisDigest: Schema.String;
983
+ readonly evaluationDigest: Schema.String;
984
+ readonly target: Schema.Struct<{
985
+ readonly kind: Schema.Literals<readonly ["configured", "external"]>;
986
+ readonly identity: Schema.String;
987
+ readonly publishAllowed: Schema.Boolean;
988
+ }>;
989
+ readonly cleanup: Schema.Struct<{
990
+ readonly sessionOpened: Schema.Boolean;
991
+ readonly sessionClosed: Schema.Boolean;
992
+ readonly detail: Schema.optionalKey<Schema.String>;
993
+ }>;
994
+ readonly comparisons: Schema.$Array<Schema.Struct<{
995
+ readonly version: Schema.Literal<1>;
996
+ readonly comparisonId: Schema.String;
997
+ readonly profileId: Schema.String;
998
+ readonly suiteDigest: Schema.String;
999
+ readonly judgeModel: Schema.String;
1000
+ readonly startedAt: Schema.String;
1001
+ readonly finishedAt: Schema.String;
1002
+ readonly models: Schema.$Array<Schema.Struct<{
1003
+ readonly model: Schema.String;
1004
+ readonly cases: Schema.$Array<Schema.Struct<{
1005
+ readonly caseId: Schema.String;
1006
+ readonly outcome: Schema.Literals<readonly ["passed", "failed", "unknown", "cutoff"]>;
1007
+ readonly measurement: Schema.Struct<{
1008
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
1009
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
1010
+ readonly judgeScore: Schema.optionalKey<Schema.Finite>;
1011
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
1012
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
1013
+ }>;
1014
+ readonly error: Schema.optionalKey<Schema.String>;
1015
+ }>>;
1016
+ }>>;
1017
+ readonly calls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
1018
+ readonly role: Schema.Literals<readonly ["candidate", "judge"]>;
1019
+ readonly model: Schema.String;
1020
+ readonly caseId: Schema.String;
1021
+ readonly measurement: Schema.Struct<{
1022
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
1023
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
1024
+ readonly judgeScore: Schema.optionalKey<Schema.Finite>;
1025
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
1026
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
1027
+ }>;
1028
+ }>>>;
1029
+ }>>;
1030
+ readonly qualification: Schema.optionalKey<Schema.Struct<{
1031
+ readonly decomposition: Schema.Struct<{
1032
+ readonly expectedCases: Schema.Finite;
1033
+ readonly passedCases: Schema.Finite;
1034
+ readonly maximumObservedL1Error: Schema.Finite;
1035
+ readonly observations: Schema.$Array<Schema.Struct<{
1036
+ readonly caseId: Schema.String;
1037
+ readonly weights: Schema.$Array<Schema.Struct<{
1038
+ readonly dimensionId: Schema.String;
1039
+ readonly weight: Schema.Finite;
1040
+ }>>;
1041
+ readonly unknownWeight: Schema.Finite;
1042
+ readonly vectorL1Error: Schema.Finite;
1043
+ readonly passed: Schema.Boolean;
1044
+ readonly classifierCallId: Schema.optionalKey<Schema.String>;
1045
+ readonly measurement: Schema.Struct<{
1046
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
1047
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
1048
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
1049
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
1050
+ }>;
1051
+ }>>;
1052
+ }>;
1053
+ readonly composition: Schema.Struct<{
1054
+ readonly expectedCases: Schema.Finite;
1055
+ readonly comparableCases: Schema.Finite;
1056
+ readonly agreeingCases: Schema.Finite;
1057
+ readonly winnerAgreement: Schema.Finite;
1058
+ readonly cases: Schema.$Array<Schema.Struct<{
1059
+ readonly caseId: Schema.String;
1060
+ readonly predictedWinner: Schema.String;
1061
+ readonly observedWinner: Schema.String;
1062
+ readonly predictedScoreGap: Schema.Finite;
1063
+ readonly observedScoreGap: Schema.Finite;
1064
+ readonly passed: Schema.Boolean;
1065
+ }>>;
1066
+ }>;
1067
+ }>>;
1068
+ readonly ledger: Schema.Struct<{
1069
+ readonly expectedCalls: Schema.Finite;
1070
+ readonly observedCalls: Schema.Finite;
1071
+ readonly observedCandidateRows: Schema.Finite;
1072
+ readonly knownInputTokens: Schema.Finite;
1073
+ readonly knownOutputTokens: Schema.Finite;
1074
+ readonly unknownTokenMeasurements: Schema.Finite;
1075
+ readonly knownPricedSubtotalUsd: Schema.Finite;
1076
+ readonly unpricedCalls: Schema.Finite;
1077
+ }>;
1078
+ }>, Schema.Struct<{
1079
+ readonly status: Schema.Literal<"failed">;
1080
+ readonly failure: Schema.String;
1081
+ readonly version: Schema.Literal<1>;
1082
+ readonly runId: Schema.String;
1083
+ readonly planId: Schema.String;
1084
+ readonly projectId: Schema.String;
1085
+ readonly startedAt: Schema.String;
1086
+ readonly finishedAt: Schema.String;
1087
+ readonly basisDigest: Schema.String;
1088
+ readonly evaluationDigest: Schema.String;
1089
+ readonly target: Schema.Struct<{
1090
+ readonly kind: Schema.Literals<readonly ["configured", "external"]>;
1091
+ readonly identity: Schema.String;
1092
+ readonly publishAllowed: Schema.Boolean;
1093
+ }>;
1094
+ readonly cleanup: Schema.Struct<{
1095
+ readonly sessionOpened: Schema.Boolean;
1096
+ readonly sessionClosed: Schema.Boolean;
1097
+ readonly detail: Schema.optionalKey<Schema.String>;
1098
+ }>;
1099
+ readonly comparisons: Schema.$Array<Schema.Struct<{
1100
+ readonly version: Schema.Literal<1>;
1101
+ readonly comparisonId: Schema.String;
1102
+ readonly profileId: Schema.String;
1103
+ readonly suiteDigest: Schema.String;
1104
+ readonly judgeModel: Schema.String;
1105
+ readonly startedAt: Schema.String;
1106
+ readonly finishedAt: Schema.String;
1107
+ readonly models: Schema.$Array<Schema.Struct<{
1108
+ readonly model: Schema.String;
1109
+ readonly cases: Schema.$Array<Schema.Struct<{
1110
+ readonly caseId: Schema.String;
1111
+ readonly outcome: Schema.Literals<readonly ["passed", "failed", "unknown", "cutoff"]>;
1112
+ readonly measurement: Schema.Struct<{
1113
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
1114
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
1115
+ readonly judgeScore: Schema.optionalKey<Schema.Finite>;
1116
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
1117
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
1118
+ }>;
1119
+ readonly error: Schema.optionalKey<Schema.String>;
1120
+ }>>;
1121
+ }>>;
1122
+ readonly calls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
1123
+ readonly role: Schema.Literals<readonly ["candidate", "judge"]>;
1124
+ readonly model: Schema.String;
1125
+ readonly caseId: Schema.String;
1126
+ readonly measurement: Schema.Struct<{
1127
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
1128
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
1129
+ readonly judgeScore: Schema.optionalKey<Schema.Finite>;
1130
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
1131
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
1132
+ }>;
1133
+ }>>>;
1134
+ }>>;
1135
+ readonly qualification: Schema.optionalKey<Schema.Struct<{
1136
+ readonly decomposition: Schema.Struct<{
1137
+ readonly expectedCases: Schema.Finite;
1138
+ readonly passedCases: Schema.Finite;
1139
+ readonly maximumObservedL1Error: Schema.Finite;
1140
+ readonly observations: Schema.$Array<Schema.Struct<{
1141
+ readonly caseId: Schema.String;
1142
+ readonly weights: Schema.$Array<Schema.Struct<{
1143
+ readonly dimensionId: Schema.String;
1144
+ readonly weight: Schema.Finite;
1145
+ }>>;
1146
+ readonly unknownWeight: Schema.Finite;
1147
+ readonly vectorL1Error: Schema.Finite;
1148
+ readonly passed: Schema.Boolean;
1149
+ readonly classifierCallId: Schema.optionalKey<Schema.String>;
1150
+ readonly measurement: Schema.Struct<{
1151
+ readonly costUsd: Schema.optionalKey<Schema.Finite>;
1152
+ readonly durationMs: Schema.optionalKey<Schema.Finite>;
1153
+ readonly inputTokens: Schema.optionalKey<Schema.Finite>;
1154
+ readonly outputTokens: Schema.optionalKey<Schema.Finite>;
1155
+ }>;
1156
+ }>>;
1157
+ }>;
1158
+ readonly composition: Schema.Struct<{
1159
+ readonly expectedCases: Schema.Finite;
1160
+ readonly comparableCases: Schema.Finite;
1161
+ readonly agreeingCases: Schema.Finite;
1162
+ readonly winnerAgreement: Schema.Finite;
1163
+ readonly cases: Schema.$Array<Schema.Struct<{
1164
+ readonly caseId: Schema.String;
1165
+ readonly predictedWinner: Schema.String;
1166
+ readonly observedWinner: Schema.String;
1167
+ readonly predictedScoreGap: Schema.Finite;
1168
+ readonly observedScoreGap: Schema.Finite;
1169
+ readonly passed: Schema.Boolean;
1170
+ }>>;
1171
+ }>;
1172
+ }>>;
1173
+ readonly ledger: Schema.Struct<{
1174
+ readonly expectedCalls: Schema.Finite;
1175
+ readonly observedCalls: Schema.Finite;
1176
+ readonly observedCandidateRows: Schema.Finite;
1177
+ readonly knownInputTokens: Schema.Finite;
1178
+ readonly knownOutputTokens: Schema.Finite;
1179
+ readonly unknownTokenMeasurements: Schema.Finite;
1180
+ readonly knownPricedSubtotalUsd: Schema.Finite;
1181
+ readonly unpricedCalls: Schema.Finite;
1182
+ }>;
1183
+ }>]>;
1184
+ export type EvalRunReport = typeof EvalRunReport.Type;
1185
+ export type EvalProjectArtifactsStatus = {
1186
+ readonly basisProposalDigest?: string;
1187
+ readonly basisApproved: boolean;
1188
+ readonly evaluationProposalDigest?: string;
1189
+ readonly evaluationsApproved: boolean;
1190
+ readonly plans: readonly string[];
1191
+ };
1192
+ export type EvalProjectStatus = {
1193
+ readonly state: EvalProjectState;
1194
+ readonly question?: EvalProjectQuestion;
1195
+ readonly artifacts?: EvalProjectArtifactsStatus;
1196
+ readonly nextAction: "answer" | "propose-dimensions" | "approve-dimensions" | "propose-evaluations" | "approve-evaluations" | "run" | "publish" | "none";
1197
+ };