@openpond/evals 0.11.16 → 0.11.17

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 (158) hide show
  1. package/dist/connected-evidence/contracts.js +72 -0
  2. package/dist/connected-evidence/imports.js +71 -0
  3. package/dist/connected-evidence/index.js +5 -0
  4. package/dist/connected-evidence/normalize.js +97 -0
  5. package/dist/connected-evidence/python-correctness.js +73 -0
  6. package/dist/connected-evidence/sources/claude-code.js +71 -0
  7. package/dist/connected-evidence/sources/codex.js +52 -0
  8. package/dist/connected-evidence/sources/history.js +86 -0
  9. package/dist/connected-evidence/sources/usage.js +15 -0
  10. package/dist/connected-evidence/usage-summary.js +52 -0
  11. package/dist/experiments.js +16 -4
  12. package/dist/graders.js +9 -6
  13. package/dist/human-review/comparison.js +161 -0
  14. package/dist/human-review/contracts.js +110 -0
  15. package/dist/human-review/execution.js +61 -0
  16. package/dist/human-review/handoff.js +18 -0
  17. package/dist/human-review/index.js +9 -0
  18. package/dist/human-review/local-publication.js +18 -0
  19. package/dist/human-review/repository.js +1 -0
  20. package/dist/human-review/results.js +18 -0
  21. package/dist/human-review/service.js +268 -0
  22. package/dist/human-review/task-proposals.js +39 -0
  23. package/dist/index.js +1 -0
  24. package/dist/javascript-verifier-contract.js +2 -2
  25. package/dist/learning/authoring.js +2 -1
  26. package/dist/learning/contracts.js +353 -81
  27. package/dist/learning/index.js +1 -0
  28. package/dist/learning/reward-authoring.js +5 -2
  29. package/dist/learning/training-handoff.js +32 -0
  30. package/dist/metrics-node.js +2 -0
  31. package/dist/profile-evaluation-target.js +7 -0
  32. package/dist/profile-evaluations.js +7 -10
  33. package/dist/profile-external-evaluation.js +35 -0
  34. package/dist/profile-source-candidate.js +5 -0
  35. package/dist/rewards.js +5 -3
  36. package/dist/taskset-run-contract.js +10 -0
  37. package/dist/tasksets.js +2 -0
  38. package/dist/types/builtin-benchmarks/harness-refiner.d.ts +75 -1
  39. package/dist/types/builtin-benchmarks/harness-refiner.d.ts.map +1 -1
  40. package/dist/types/conformance.d.ts +150 -2
  41. package/dist/types/conformance.d.ts.map +1 -1
  42. package/dist/types/connected-evidence/contracts.d.ts +317 -0
  43. package/dist/types/connected-evidence/contracts.d.ts.map +1 -0
  44. package/dist/types/connected-evidence/imports.d.ts +19 -0
  45. package/dist/types/connected-evidence/imports.d.ts.map +1 -0
  46. package/dist/types/connected-evidence/index.d.ts +6 -0
  47. package/dist/types/connected-evidence/index.d.ts.map +1 -0
  48. package/dist/types/connected-evidence/normalize.d.ts +25 -0
  49. package/dist/types/connected-evidence/normalize.d.ts.map +1 -0
  50. package/dist/types/connected-evidence/python-correctness.d.ts +196 -0
  51. package/dist/types/connected-evidence/python-correctness.d.ts.map +1 -0
  52. package/dist/types/connected-evidence/sources/claude-code.d.ts +74 -0
  53. package/dist/types/connected-evidence/sources/claude-code.d.ts.map +1 -0
  54. package/dist/types/connected-evidence/sources/codex.d.ts +74 -0
  55. package/dist/types/connected-evidence/sources/codex.d.ts.map +1 -0
  56. package/dist/types/connected-evidence/sources/history.d.ts +146 -0
  57. package/dist/types/connected-evidence/sources/history.d.ts.map +1 -0
  58. package/dist/types/connected-evidence/sources/usage.d.ts +3 -0
  59. package/dist/types/connected-evidence/sources/usage.d.ts.map +1 -0
  60. package/dist/types/connected-evidence/usage-summary.d.ts +162 -0
  61. package/dist/types/connected-evidence/usage-summary.d.ts.map +1 -0
  62. package/dist/types/evidence/conformance.d.ts +28 -28
  63. package/dist/types/evidence/contracts.d.ts +27 -27
  64. package/dist/types/execution-contracts.d.ts +159 -5
  65. package/dist/types/execution-contracts.d.ts.map +1 -1
  66. package/dist/types/experiments.d.ts +476 -7
  67. package/dist/types/experiments.d.ts.map +1 -1
  68. package/dist/types/graders.d.ts +21 -1
  69. package/dist/types/graders.d.ts.map +1 -1
  70. package/dist/types/human-review/comparison.d.ts +1959 -0
  71. package/dist/types/human-review/comparison.d.ts.map +1 -0
  72. package/dist/types/human-review/contracts.d.ts +2719 -0
  73. package/dist/types/human-review/contracts.d.ts.map +1 -0
  74. package/dist/types/human-review/execution.d.ts +298 -0
  75. package/dist/types/human-review/execution.d.ts.map +1 -0
  76. package/dist/types/human-review/handoff.d.ts +124 -0
  77. package/dist/types/human-review/handoff.d.ts.map +1 -0
  78. package/dist/types/human-review/index.d.ts +10 -0
  79. package/dist/types/human-review/index.d.ts.map +1 -0
  80. package/dist/types/human-review/local-publication.d.ts +165 -0
  81. package/dist/types/human-review/local-publication.d.ts.map +1 -0
  82. package/dist/types/human-review/repository.d.ts +43 -0
  83. package/dist/types/human-review/repository.d.ts.map +1 -0
  84. package/dist/types/human-review/results.d.ts +118 -0
  85. package/dist/types/human-review/results.d.ts.map +1 -0
  86. package/dist/types/human-review/service.d.ts +878 -0
  87. package/dist/types/human-review/service.d.ts.map +1 -0
  88. package/dist/types/human-review/task-proposals.d.ts +70 -0
  89. package/dist/types/human-review/task-proposals.d.ts.map +1 -0
  90. package/dist/types/index.d.ts +1 -0
  91. package/dist/types/index.d.ts.map +1 -1
  92. package/dist/types/javascript-environment-contract.d.ts +1 -1
  93. package/dist/types/javascript-verifier-contract.d.ts +1 -1
  94. package/dist/types/javascript-verifier-contract.d.ts.map +1 -1
  95. package/dist/types/learned-preference.d.ts +2 -2
  96. package/dist/types/learning/acceptance-group.d.ts +13 -13
  97. package/dist/types/learning/acceptance-plan.d.ts +1 -1
  98. package/dist/types/learning/admission.d.ts +78 -1
  99. package/dist/types/learning/admission.d.ts.map +1 -1
  100. package/dist/types/learning/authoring-service.d.ts +2 -0
  101. package/dist/types/learning/authoring-service.d.ts.map +1 -1
  102. package/dist/types/learning/authoring.d.ts +20 -0
  103. package/dist/types/learning/authoring.d.ts.map +1 -1
  104. package/dist/types/learning/batch-service.d.ts +76 -2
  105. package/dist/types/learning/batch-service.d.ts.map +1 -1
  106. package/dist/types/learning/contracts.d.ts +759 -24
  107. package/dist/types/learning/contracts.d.ts.map +1 -1
  108. package/dist/types/learning/index.d.ts +1 -0
  109. package/dist/types/learning/index.d.ts.map +1 -1
  110. package/dist/types/learning/iteration-candidate-decision.d.ts +2 -2
  111. package/dist/types/learning/iteration-dispatch-contracts.d.ts +2 -2
  112. package/dist/types/learning/operations.d.ts +1118 -8
  113. package/dist/types/learning/operations.d.ts.map +1 -1
  114. package/dist/types/learning/repository.d.ts +402 -15
  115. package/dist/types/learning/repository.d.ts.map +1 -1
  116. package/dist/types/learning/reward-authoring.d.ts.map +1 -1
  117. package/dist/types/learning/reward-calibration.d.ts +74 -0
  118. package/dist/types/learning/reward-calibration.d.ts.map +1 -1
  119. package/dist/types/learning/reward-checks.d.ts +14 -2
  120. package/dist/types/learning/reward-checks.d.ts.map +1 -1
  121. package/dist/types/learning/training-handoff.d.ts +201 -0
  122. package/dist/types/learning/training-handoff.d.ts.map +1 -0
  123. package/dist/types/learning/transport.d.ts +963 -21
  124. package/dist/types/learning/transport.d.ts.map +1 -1
  125. package/dist/types/metrics-node.d.ts +3 -1
  126. package/dist/types/metrics-node.d.ts.map +1 -1
  127. package/dist/types/model-improvement-qualification.d.ts +3 -3
  128. package/dist/types/native-harness-experiment-attempt.d.ts +2 -2
  129. package/dist/types/profile-evaluation-comparisons.d.ts +230 -0
  130. package/dist/types/profile-evaluation-comparisons.d.ts.map +1 -1
  131. package/dist/types/profile-evaluation-target.d.ts +14 -0
  132. package/dist/types/profile-evaluation-target.d.ts.map +1 -0
  133. package/dist/types/profile-evaluations.d.ts +117 -14
  134. package/dist/types/profile-evaluations.d.ts.map +1 -1
  135. package/dist/types/profile-external-evaluation.d.ts +404 -0
  136. package/dist/types/profile-external-evaluation.d.ts.map +1 -0
  137. package/dist/types/profile-source-candidate.d.ts +15 -0
  138. package/dist/types/profile-source-candidate.d.ts.map +1 -0
  139. package/dist/types/review-conformance.d.ts +7 -7
  140. package/dist/types/rewards.d.ts +253 -0
  141. package/dist/types/rewards.d.ts.map +1 -1
  142. package/dist/types/rollouts.d.ts +3 -3
  143. package/dist/types/taskset-run-contract.d.ts +230 -0
  144. package/dist/types/taskset-run-contract.d.ts.map +1 -1
  145. package/dist/types/tasksets.d.ts +311 -3
  146. package/dist/types/tasksets.d.ts.map +1 -1
  147. package/dist/types/telemetry-analysis.d.ts +2 -2
  148. package/dist/types/telemetry-bundle.d.ts +6 -6
  149. package/dist/types/telemetry.d.ts +3 -3
  150. package/package.json +9 -1
  151. package/schemas/learning/v1/command-request.schema.json +1745 -409
  152. package/schemas/learning/v1/decision.schema.json +11 -0
  153. package/schemas/learning/v1/draft.schema.json +11 -0
  154. package/schemas/learning/v1/grade.schema.json +11 -0
  155. package/schemas/learning/v1/package.schema.json +532 -12
  156. package/schemas/learning/v1/policy.schema.json +836 -28
  157. package/schemas/learning/v1/reward.schema.json +520 -0
  158. package/schemas/learning/v1/reward_check.schema.json +9 -0
@@ -2,12 +2,487 @@
2
2
  "$schema": "https://json-schema.org/draft/2020-12/schema",
3
3
  "type": "object",
4
4
  "properties": {
5
+ "origin": {
6
+ "type": "object",
7
+ "properties": {
8
+ "schemaVersion": {
9
+ "type": "string",
10
+ "const": "openpond.trainingHandoffOrigin.v1"
11
+ },
12
+ "teamId": {
13
+ "$ref": "#/$defs/__schema0"
14
+ },
15
+ "projectId": {
16
+ "anyOf": [
17
+ {
18
+ "$ref": "#/$defs/__schema0"
19
+ },
20
+ {
21
+ "type": "null"
22
+ }
23
+ ]
24
+ },
25
+ "configurationId": {
26
+ "$ref": "#/$defs/__schema0"
27
+ },
28
+ "source": {
29
+ "type": "object",
30
+ "properties": {
31
+ "executionId": {
32
+ "$ref": "#/$defs/__schema0"
33
+ },
34
+ "passId": {
35
+ "anyOf": [
36
+ {
37
+ "$ref": "#/$defs/__schema0"
38
+ },
39
+ {
40
+ "type": "null"
41
+ }
42
+ ]
43
+ },
44
+ "manifestHash": {
45
+ "$ref": "#/$defs/__schema1"
46
+ },
47
+ "resultHash": {
48
+ "$ref": "#/$defs/__schema1"
49
+ }
50
+ },
51
+ "required": [
52
+ "executionId",
53
+ "passId",
54
+ "manifestHash",
55
+ "resultHash"
56
+ ],
57
+ "additionalProperties": false
58
+ },
59
+ "dataset": {
60
+ "type": "object",
61
+ "properties": {
62
+ "id": {
63
+ "$ref": "#/$defs/__schema0"
64
+ },
65
+ "revision": {
66
+ "type": "integer",
67
+ "exclusiveMinimum": 0,
68
+ "maximum": 9007199254740991
69
+ },
70
+ "contentHash": {
71
+ "$ref": "#/$defs/__schema1"
72
+ }
73
+ },
74
+ "required": [
75
+ "id",
76
+ "revision",
77
+ "contentHash"
78
+ ],
79
+ "additionalProperties": false
80
+ },
81
+ "target": {
82
+ "oneOf": [
83
+ {
84
+ "type": "object",
85
+ "properties": {
86
+ "kind": {
87
+ "type": "string",
88
+ "const": "model"
89
+ },
90
+ "modelId": {
91
+ "$ref": "#/$defs/__schema2"
92
+ },
93
+ "configurationHash": {
94
+ "$ref": "#/$defs/__schema3"
95
+ }
96
+ },
97
+ "required": [
98
+ "kind",
99
+ "modelId",
100
+ "configurationHash"
101
+ ],
102
+ "additionalProperties": false
103
+ },
104
+ {
105
+ "type": "object",
106
+ "properties": {
107
+ "kind": {
108
+ "type": "string",
109
+ "const": "harness"
110
+ },
111
+ "source": {
112
+ "type": "object",
113
+ "properties": {
114
+ "profileId": {
115
+ "$ref": "#/$defs/__schema2"
116
+ },
117
+ "sourceRevision": {
118
+ "type": "string",
119
+ "minLength": 1,
120
+ "maxLength": 240
121
+ },
122
+ "harnessRelease": {
123
+ "$ref": "#/$defs/__schema4"
124
+ },
125
+ "catalogHash": {
126
+ "$ref": "#/$defs/__schema3"
127
+ },
128
+ "definitionId": {
129
+ "$ref": "#/$defs/__schema2"
130
+ },
131
+ "definitionHash": {
132
+ "$ref": "#/$defs/__schema3"
133
+ },
134
+ "target": {
135
+ "$ref": "#/$defs/__schema5"
136
+ },
137
+ "environmentHash": {
138
+ "$ref": "#/$defs/__schema3"
139
+ },
140
+ "externalDatasetBinding": {
141
+ "type": "object",
142
+ "properties": {
143
+ "schemaVersion": {
144
+ "type": "string",
145
+ "const": "openpond.profileExternalDatasetBinding.v1"
146
+ },
147
+ "dataset": {
148
+ "type": "object",
149
+ "properties": {
150
+ "id": {
151
+ "$ref": "#/$defs/__schema2"
152
+ },
153
+ "contentHash": {
154
+ "$ref": "#/$defs/__schema3"
155
+ },
156
+ "revision": {
157
+ "type": "integer",
158
+ "exclusiveMinimum": 0,
159
+ "maximum": 9007199254740991
160
+ }
161
+ },
162
+ "required": [
163
+ "id",
164
+ "contentHash",
165
+ "revision"
166
+ ],
167
+ "additionalProperties": false
168
+ },
169
+ "packageHash": {
170
+ "$ref": "#/$defs/__schema3"
171
+ },
172
+ "profileId": {
173
+ "$ref": "#/$defs/__schema2"
174
+ },
175
+ "target": {
176
+ "$ref": "#/$defs/__schema5"
177
+ },
178
+ "declaredProfileCatalogHash": {
179
+ "anyOf": [
180
+ {
181
+ "$ref": "#/$defs/__schema3"
182
+ },
183
+ {
184
+ "type": "null"
185
+ }
186
+ ]
187
+ },
188
+ "protectedProfileClosureHash": {
189
+ "$ref": "#/$defs/__schema3"
190
+ },
191
+ "recordedOrigin": {
192
+ "type": "object",
193
+ "properties": {
194
+ "execution": {
195
+ "$ref": "#/$defs/__schema4"
196
+ },
197
+ "originalEvidence": {
198
+ "$ref": "#/$defs/__schema4"
199
+ },
200
+ "members": {
201
+ "minItems": 1,
202
+ "maxItems": 500,
203
+ "type": "array",
204
+ "items": {
205
+ "$ref": "#/$defs/__schema6"
206
+ }
207
+ }
208
+ },
209
+ "required": [
210
+ "execution",
211
+ "originalEvidence",
212
+ "members"
213
+ ],
214
+ "additionalProperties": false
215
+ },
216
+ "gradingSource": {
217
+ "type": "object",
218
+ "properties": {
219
+ "pass": {
220
+ "$ref": "#/$defs/__schema4"
221
+ },
222
+ "evidence": {
223
+ "$ref": "#/$defs/__schema4"
224
+ }
225
+ },
226
+ "required": [
227
+ "pass",
228
+ "evidence"
229
+ ],
230
+ "additionalProperties": false
231
+ },
232
+ "split": {
233
+ "type": "string",
234
+ "enum": [
235
+ "train",
236
+ "validation",
237
+ "test",
238
+ "frozen_eval"
239
+ ]
240
+ },
241
+ "population": {
242
+ "minItems": 1,
243
+ "maxItems": 10000,
244
+ "type": "array",
245
+ "items": {
246
+ "$ref": "#/$defs/__schema7"
247
+ }
248
+ },
249
+ "evaluators": {
250
+ "minItems": 1,
251
+ "maxItems": 100,
252
+ "type": "array",
253
+ "items": {
254
+ "$ref": "#/$defs/__schema8"
255
+ }
256
+ },
257
+ "fieldMappingsHash": {
258
+ "$ref": "#/$defs/__schema3"
259
+ },
260
+ "fieldMappings": {
261
+ "maxItems": 100,
262
+ "type": "array",
263
+ "items": {
264
+ "$ref": "#/$defs/__schema10"
265
+ }
266
+ },
267
+ "environmentHash": {
268
+ "$ref": "#/$defs/__schema3"
269
+ },
270
+ "privateDatasetClosureHash": {
271
+ "$ref": "#/$defs/__schema3"
272
+ },
273
+ "criterion": {
274
+ "type": "object",
275
+ "properties": {
276
+ "minimumPassRate": {
277
+ "type": "number",
278
+ "minimum": 0,
279
+ "maximum": 1
280
+ },
281
+ "requireComplete": {
282
+ "type": "boolean"
283
+ }
284
+ },
285
+ "required": [
286
+ "minimumPassRate",
287
+ "requireComplete"
288
+ ],
289
+ "additionalProperties": false
290
+ },
291
+ "contentHash": {
292
+ "$ref": "#/$defs/__schema3"
293
+ }
294
+ },
295
+ "required": [
296
+ "schemaVersion",
297
+ "dataset",
298
+ "packageHash",
299
+ "profileId",
300
+ "target",
301
+ "declaredProfileCatalogHash",
302
+ "protectedProfileClosureHash",
303
+ "split",
304
+ "population",
305
+ "evaluators",
306
+ "fieldMappingsHash",
307
+ "fieldMappings",
308
+ "environmentHash",
309
+ "privateDatasetClosureHash",
310
+ "criterion",
311
+ "contentHash"
312
+ ],
313
+ "additionalProperties": false
314
+ },
315
+ "sourceCandidate": {
316
+ "type": "object",
317
+ "properties": {
318
+ "candidateId": {
319
+ "type": "string",
320
+ "minLength": 1,
321
+ "maxLength": 240
322
+ },
323
+ "frozenRevision": {
324
+ "type": "integer",
325
+ "exclusiveMinimum": 0,
326
+ "maximum": 9007199254740991
327
+ },
328
+ "frozenHash": {
329
+ "type": "string",
330
+ "pattern": "^[a-f0-9]{64}$"
331
+ },
332
+ "harnessRelease": {
333
+ "$ref": "#/$defs/__schema4"
334
+ },
335
+ "protectedClosureHash": {
336
+ "type": "string",
337
+ "pattern": "^[a-f0-9]{64}$"
338
+ }
339
+ },
340
+ "required": [
341
+ "candidateId",
342
+ "frozenRevision",
343
+ "frozenHash",
344
+ "harnessRelease",
345
+ "protectedClosureHash"
346
+ ],
347
+ "additionalProperties": false
348
+ }
349
+ },
350
+ "required": [
351
+ "profileId",
352
+ "sourceRevision",
353
+ "harnessRelease",
354
+ "catalogHash",
355
+ "definitionId",
356
+ "definitionHash",
357
+ "target",
358
+ "environmentHash"
359
+ ],
360
+ "additionalProperties": false
361
+ },
362
+ "model": {
363
+ "type": "object",
364
+ "properties": {
365
+ "modelId": {
366
+ "$ref": "#/$defs/__schema2"
367
+ },
368
+ "configurationHash": {
369
+ "$ref": "#/$defs/__schema3"
370
+ }
371
+ },
372
+ "required": [
373
+ "modelId",
374
+ "configurationHash"
375
+ ],
376
+ "additionalProperties": false
377
+ }
378
+ },
379
+ "required": [
380
+ "kind",
381
+ "source",
382
+ "model"
383
+ ],
384
+ "additionalProperties": false
385
+ },
386
+ {
387
+ "type": "object",
388
+ "properties": {
389
+ "kind": {
390
+ "type": "string",
391
+ "const": "agent"
392
+ },
393
+ "source": {
394
+ "type": "object",
395
+ "properties": {
396
+ "harnessRelease": {
397
+ "$ref": "#/$defs/__schema4"
398
+ },
399
+ "agentSnapshot": {
400
+ "$ref": "#/$defs/__schema4"
401
+ },
402
+ "sourcePackageHash": {
403
+ "$ref": "#/$defs/__schema3"
404
+ }
405
+ },
406
+ "required": [
407
+ "harnessRelease",
408
+ "agentSnapshot",
409
+ "sourcePackageHash"
410
+ ],
411
+ "additionalProperties": false
412
+ },
413
+ "model": {
414
+ "type": "object",
415
+ "properties": {
416
+ "modelId": {
417
+ "$ref": "#/$defs/__schema2"
418
+ },
419
+ "configurationHash": {
420
+ "$ref": "#/$defs/__schema3"
421
+ }
422
+ },
423
+ "required": [
424
+ "modelId",
425
+ "configurationHash"
426
+ ],
427
+ "additionalProperties": false
428
+ }
429
+ },
430
+ "required": [
431
+ "kind",
432
+ "source",
433
+ "model"
434
+ ],
435
+ "additionalProperties": false
436
+ },
437
+ {
438
+ "type": "object",
439
+ "properties": {
440
+ "kind": {
441
+ "type": "string",
442
+ "const": "fixture"
443
+ },
444
+ "configurationHash": {
445
+ "$ref": "#/$defs/__schema3"
446
+ }
447
+ },
448
+ "required": [
449
+ "kind",
450
+ "configurationHash"
451
+ ],
452
+ "additionalProperties": false
453
+ }
454
+ ]
455
+ },
456
+ "graders": {
457
+ "maxItems": 100,
458
+ "type": "array",
459
+ "items": {
460
+ "$ref": "#/$defs/__schema14"
461
+ }
462
+ },
463
+ "contentHash": {
464
+ "$ref": "#/$defs/__schema1"
465
+ }
466
+ },
467
+ "required": [
468
+ "schemaVersion",
469
+ "teamId",
470
+ "projectId",
471
+ "configurationId",
472
+ "source",
473
+ "dataset",
474
+ "target",
475
+ "graders",
476
+ "contentHash"
477
+ ],
478
+ "additionalProperties": false
479
+ },
5
480
  "schemaVersion": {
6
481
  "type": "string",
7
482
  "const": "openpond.learningPolicy.v1"
8
483
  },
9
484
  "id": {
10
- "$ref": "#/$defs/__schema0"
485
+ "$ref": "#/$defs/__schema2"
11
486
  },
12
487
  "revision": {
13
488
  "type": "integer",
@@ -15,7 +490,7 @@
15
490
  "maximum": 9007199254740991
16
491
  },
17
492
  "modelProjectId": {
18
- "$ref": "#/$defs/__schema0"
493
+ "$ref": "#/$defs/__schema2"
19
494
  },
20
495
  "executionOwner": {
21
496
  "type": "string",
@@ -32,14 +507,14 @@
32
507
  "maxItems": 100,
33
508
  "type": "array",
34
509
  "items": {
35
- "$ref": "#/$defs/__schema1"
510
+ "$ref": "#/$defs/__schema15"
36
511
  }
37
512
  },
38
513
  "taskDefinition": {
39
- "$ref": "#/$defs/__schema1"
514
+ "$ref": "#/$defs/__schema15"
40
515
  },
41
516
  "rewardBinding": {
42
- "$ref": "#/$defs/__schema1"
517
+ "$ref": "#/$defs/__schema15"
43
518
  },
44
519
  "admission": {
45
520
  "type": "object",
@@ -54,7 +529,7 @@
54
529
  "qualification": {
55
530
  "anyOf": [
56
531
  {
57
- "$ref": "#/$defs/__schema3"
532
+ "$ref": "#/$defs/__schema4"
58
533
  },
59
534
  {
60
535
  "type": "null"
@@ -153,7 +628,7 @@
153
628
  "const": "upstream_accepted"
154
629
  },
155
630
  "modelProjectId": {
156
- "$ref": "#/$defs/__schema0"
631
+ "$ref": "#/$defs/__schema2"
157
632
  }
158
633
  },
159
634
  "required": [
@@ -165,12 +640,12 @@
165
640
  ]
166
641
  },
167
642
  "trainingParent": {
168
- "$ref": "#/$defs/__schema3"
643
+ "$ref": "#/$defs/__schema4"
169
644
  },
170
645
  "teacher": {
171
646
  "anyOf": [
172
647
  {
173
- "$ref": "#/$defs/__schema3"
648
+ "$ref": "#/$defs/__schema4"
174
649
  },
175
650
  {
176
651
  "type": "null"
@@ -194,16 +669,16 @@
194
669
  ]
195
670
  },
196
671
  "recipe": {
197
- "$ref": "#/$defs/__schema3"
672
+ "$ref": "#/$defs/__schema4"
198
673
  },
199
674
  "retentionEvaluation": {
200
- "$ref": "#/$defs/__schema3"
675
+ "$ref": "#/$defs/__schema4"
201
676
  },
202
677
  "replayBatches": {
203
678
  "maxItems": 100,
204
679
  "type": "array",
205
680
  "items": {
206
- "$ref": "#/$defs/__schema3"
681
+ "$ref": "#/$defs/__schema4"
207
682
  }
208
683
  }
209
684
  },
@@ -302,7 +777,7 @@
302
777
  "rollbackVersion": {
303
778
  "anyOf": [
304
779
  {
305
- "$ref": "#/$defs/__schema3"
780
+ "$ref": "#/$defs/__schema4"
306
781
  },
307
782
  {
308
783
  "type": "null"
@@ -319,7 +794,7 @@
319
794
  "additionalProperties": false
320
795
  },
321
796
  "contentHash": {
322
- "$ref": "#/$defs/__schema2"
797
+ "$ref": "#/$defs/__schema3"
323
798
  }
324
799
  },
325
800
  "required": [
@@ -347,47 +822,380 @@
347
822
  "__schema0": {
348
823
  "type": "string",
349
824
  "minLength": 1,
350
- "maxLength": 240
825
+ "maxLength": 200
351
826
  },
352
827
  "__schema1": {
828
+ "type": "string",
829
+ "pattern": "^[a-f0-9]{64}$"
830
+ },
831
+ "__schema2": {
832
+ "type": "string",
833
+ "minLength": 1,
834
+ "maxLength": 240
835
+ },
836
+ "__schema3": {
837
+ "type": "string",
838
+ "pattern": "^[a-f0-9]{64}$"
839
+ },
840
+ "__schema4": {
353
841
  "type": "object",
354
842
  "properties": {
355
843
  "id": {
356
- "$ref": "#/$defs/__schema0"
844
+ "$ref": "#/$defs/__schema2"
357
845
  },
358
846
  "contentHash": {
847
+ "$ref": "#/$defs/__schema3"
848
+ }
849
+ },
850
+ "required": [
851
+ "id",
852
+ "contentHash"
853
+ ],
854
+ "additionalProperties": false
855
+ },
856
+ "__schema5": {
857
+ "oneOf": [
858
+ {
859
+ "type": "object",
860
+ "properties": {
861
+ "kind": {
862
+ "type": "string",
863
+ "const": "profile"
864
+ }
865
+ },
866
+ "required": [
867
+ "kind"
868
+ ],
869
+ "additionalProperties": false
870
+ },
871
+ {
872
+ "type": "object",
873
+ "properties": {
874
+ "kind": {
875
+ "type": "string",
876
+ "const": "workflow"
877
+ },
878
+ "workflowId": {
879
+ "$ref": "#/$defs/__schema2"
880
+ }
881
+ },
882
+ "required": [
883
+ "kind",
884
+ "workflowId"
885
+ ],
886
+ "additionalProperties": false
887
+ },
888
+ {
889
+ "type": "object",
890
+ "properties": {
891
+ "kind": {
892
+ "type": "string",
893
+ "const": "skill"
894
+ },
895
+ "skillPath": {
896
+ "type": "string",
897
+ "minLength": 1,
898
+ "maxLength": 2000
899
+ }
900
+ },
901
+ "required": [
902
+ "kind",
903
+ "skillPath"
904
+ ],
905
+ "additionalProperties": false
906
+ },
907
+ {
908
+ "type": "object",
909
+ "properties": {
910
+ "kind": {
911
+ "type": "string",
912
+ "const": "agent_action"
913
+ },
914
+ "actionId": {
915
+ "$ref": "#/$defs/__schema2"
916
+ }
917
+ },
918
+ "required": [
919
+ "kind",
920
+ "actionId"
921
+ ],
922
+ "additionalProperties": false
923
+ }
924
+ ]
925
+ },
926
+ "__schema6": {
927
+ "type": "object",
928
+ "properties": {
929
+ "caseId": {
359
930
  "$ref": "#/$defs/__schema2"
360
931
  },
361
- "revision": {
932
+ "seed": {
933
+ "type": "string",
934
+ "minLength": 1,
935
+ "maxLength": 500
936
+ },
937
+ "fixtureId": {
938
+ "$ref": "#/$defs/__schema3"
939
+ },
940
+ "sourceId": {
941
+ "$ref": "#/$defs/__schema2"
942
+ },
943
+ "snapshotHash": {
944
+ "$ref": "#/$defs/__schema3"
945
+ },
946
+ "boundaryRevisionHash": {
947
+ "$ref": "#/$defs/__schema3"
948
+ }
949
+ },
950
+ "required": [
951
+ "caseId",
952
+ "seed",
953
+ "fixtureId",
954
+ "sourceId",
955
+ "snapshotHash",
956
+ "boundaryRevisionHash"
957
+ ],
958
+ "additionalProperties": false
959
+ },
960
+ "__schema7": {
961
+ "type": "object",
962
+ "properties": {
963
+ "taskId": {
964
+ "$ref": "#/$defs/__schema2"
965
+ },
966
+ "seed": {
967
+ "type": "string",
968
+ "minLength": 1,
969
+ "maxLength": 500
970
+ },
971
+ "fixtureId": {
972
+ "anyOf": [
973
+ {
974
+ "$ref": "#/$defs/__schema2"
975
+ },
976
+ {
977
+ "type": "null"
978
+ }
979
+ ]
980
+ }
981
+ },
982
+ "required": [
983
+ "taskId",
984
+ "seed",
985
+ "fixtureId"
986
+ ],
987
+ "additionalProperties": false
988
+ },
989
+ "__schema8": {
990
+ "type": "object",
991
+ "properties": {
992
+ "release": {
993
+ "type": "object",
994
+ "properties": {
995
+ "id": {
996
+ "$ref": "#/$defs/__schema2"
997
+ },
998
+ "contentHash": {
999
+ "$ref": "#/$defs/__schema3"
1000
+ },
1001
+ "revision": {
1002
+ "anyOf": [
1003
+ {
1004
+ "type": "integer",
1005
+ "exclusiveMinimum": 0,
1006
+ "maximum": 9007199254740991
1007
+ },
1008
+ {
1009
+ "type": "string",
1010
+ "minLength": 1,
1011
+ "maxLength": 200
1012
+ }
1013
+ ]
1014
+ }
1015
+ },
1016
+ "required": [
1017
+ "id",
1018
+ "contentHash",
1019
+ "revision"
1020
+ ],
1021
+ "additionalProperties": false
1022
+ },
1023
+ "feedbackKey": {
1024
+ "type": "string",
1025
+ "minLength": 1,
1026
+ "maxLength": 240
1027
+ },
1028
+ "configurationHash": {
1029
+ "$ref": "#/$defs/__schema3"
1030
+ },
1031
+ "output": {
1032
+ "type": "string",
1033
+ "enum": [
1034
+ "boolean",
1035
+ "score",
1036
+ "category"
1037
+ ]
1038
+ },
1039
+ "categories": {
1040
+ "maxItems": 50,
1041
+ "type": "array",
1042
+ "items": {
1043
+ "$ref": "#/$defs/__schema9"
1044
+ }
1045
+ }
1046
+ },
1047
+ "required": [
1048
+ "release",
1049
+ "feedbackKey",
1050
+ "configurationHash",
1051
+ "output",
1052
+ "categories"
1053
+ ],
1054
+ "additionalProperties": false
1055
+ },
1056
+ "__schema9": {
1057
+ "type": "string",
1058
+ "minLength": 1,
1059
+ "maxLength": 120
1060
+ },
1061
+ "__schema10": {
1062
+ "type": "object",
1063
+ "properties": {
1064
+ "graderId": {
1065
+ "$ref": "#/$defs/__schema2"
1066
+ },
1067
+ "fields": {
1068
+ "maxItems": 100,
1069
+ "type": "array",
1070
+ "items": {
1071
+ "$ref": "#/$defs/__schema11"
1072
+ }
1073
+ }
1074
+ },
1075
+ "required": [
1076
+ "graderId",
1077
+ "fields"
1078
+ ],
1079
+ "additionalProperties": false
1080
+ },
1081
+ "__schema11": {
1082
+ "type": "object",
1083
+ "properties": {
1084
+ "destination": {
1085
+ "$ref": "#/$defs/__schema12"
1086
+ },
1087
+ "path": {
1088
+ "minItems": 1,
1089
+ "maxItems": 20,
1090
+ "type": "array",
1091
+ "items": {
1092
+ "$ref": "#/$defs/__schema13"
1093
+ }
1094
+ },
1095
+ "source": {
1096
+ "$ref": "#/$defs/__schema12"
1097
+ },
1098
+ "sourcePath": {
1099
+ "maxItems": 20,
1100
+ "type": "array",
1101
+ "items": {
1102
+ "$ref": "#/$defs/__schema13"
1103
+ }
1104
+ }
1105
+ },
1106
+ "required": [
1107
+ "destination",
1108
+ "path",
1109
+ "source",
1110
+ "sourcePath"
1111
+ ],
1112
+ "additionalProperties": false
1113
+ },
1114
+ "__schema12": {
1115
+ "type": "string",
1116
+ "enum": [
1117
+ "input",
1118
+ "output",
1119
+ "expectedOutput",
1120
+ "evaluatorContext"
1121
+ ]
1122
+ },
1123
+ "__schema13": {
1124
+ "type": "string",
1125
+ "minLength": 1,
1126
+ "maxLength": 200
1127
+ },
1128
+ "__schema14": {
1129
+ "type": "object",
1130
+ "properties": {
1131
+ "feedbackKey": {
1132
+ "$ref": "#/$defs/__schema0"
1133
+ },
1134
+ "evaluator": {
1135
+ "type": "object",
1136
+ "properties": {
1137
+ "id": {
1138
+ "$ref": "#/$defs/__schema0"
1139
+ },
1140
+ "contentHash": {
1141
+ "$ref": "#/$defs/__schema1"
1142
+ }
1143
+ },
1144
+ "required": [
1145
+ "id",
1146
+ "contentHash"
1147
+ ],
1148
+ "additionalProperties": false
1149
+ },
1150
+ "expectedCount": {
362
1151
  "type": "integer",
363
1152
  "exclusiveMinimum": 0,
364
1153
  "maximum": 9007199254740991
1154
+ },
1155
+ "scoredCount": {
1156
+ "type": "integer",
1157
+ "minimum": 0,
1158
+ "maximum": 9007199254740991
1159
+ },
1160
+ "score": {
1161
+ "anyOf": [
1162
+ {
1163
+ "type": "number"
1164
+ },
1165
+ {
1166
+ "type": "null"
1167
+ }
1168
+ ]
365
1169
  }
366
1170
  },
367
1171
  "required": [
368
- "id",
369
- "contentHash",
370
- "revision"
1172
+ "feedbackKey",
1173
+ "evaluator",
1174
+ "expectedCount",
1175
+ "scoredCount",
1176
+ "score"
371
1177
  ],
372
1178
  "additionalProperties": false
373
1179
  },
374
- "__schema2": {
375
- "type": "string",
376
- "pattern": "^[a-f0-9]{64}$"
377
- },
378
- "__schema3": {
1180
+ "__schema15": {
379
1181
  "type": "object",
380
1182
  "properties": {
381
1183
  "id": {
382
- "$ref": "#/$defs/__schema0"
1184
+ "$ref": "#/$defs/__schema2"
383
1185
  },
384
1186
  "contentHash": {
385
- "$ref": "#/$defs/__schema2"
1187
+ "$ref": "#/$defs/__schema3"
1188
+ },
1189
+ "revision": {
1190
+ "type": "integer",
1191
+ "exclusiveMinimum": 0,
1192
+ "maximum": 9007199254740991
386
1193
  }
387
1194
  },
388
1195
  "required": [
389
1196
  "id",
390
- "contentHash"
1197
+ "contentHash",
1198
+ "revision"
391
1199
  ],
392
1200
  "additionalProperties": false
393
1201
  }