@openpond/evals 0.11.16 → 0.11.18

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 +13 -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
@@ -0,0 +1,404 @@
1
+ import { z } from "zod";
2
+ export declare const ProfileExternalFieldMappingsSchema: z.ZodArray<z.ZodObject<{
3
+ graderId: z.ZodString;
4
+ fields: z.ZodArray<z.ZodObject<{
5
+ destination: z.ZodEnum<{
6
+ output: "output";
7
+ input: "input";
8
+ expectedOutput: "expectedOutput";
9
+ evaluatorContext: "evaluatorContext";
10
+ }>;
11
+ path: z.ZodArray<z.ZodString>;
12
+ source: z.ZodEnum<{
13
+ output: "output";
14
+ input: "input";
15
+ expectedOutput: "expectedOutput";
16
+ evaluatorContext: "evaluatorContext";
17
+ }>;
18
+ sourcePath: z.ZodArray<z.ZodString>;
19
+ }, z.core.$strict>>;
20
+ }, z.core.$strict>>;
21
+ export declare const ProfileExternalDatasetBindingContentSchema: z.ZodObject<{
22
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
23
+ dataset: z.ZodObject<{
24
+ id: z.ZodString;
25
+ contentHash: z.ZodString;
26
+ revision: z.ZodNumber;
27
+ }, z.core.$strict>;
28
+ packageHash: z.ZodString;
29
+ profileId: z.ZodString;
30
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
31
+ kind: z.ZodLiteral<"profile">;
32
+ }, z.core.$strict>, z.ZodObject<{
33
+ kind: z.ZodLiteral<"workflow">;
34
+ workflowId: z.ZodString;
35
+ }, z.core.$strict>, z.ZodObject<{
36
+ kind: z.ZodLiteral<"skill">;
37
+ skillPath: z.ZodString;
38
+ }, z.core.$strict>, z.ZodObject<{
39
+ kind: z.ZodLiteral<"agent_action">;
40
+ actionId: z.ZodString;
41
+ }, z.core.$strict>], "kind">;
42
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
43
+ protectedProfileClosureHash: z.ZodString;
44
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
45
+ execution: z.ZodObject<{
46
+ id: z.ZodString;
47
+ contentHash: z.ZodString;
48
+ }, z.core.$strict>;
49
+ originalEvidence: z.ZodObject<{
50
+ id: z.ZodString;
51
+ contentHash: z.ZodString;
52
+ }, z.core.$strict>;
53
+ members: z.ZodArray<z.ZodObject<{
54
+ caseId: z.ZodString;
55
+ seed: z.ZodString;
56
+ fixtureId: z.ZodString;
57
+ sourceId: z.ZodString;
58
+ snapshotHash: z.ZodString;
59
+ boundaryRevisionHash: z.ZodString;
60
+ }, z.core.$strict>>;
61
+ }, z.core.$strict>>;
62
+ gradingSource: z.ZodOptional<z.ZodObject<{
63
+ pass: z.ZodObject<{
64
+ id: z.ZodString;
65
+ contentHash: z.ZodString;
66
+ }, z.core.$strict>;
67
+ evidence: z.ZodObject<{
68
+ id: z.ZodString;
69
+ contentHash: z.ZodString;
70
+ }, z.core.$strict>;
71
+ }, z.core.$strict>>;
72
+ split: z.ZodEnum<{
73
+ train: "train";
74
+ validation: "validation";
75
+ test: "test";
76
+ frozen_eval: "frozen_eval";
77
+ }>;
78
+ population: z.ZodArray<z.ZodObject<{
79
+ taskId: z.ZodString;
80
+ seed: z.ZodString;
81
+ fixtureId: z.ZodNullable<z.ZodString>;
82
+ }, z.core.$strict>>;
83
+ evaluators: z.ZodArray<z.ZodObject<{
84
+ release: z.ZodObject<{
85
+ id: z.ZodString;
86
+ contentHash: z.ZodString;
87
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
88
+ }, z.core.$strict>;
89
+ feedbackKey: z.ZodString;
90
+ configurationHash: z.ZodString;
91
+ output: z.ZodEnum<{
92
+ boolean: "boolean";
93
+ score: "score";
94
+ category: "category";
95
+ }>;
96
+ categories: z.ZodArray<z.ZodString>;
97
+ }, z.core.$strict>>;
98
+ fieldMappingsHash: z.ZodString;
99
+ fieldMappings: z.ZodArray<z.ZodObject<{
100
+ graderId: z.ZodString;
101
+ fields: z.ZodArray<z.ZodObject<{
102
+ destination: z.ZodEnum<{
103
+ output: "output";
104
+ input: "input";
105
+ expectedOutput: "expectedOutput";
106
+ evaluatorContext: "evaluatorContext";
107
+ }>;
108
+ path: z.ZodArray<z.ZodString>;
109
+ source: z.ZodEnum<{
110
+ output: "output";
111
+ input: "input";
112
+ expectedOutput: "expectedOutput";
113
+ evaluatorContext: "evaluatorContext";
114
+ }>;
115
+ sourcePath: z.ZodArray<z.ZodString>;
116
+ }, z.core.$strict>>;
117
+ }, z.core.$strict>>;
118
+ environmentHash: z.ZodString;
119
+ privateDatasetClosureHash: z.ZodString;
120
+ criterion: z.ZodObject<{
121
+ minimumPassRate: z.ZodNumber;
122
+ requireComplete: z.ZodBoolean;
123
+ }, z.core.$strict>;
124
+ }, z.core.$strict>;
125
+ export declare const ProfileExternalDatasetBindingSchema: z.ZodObject<{
126
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
127
+ dataset: z.ZodObject<{
128
+ id: z.ZodString;
129
+ contentHash: z.ZodString;
130
+ revision: z.ZodNumber;
131
+ }, z.core.$strict>;
132
+ packageHash: z.ZodString;
133
+ profileId: z.ZodString;
134
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
135
+ kind: z.ZodLiteral<"profile">;
136
+ }, z.core.$strict>, z.ZodObject<{
137
+ kind: z.ZodLiteral<"workflow">;
138
+ workflowId: z.ZodString;
139
+ }, z.core.$strict>, z.ZodObject<{
140
+ kind: z.ZodLiteral<"skill">;
141
+ skillPath: z.ZodString;
142
+ }, z.core.$strict>, z.ZodObject<{
143
+ kind: z.ZodLiteral<"agent_action">;
144
+ actionId: z.ZodString;
145
+ }, z.core.$strict>], "kind">;
146
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
147
+ protectedProfileClosureHash: z.ZodString;
148
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
149
+ execution: z.ZodObject<{
150
+ id: z.ZodString;
151
+ contentHash: z.ZodString;
152
+ }, z.core.$strict>;
153
+ originalEvidence: z.ZodObject<{
154
+ id: z.ZodString;
155
+ contentHash: z.ZodString;
156
+ }, z.core.$strict>;
157
+ members: z.ZodArray<z.ZodObject<{
158
+ caseId: z.ZodString;
159
+ seed: z.ZodString;
160
+ fixtureId: z.ZodString;
161
+ sourceId: z.ZodString;
162
+ snapshotHash: z.ZodString;
163
+ boundaryRevisionHash: z.ZodString;
164
+ }, z.core.$strict>>;
165
+ }, z.core.$strict>>;
166
+ gradingSource: z.ZodOptional<z.ZodObject<{
167
+ pass: z.ZodObject<{
168
+ id: z.ZodString;
169
+ contentHash: z.ZodString;
170
+ }, z.core.$strict>;
171
+ evidence: z.ZodObject<{
172
+ id: z.ZodString;
173
+ contentHash: z.ZodString;
174
+ }, z.core.$strict>;
175
+ }, z.core.$strict>>;
176
+ split: z.ZodEnum<{
177
+ train: "train";
178
+ validation: "validation";
179
+ test: "test";
180
+ frozen_eval: "frozen_eval";
181
+ }>;
182
+ population: z.ZodArray<z.ZodObject<{
183
+ taskId: z.ZodString;
184
+ seed: z.ZodString;
185
+ fixtureId: z.ZodNullable<z.ZodString>;
186
+ }, z.core.$strict>>;
187
+ evaluators: z.ZodArray<z.ZodObject<{
188
+ release: z.ZodObject<{
189
+ id: z.ZodString;
190
+ contentHash: z.ZodString;
191
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
192
+ }, z.core.$strict>;
193
+ feedbackKey: z.ZodString;
194
+ configurationHash: z.ZodString;
195
+ output: z.ZodEnum<{
196
+ boolean: "boolean";
197
+ score: "score";
198
+ category: "category";
199
+ }>;
200
+ categories: z.ZodArray<z.ZodString>;
201
+ }, z.core.$strict>>;
202
+ fieldMappingsHash: z.ZodString;
203
+ fieldMappings: z.ZodArray<z.ZodObject<{
204
+ graderId: z.ZodString;
205
+ fields: z.ZodArray<z.ZodObject<{
206
+ destination: z.ZodEnum<{
207
+ output: "output";
208
+ input: "input";
209
+ expectedOutput: "expectedOutput";
210
+ evaluatorContext: "evaluatorContext";
211
+ }>;
212
+ path: z.ZodArray<z.ZodString>;
213
+ source: z.ZodEnum<{
214
+ output: "output";
215
+ input: "input";
216
+ expectedOutput: "expectedOutput";
217
+ evaluatorContext: "evaluatorContext";
218
+ }>;
219
+ sourcePath: z.ZodArray<z.ZodString>;
220
+ }, z.core.$strict>>;
221
+ }, z.core.$strict>>;
222
+ environmentHash: z.ZodString;
223
+ privateDatasetClosureHash: z.ZodString;
224
+ criterion: z.ZodObject<{
225
+ minimumPassRate: z.ZodNumber;
226
+ requireComplete: z.ZodBoolean;
227
+ }, z.core.$strict>;
228
+ contentHash: z.ZodString;
229
+ }, z.core.$strict>;
230
+ export type ProfileExternalDatasetBinding = z.infer<typeof ProfileExternalDatasetBindingSchema>;
231
+ export declare function verifyProfileExternalDatasetBinding(raw: unknown): {
232
+ schemaVersion: "openpond.profileExternalDatasetBinding.v1";
233
+ dataset: {
234
+ id: string;
235
+ contentHash: string;
236
+ revision: number;
237
+ };
238
+ packageHash: string;
239
+ profileId: string;
240
+ target: {
241
+ kind: "profile";
242
+ } | {
243
+ kind: "workflow";
244
+ workflowId: string;
245
+ } | {
246
+ kind: "skill";
247
+ skillPath: string;
248
+ } | {
249
+ kind: "agent_action";
250
+ actionId: string;
251
+ };
252
+ declaredProfileCatalogHash: string | null;
253
+ protectedProfileClosureHash: string;
254
+ split: "train" | "validation" | "test" | "frozen_eval";
255
+ population: {
256
+ taskId: string;
257
+ seed: string;
258
+ fixtureId: string | null;
259
+ }[];
260
+ evaluators: {
261
+ release: {
262
+ id: string;
263
+ contentHash: string;
264
+ revision: string | number;
265
+ };
266
+ feedbackKey: string;
267
+ configurationHash: string;
268
+ output: "boolean" | "score" | "category";
269
+ categories: string[];
270
+ }[];
271
+ fieldMappingsHash: string;
272
+ fieldMappings: {
273
+ graderId: string;
274
+ fields: {
275
+ destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
276
+ path: string[];
277
+ source: "output" | "input" | "expectedOutput" | "evaluatorContext";
278
+ sourcePath: string[];
279
+ }[];
280
+ }[];
281
+ environmentHash: string;
282
+ privateDatasetClosureHash: string;
283
+ criterion: {
284
+ minimumPassRate: number;
285
+ requireComplete: boolean;
286
+ };
287
+ contentHash: string;
288
+ recordedOrigin?: {
289
+ execution: {
290
+ id: string;
291
+ contentHash: string;
292
+ };
293
+ originalEvidence: {
294
+ id: string;
295
+ contentHash: string;
296
+ };
297
+ members: {
298
+ caseId: string;
299
+ seed: string;
300
+ fixtureId: string;
301
+ sourceId: string;
302
+ snapshotHash: string;
303
+ boundaryRevisionHash: string;
304
+ }[];
305
+ } | undefined;
306
+ gradingSource?: {
307
+ pass: {
308
+ id: string;
309
+ contentHash: string;
310
+ };
311
+ evidence: {
312
+ id: string;
313
+ contentHash: string;
314
+ };
315
+ } | undefined;
316
+ };
317
+ export declare function createProfileExternalDatasetBinding(raw: z.input<typeof ProfileExternalDatasetBindingContentSchema>): {
318
+ schemaVersion: "openpond.profileExternalDatasetBinding.v1";
319
+ dataset: {
320
+ id: string;
321
+ contentHash: string;
322
+ revision: number;
323
+ };
324
+ packageHash: string;
325
+ profileId: string;
326
+ target: {
327
+ kind: "profile";
328
+ } | {
329
+ kind: "workflow";
330
+ workflowId: string;
331
+ } | {
332
+ kind: "skill";
333
+ skillPath: string;
334
+ } | {
335
+ kind: "agent_action";
336
+ actionId: string;
337
+ };
338
+ declaredProfileCatalogHash: string | null;
339
+ protectedProfileClosureHash: string;
340
+ split: "train" | "validation" | "test" | "frozen_eval";
341
+ population: {
342
+ taskId: string;
343
+ seed: string;
344
+ fixtureId: string | null;
345
+ }[];
346
+ evaluators: {
347
+ release: {
348
+ id: string;
349
+ contentHash: string;
350
+ revision: string | number;
351
+ };
352
+ feedbackKey: string;
353
+ configurationHash: string;
354
+ output: "boolean" | "score" | "category";
355
+ categories: string[];
356
+ }[];
357
+ fieldMappingsHash: string;
358
+ fieldMappings: {
359
+ graderId: string;
360
+ fields: {
361
+ destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
362
+ path: string[];
363
+ source: "output" | "input" | "expectedOutput" | "evaluatorContext";
364
+ sourcePath: string[];
365
+ }[];
366
+ }[];
367
+ environmentHash: string;
368
+ privateDatasetClosureHash: string;
369
+ criterion: {
370
+ minimumPassRate: number;
371
+ requireComplete: boolean;
372
+ };
373
+ contentHash: string;
374
+ recordedOrigin?: {
375
+ execution: {
376
+ id: string;
377
+ contentHash: string;
378
+ };
379
+ originalEvidence: {
380
+ id: string;
381
+ contentHash: string;
382
+ };
383
+ members: {
384
+ caseId: string;
385
+ seed: string;
386
+ fixtureId: string;
387
+ sourceId: string;
388
+ snapshotHash: string;
389
+ boundaryRevisionHash: string;
390
+ }[];
391
+ } | undefined;
392
+ gradingSource?: {
393
+ pass: {
394
+ id: string;
395
+ contentHash: string;
396
+ };
397
+ evidence: {
398
+ id: string;
399
+ contentHash: string;
400
+ };
401
+ } | undefined;
402
+ };
403
+ export declare const externalDatasetDefinitionId: (binding: ProfileExternalDatasetBinding) => string;
404
+ //# sourceMappingURL=profile-external-evaluation.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"profile-external-evaluation.d.ts","sourceRoot":"","sources":["../../../../src/profile-external-evaluation.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAOxB,eAAO,MAAM,kCAAkC;;;;;;;;;;;;;;;;;;mBAAoN,CAAC;AAKpQ,eAAO,MAAM,0CAA0C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAUnD,CAAC;AACL,eAAO,MAAM,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAgG,CAAC;AACjJ,MAAM,MAAM,6BAA6B,GAAC,CAAC,CAAC,KAAK,CAAC,OAAO,mCAAmC,CAAC,CAAC;AAC9F,wBAAgB,mCAAmC,CAAC,GAAG,EAAC,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAAsM;AACrQ,wBAAgB,mCAAmC,CAAC,GAAG,EAAC,CAAC,CAAC,KAAK,CAAC,OAAO,0CAA0C,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAA2J;AAC7Q,eAAO,MAAM,2BAA2B,GAAE,SAAQ,6BAA6B,WAA6F,CAAC"}
@@ -0,0 +1,15 @@
1
+ import { z } from "zod";
2
+ /** An identity to resolve through the actual private hosted source owner. It is
3
+ * never itself an authorization or a model-artifact declaration. */
4
+ export declare const ProfileSourceCandidateSchema: z.ZodObject<{
5
+ candidateId: z.ZodString;
6
+ frozenRevision: z.ZodNumber;
7
+ frozenHash: z.ZodString;
8
+ harnessRelease: z.ZodObject<{
9
+ id: z.ZodString;
10
+ contentHash: z.ZodString;
11
+ }, z.core.$strict>;
12
+ protectedClosureHash: z.ZodString;
13
+ }, z.core.$strict>;
14
+ export type ProfileSourceCandidate = z.infer<typeof ProfileSourceCandidateSchema>;
15
+ //# sourceMappingURL=profile-source-candidate.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"profile-source-candidate.d.ts","sourceRoot":"","sources":["../../../../src/profile-source-candidate.ts"],"names":[],"mappings":"AAAA,OAAO,EAAC,CAAC,EAAC,MAAM,KAAK,CAAC;AAEtB;oEACoE;AACpE,eAAO,MAAM,4BAA4B;;;;;;;;;kBAAsP,CAAC;AAChS,MAAM,MAAM,sBAAsB,GAAC,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC"}
@@ -3,7 +3,7 @@ export declare const harnessEvaluationReviewConformance: {
3
3
  schemaVersion: "openpond.harnessEvaluationReviewReceipt.v1";
4
4
  id: string;
5
5
  ownerScope: {
6
- kind: "personal" | "team";
6
+ kind: "team" | "personal";
7
7
  id: string;
8
8
  };
9
9
  workspaceRef: string;
@@ -94,7 +94,7 @@ export declare const harnessEvaluationReviewConformance: {
94
94
  schemaVersion: "openpond.harnessEvaluationReviewReceipt.v1";
95
95
  id: string;
96
96
  ownerScope: {
97
- kind: "personal" | "team";
97
+ kind: "team" | "personal";
98
98
  id: string;
99
99
  };
100
100
  workspaceRef: string;
@@ -185,7 +185,7 @@ export declare const harnessEvaluationReviewConformance: {
185
185
  schemaVersion: "openpond.harnessEvaluationReviewReceipt.v1";
186
186
  id: string;
187
187
  ownerScope: {
188
- kind: "personal" | "team";
188
+ kind: "team" | "personal";
189
189
  id: string;
190
190
  };
191
191
  workspaceRef: string;
@@ -276,7 +276,7 @@ export declare const harnessEvaluationReviewConformance: {
276
276
  schemaVersion: "openpond.harnessEvaluationReviewReceipt.v1";
277
277
  id: string;
278
278
  ownerScope: {
279
- kind: "personal" | "team";
279
+ kind: "team" | "personal";
280
280
  id: string;
281
281
  };
282
282
  workspaceRef: string;
@@ -434,7 +434,7 @@ export declare const harnessEvaluationReviewConformance: {
434
434
  contentHash: string;
435
435
  }[];
436
436
  };
437
- decision: "no_training" | "sft" | "preference" | "rl";
437
+ decision: "preference" | "no_training" | "sft" | "rl";
438
438
  reasons: string[];
439
439
  createdAt: string;
440
440
  metadata: Record<string, unknown>;
@@ -511,7 +511,7 @@ export declare const harnessEvaluationReviewConformance: {
511
511
  contentHash: string;
512
512
  }[];
513
513
  };
514
- decision: "no_training" | "sft" | "preference" | "rl";
514
+ decision: "preference" | "no_training" | "sft" | "rl";
515
515
  reasons: string[];
516
516
  createdAt: string;
517
517
  metadata: Record<string, unknown>;
@@ -521,7 +521,7 @@ export declare const harnessEvaluationReviewConformance: {
521
521
  schemaVersion: "openpond.harnessEvaluationReviewReceipt.v1";
522
522
  id: string;
523
523
  ownerScope: {
524
- kind: "personal" | "team";
524
+ kind: "team" | "personal";
525
525
  id: string;
526
526
  };
527
527
  workspaceRef: string;