openpond-sdk 0.9.2 → 0.9.4

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 (44) hide show
  1. package/README.md +24 -0
  2. package/dist/chunk-62LWLACP.js +155 -0
  3. package/dist/chunk-62LWLACP.js.map +7 -0
  4. package/dist/chunk-EMANJDJR.js +393 -0
  5. package/dist/chunk-EMANJDJR.js.map +7 -0
  6. package/dist/{chunk-QABCKPW6.js → chunk-TQ5ASK7D.js} +4 -4
  7. package/dist/{chunk-IKTZ2CD4.js → chunk-XR6CLSCS.js} +85 -43
  8. package/dist/chunk-XR6CLSCS.js.map +7 -0
  9. package/dist/dataset-workspaces.js +3 -3
  10. package/dist/experiments.js +57 -0
  11. package/dist/experiments.js.map +7 -0
  12. package/dist/index.js +12 -3
  13. package/dist/index.js.map +2 -2
  14. package/dist/learning.js +3 -3
  15. package/dist/model-starter-attempts.js +3 -1
  16. package/dist/model-taskset-runs.js +21 -113
  17. package/dist/model-taskset-runs.js.map +4 -4
  18. package/dist/taskset-drafts.js +2 -2
  19. package/dist/taskset-packages.js +2 -2
  20. package/dist/types/packages/sdk/src/experiment-client.d.ts +2441 -0
  21. package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -0
  22. package/dist/types/packages/sdk/src/experiment-contracts.d.ts +570 -0
  23. package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -0
  24. package/dist/types/packages/sdk/src/experiment-field-mappings.d.ts +43 -0
  25. package/dist/types/packages/sdk/src/experiment-field-mappings.d.ts.map +1 -0
  26. package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +79 -0
  27. package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts.map +1 -0
  28. package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +192 -0
  29. package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts.map +1 -0
  30. package/dist/types/packages/sdk/src/experiments.d.ts +7 -0
  31. package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -0
  32. package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +44 -0
  33. package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -0
  34. package/dist/types/packages/sdk/src/index.d.ts +2 -0
  35. package/dist/types/packages/sdk/src/index.d.ts.map +1 -1
  36. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +123 -0
  37. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
  38. package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +159 -0
  39. package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
  40. package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +129 -5
  41. package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
  42. package/package.json +6 -2
  43. package/dist/chunk-IKTZ2CD4.js.map +0 -7
  44. /package/dist/{chunk-QABCKPW6.js.map → chunk-TQ5ASK7D.js.map} +0 -0
@@ -0,0 +1 @@
1
+ {"version":3,"file":"experiment-client.d.ts","sourceRoot":"","sources":["../../../../../src/experiment-client.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAGxB,OAAO,EAAiC,yBAAyB,EAAE,oBAAoB,EAAE,qBAAqB,EAChF,KAAK,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AAC9F,OAAO,EAAE,8BAA8B,EAAmC,MAAM,2BAA2B,CAAC;AAE5G,OAAO,EAA+D,KAAK,wBAAwB,EAAE,MAAM,mCAAmC,CAAC;AAG/I,qBAAa,uBAAwB,SAAQ,KAAK;IACpC,QAAQ,CAAC,MAAM,EAAE,MAAM;IAAE,QAAQ,CAAC,IAAI,EAAE,MAAM;gBAArC,MAAM,EAAE,MAAM,EAAW,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM;CAC5E;AAED,6EAA6E;AAC7E,qBAAa,yBAAyB;IAExB,OAAO,CAAC,QAAQ,CAAC,OAAO;IADpC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;gBACJ,OAAO,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,OAAO,KAAK,CAAA;KAAE;IAM/G,gFAAgF;IAC1E,cAAc,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,8BAA8B,CAAC,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAa1F,IAAI,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,oBAAoB,CAAC,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAUtE,GAAG,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,SAAS,CAAC,EAAE,uBAAuB,CAAC;QAAC,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAQ3F,IAAI,CAAC,KAAK,GAAE,CAAC,CAAC,KAAK,CAAC,OAAO,yBAAyB,CAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAUhF,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,qBAAqB,CAAC,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAexE,MAAM,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iCAuG24M,CAAC;;;;;;;;;qBAA2O,CAAC;;;;;mBAAwJ,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;0BAA6gD,CAAC;;;;;;;;;;;;;;;;;;;;;;IA9F/0Q,UAAU,CAAC,YAAY,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,OAAO,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAWzF,SAAS,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAG1C,MAAM,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAG7C,qGAAqG;IAC/F,KAAK,CAAC,EAAE,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAK3D,KAAK,CAAC,KAAK,EAAE,wBAAwB,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAM3D,WAAW,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAG5C,aAAa,CAAC,WAAW,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,OAAO,CAAC,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAS3F,iBAAiB,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAGlD,aAAa,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iCAkD64M,CAAC;;;;;;;;;qBAA2O,CAAC;;;;;mBAAwJ,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;0BAA6gD,CAAC;;;;;;;;;;;;;;;;;;;;;;IAzC/0Q,OAAO,CAAC,UAAU,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qCAyCs3M,CAAC;;;;;;;;;yBAA2O,CAAC;;;;;uBAAwJ,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;8BAA6gD,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qCAAp5D,CAAC;;;;;;;;;yBAA2O,CAAC;;;;;uBAAwJ,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;8BAA6gD,CAAC;;;;;;;;;;;;;;;;;;;;;;;;IApCr1Q,OAAO,CAAC,UAAU;YAKJ,gBAAgB;IAK9B,OAAO,CAAC,IAAI;YAKE,OAAO;CAYtB"}
@@ -0,0 +1,570 @@
1
+ import { z } from "zod";
2
+ export declare const ExperimentDefinitionRefSchema: z.ZodObject<{
3
+ id: z.ZodString;
4
+ revision: z.ZodNumber;
5
+ contentHash: z.ZodString;
6
+ }, z.core.$strict>;
7
+ export declare const ExperimentGraderPinSchema: z.ZodObject<{
8
+ id: z.ZodString;
9
+ version: z.ZodString;
10
+ contentHash: z.ZodString;
11
+ feedbackKey: z.ZodString;
12
+ release: z.ZodNullable<z.ZodObject<{
13
+ id: z.ZodString;
14
+ revision: z.ZodNumber;
15
+ contentHash: z.ZodString;
16
+ }, z.core.$strict>>;
17
+ mappings: z.ZodOptional<z.ZodArray<z.ZodObject<{
18
+ destination: z.ZodEnum<{
19
+ input: "input";
20
+ output: "output";
21
+ expectedOutput: "expectedOutput";
22
+ evaluatorContext: "evaluatorContext";
23
+ }>;
24
+ path: z.ZodArray<z.ZodString>;
25
+ source: z.ZodEnum<{
26
+ input: "input";
27
+ output: "output";
28
+ expectedOutput: "expectedOutput";
29
+ evaluatorContext: "evaluatorContext";
30
+ }>;
31
+ sourcePath: z.ZodArray<z.ZodString>;
32
+ }, z.core.$strict>>>;
33
+ }, z.core.$strict>;
34
+ export declare const ExperimentGraderSelectionSchema: z.ZodObject<{
35
+ id: z.ZodString;
36
+ version: z.ZodString;
37
+ contentHash: z.ZodString;
38
+ mappings: z.ZodArray<z.ZodObject<{
39
+ destination: z.ZodEnum<{
40
+ input: "input";
41
+ output: "output";
42
+ expectedOutput: "expectedOutput";
43
+ evaluatorContext: "evaluatorContext";
44
+ }>;
45
+ path: z.ZodArray<z.ZodString>;
46
+ source: z.ZodEnum<{
47
+ input: "input";
48
+ output: "output";
49
+ expectedOutput: "expectedOutput";
50
+ evaluatorContext: "evaluatorContext";
51
+ }>;
52
+ sourcePath: z.ZodArray<z.ZodString>;
53
+ }, z.core.$strict>>;
54
+ }, z.core.$strict>;
55
+ export declare const SaveExperimentSchema: z.ZodObject<{
56
+ operationId: z.ZodString;
57
+ id: z.ZodOptional<z.ZodString>;
58
+ expectedRevision: z.ZodNumber;
59
+ request: z.ZodObject<{
60
+ schemaVersion: z.ZodLiteral<"openpond.modelTasksetRunRequest.v1">;
61
+ operationId: z.ZodString;
62
+ teamId: z.ZodString;
63
+ modelProjectId: z.ZodNullable<z.ZodString>;
64
+ name: z.ZodOptional<z.ZodString>;
65
+ project: z.ZodOptional<z.ZodObject<{
66
+ id: z.ZodString;
67
+ revision: z.ZodNumber;
68
+ contentHash: z.ZodString;
69
+ targetId: z.ZodNullable<z.ZodString>;
70
+ }, z.core.$strict>>;
71
+ taskset: z.ZodObject<{
72
+ id: z.ZodString;
73
+ contentHash: z.ZodString;
74
+ revision: z.ZodNumber;
75
+ }, z.core.$strict>;
76
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
77
+ kind: z.ZodLiteral<"hosted_chat">;
78
+ modelId: z.ZodString;
79
+ messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
80
+ role: z.ZodEnum<{
81
+ user: "user";
82
+ system: "system";
83
+ assistant: "assistant";
84
+ }>;
85
+ content: z.ZodString;
86
+ }, z.core.$strict>>>;
87
+ maxOutputTokens: z.ZodDefault<z.ZodNumber>;
88
+ temperature: z.ZodDefault<z.ZodNumber>;
89
+ topP: z.ZodDefault<z.ZodNumber>;
90
+ }, z.core.$strict>, z.ZodObject<{
91
+ kind: z.ZodLiteral<"fixture">;
92
+ }, z.core.$strict>, z.ZodObject<{
93
+ kind: z.ZodLiteral<"hosted_harness">;
94
+ modelId: z.ZodString;
95
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
96
+ high: "high";
97
+ low: "low";
98
+ off: "off";
99
+ medium: "medium";
100
+ xhigh: "xhigh";
101
+ max: "max";
102
+ }>>;
103
+ profileRepositoryId: z.ZodString;
104
+ source: z.ZodObject<{
105
+ profileId: z.ZodString;
106
+ sourceRevision: z.ZodString;
107
+ harnessRelease: z.ZodObject<{
108
+ id: z.ZodString;
109
+ contentHash: z.ZodString;
110
+ }, z.core.$strict>;
111
+ catalogHash: z.ZodString;
112
+ definitionId: z.ZodString;
113
+ definitionHash: z.ZodString;
114
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
115
+ kind: z.ZodLiteral<"profile">;
116
+ }, z.core.$strict>, z.ZodObject<{
117
+ kind: z.ZodLiteral<"workflow">;
118
+ workflowId: z.ZodString;
119
+ }, z.core.$strict>, z.ZodObject<{
120
+ kind: z.ZodLiteral<"skill">;
121
+ skillPath: z.ZodString;
122
+ }, z.core.$strict>, z.ZodObject<{
123
+ kind: z.ZodLiteral<"agent_action">;
124
+ actionId: z.ZodString;
125
+ }, z.core.$strict>], "kind">;
126
+ environmentHash: z.ZodString;
127
+ }, z.core.$strict>;
128
+ modelConfigurationHash: z.ZodString;
129
+ packageHash: z.ZodString;
130
+ }, z.core.$strict>], "kind">;
131
+ population: z.ZodArray<z.ZodObject<{
132
+ receiptId: z.ZodString;
133
+ taskId: z.ZodString;
134
+ seed: z.ZodString;
135
+ fixtureId: z.ZodNullable<z.ZodString>;
136
+ }, z.core.$strict>>;
137
+ }, z.core.$strict>;
138
+ maximumCostUsd: z.ZodNumber;
139
+ graders: z.ZodOptional<z.ZodArray<z.ZodObject<{
140
+ id: z.ZodString;
141
+ version: z.ZodString;
142
+ contentHash: z.ZodString;
143
+ mappings: z.ZodArray<z.ZodObject<{
144
+ destination: z.ZodEnum<{
145
+ input: "input";
146
+ output: "output";
147
+ expectedOutput: "expectedOutput";
148
+ evaluatorContext: "evaluatorContext";
149
+ }>;
150
+ path: z.ZodArray<z.ZodString>;
151
+ source: z.ZodEnum<{
152
+ input: "input";
153
+ output: "output";
154
+ expectedOutput: "expectedOutput";
155
+ evaluatorContext: "evaluatorContext";
156
+ }>;
157
+ sourcePath: z.ZodArray<z.ZodString>;
158
+ }, z.core.$strict>>;
159
+ }, z.core.$strict>>>;
160
+ }, z.core.$strict>;
161
+ export declare const ExperimentDefinitionSchema: z.ZodObject<{
162
+ schemaVersion: z.ZodLiteral<"openpond.experimentDefinition.v1">;
163
+ id: z.ZodString;
164
+ teamId: z.ZodString;
165
+ revision: z.ZodNumber;
166
+ request: z.ZodObject<{
167
+ schemaVersion: z.ZodLiteral<"openpond.modelTasksetRunRequest.v1">;
168
+ operationId: z.ZodString;
169
+ teamId: z.ZodString;
170
+ modelProjectId: z.ZodNullable<z.ZodString>;
171
+ name: z.ZodOptional<z.ZodString>;
172
+ project: z.ZodOptional<z.ZodObject<{
173
+ id: z.ZodString;
174
+ revision: z.ZodNumber;
175
+ contentHash: z.ZodString;
176
+ targetId: z.ZodNullable<z.ZodString>;
177
+ }, z.core.$strict>>;
178
+ taskset: z.ZodObject<{
179
+ id: z.ZodString;
180
+ contentHash: z.ZodString;
181
+ revision: z.ZodNumber;
182
+ }, z.core.$strict>;
183
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
184
+ kind: z.ZodLiteral<"hosted_chat">;
185
+ modelId: z.ZodString;
186
+ messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
187
+ role: z.ZodEnum<{
188
+ user: "user";
189
+ system: "system";
190
+ assistant: "assistant";
191
+ }>;
192
+ content: z.ZodString;
193
+ }, z.core.$strict>>>;
194
+ maxOutputTokens: z.ZodDefault<z.ZodNumber>;
195
+ temperature: z.ZodDefault<z.ZodNumber>;
196
+ topP: z.ZodDefault<z.ZodNumber>;
197
+ }, z.core.$strict>, z.ZodObject<{
198
+ kind: z.ZodLiteral<"fixture">;
199
+ }, z.core.$strict>, z.ZodObject<{
200
+ kind: z.ZodLiteral<"hosted_harness">;
201
+ modelId: z.ZodString;
202
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
203
+ high: "high";
204
+ low: "low";
205
+ off: "off";
206
+ medium: "medium";
207
+ xhigh: "xhigh";
208
+ max: "max";
209
+ }>>;
210
+ profileRepositoryId: z.ZodString;
211
+ source: z.ZodObject<{
212
+ profileId: z.ZodString;
213
+ sourceRevision: z.ZodString;
214
+ harnessRelease: z.ZodObject<{
215
+ id: z.ZodString;
216
+ contentHash: z.ZodString;
217
+ }, z.core.$strict>;
218
+ catalogHash: z.ZodString;
219
+ definitionId: z.ZodString;
220
+ definitionHash: z.ZodString;
221
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
222
+ kind: z.ZodLiteral<"profile">;
223
+ }, z.core.$strict>, z.ZodObject<{
224
+ kind: z.ZodLiteral<"workflow">;
225
+ workflowId: z.ZodString;
226
+ }, z.core.$strict>, z.ZodObject<{
227
+ kind: z.ZodLiteral<"skill">;
228
+ skillPath: z.ZodString;
229
+ }, z.core.$strict>, z.ZodObject<{
230
+ kind: z.ZodLiteral<"agent_action">;
231
+ actionId: z.ZodString;
232
+ }, z.core.$strict>], "kind">;
233
+ environmentHash: z.ZodString;
234
+ }, z.core.$strict>;
235
+ modelConfigurationHash: z.ZodString;
236
+ packageHash: z.ZodString;
237
+ }, z.core.$strict>], "kind">;
238
+ population: z.ZodArray<z.ZodObject<{
239
+ receiptId: z.ZodString;
240
+ taskId: z.ZodString;
241
+ seed: z.ZodString;
242
+ fixtureId: z.ZodNullable<z.ZodString>;
243
+ }, z.core.$strict>>;
244
+ }, z.core.$strict>;
245
+ maximumCostUsd: z.ZodNumber;
246
+ graders: z.ZodArray<z.ZodObject<{
247
+ id: z.ZodString;
248
+ version: z.ZodString;
249
+ contentHash: z.ZodString;
250
+ feedbackKey: z.ZodString;
251
+ release: z.ZodNullable<z.ZodObject<{
252
+ id: z.ZodString;
253
+ revision: z.ZodNumber;
254
+ contentHash: z.ZodString;
255
+ }, z.core.$strict>>;
256
+ mappings: z.ZodOptional<z.ZodArray<z.ZodObject<{
257
+ destination: z.ZodEnum<{
258
+ input: "input";
259
+ output: "output";
260
+ expectedOutput: "expectedOutput";
261
+ evaluatorContext: "evaluatorContext";
262
+ }>;
263
+ path: z.ZodArray<z.ZodString>;
264
+ source: z.ZodEnum<{
265
+ input: "input";
266
+ output: "output";
267
+ expectedOutput: "expectedOutput";
268
+ evaluatorContext: "evaluatorContext";
269
+ }>;
270
+ sourcePath: z.ZodArray<z.ZodString>;
271
+ }, z.core.$strict>>>;
272
+ }, z.core.$strict>>;
273
+ createdAt: z.ZodISODateTime;
274
+ updatedAt: z.ZodISODateTime;
275
+ contentHash: z.ZodString;
276
+ }, z.core.$strict>;
277
+ export type ExperimentDefinition = z.infer<typeof ExperimentDefinitionSchema>;
278
+ export type ExperimentDefinitionRef = z.infer<typeof ExperimentDefinitionRefSchema>;
279
+ export declare const StartExperimentSchema: z.ZodObject<{
280
+ operationId: z.ZodString;
281
+ definition: z.ZodObject<{
282
+ id: z.ZodString;
283
+ revision: z.ZodNumber;
284
+ contentHash: z.ZodString;
285
+ }, z.core.$strict>;
286
+ }, z.core.$strict>;
287
+ export declare const ExperimentListQuerySchema: z.ZodObject<{
288
+ projectId: z.ZodOptional<z.ZodString>;
289
+ datasetHash: z.ZodOptional<z.ZodString>;
290
+ search: z.ZodOptional<z.ZodString>;
291
+ afterId: z.ZodOptional<z.ZodString>;
292
+ limit: z.ZodDefault<z.ZodNumber>;
293
+ }, z.core.$strict>;
294
+ export declare const ExperimentExecutionContextSchema: z.ZodObject<{
295
+ definition: z.ZodObject<{
296
+ id: z.ZodString;
297
+ revision: z.ZodNumber;
298
+ contentHash: z.ZodString;
299
+ }, z.core.$strict>;
300
+ maximumCostUsd: z.ZodNumber;
301
+ graders: z.ZodArray<z.ZodObject<{
302
+ id: z.ZodString;
303
+ version: z.ZodString;
304
+ contentHash: z.ZodString;
305
+ feedbackKey: z.ZodString;
306
+ release: z.ZodNullable<z.ZodObject<{
307
+ id: z.ZodString;
308
+ revision: z.ZodNumber;
309
+ contentHash: z.ZodString;
310
+ }, z.core.$strict>>;
311
+ mappings: z.ZodOptional<z.ZodArray<z.ZodObject<{
312
+ destination: z.ZodEnum<{
313
+ input: "input";
314
+ output: "output";
315
+ expectedOutput: "expectedOutput";
316
+ evaluatorContext: "evaluatorContext";
317
+ }>;
318
+ path: z.ZodArray<z.ZodString>;
319
+ source: z.ZodEnum<{
320
+ input: "input";
321
+ output: "output";
322
+ expectedOutput: "expectedOutput";
323
+ evaluatorContext: "evaluatorContext";
324
+ }>;
325
+ sourcePath: z.ZodArray<z.ZodString>;
326
+ }, z.core.$strict>>>;
327
+ }, z.core.$strict>>;
328
+ }, z.core.$strict>;
329
+ export type ExperimentExecutionContext = z.infer<typeof ExperimentExecutionContextSchema>;
330
+ export declare const PrepareHarnessExperimentSchema: z.ZodObject<{
331
+ operationId: z.ZodString;
332
+ profileRepositoryId: z.ZodString;
333
+ definitionId: z.ZodString;
334
+ modelId: z.ZodString;
335
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
336
+ high: "high";
337
+ low: "low";
338
+ off: "off";
339
+ medium: "medium";
340
+ xhigh: "xhigh";
341
+ max: "max";
342
+ }>>;
343
+ maximumCostUsd: z.ZodNumber;
344
+ }, z.core.$strict>;
345
+ export declare const PreparedHarnessExperimentSchema: z.ZodObject<{
346
+ request: z.ZodObject<{
347
+ schemaVersion: z.ZodLiteral<"openpond.modelTasksetRunRequest.v1">;
348
+ operationId: z.ZodString;
349
+ teamId: z.ZodString;
350
+ modelProjectId: z.ZodNullable<z.ZodString>;
351
+ name: z.ZodOptional<z.ZodString>;
352
+ project: z.ZodOptional<z.ZodObject<{
353
+ id: z.ZodString;
354
+ revision: z.ZodNumber;
355
+ contentHash: z.ZodString;
356
+ targetId: z.ZodNullable<z.ZodString>;
357
+ }, z.core.$strict>>;
358
+ taskset: z.ZodObject<{
359
+ id: z.ZodString;
360
+ contentHash: z.ZodString;
361
+ revision: z.ZodNumber;
362
+ }, z.core.$strict>;
363
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
364
+ kind: z.ZodLiteral<"hosted_chat">;
365
+ modelId: z.ZodString;
366
+ messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
367
+ role: z.ZodEnum<{
368
+ user: "user";
369
+ system: "system";
370
+ assistant: "assistant";
371
+ }>;
372
+ content: z.ZodString;
373
+ }, z.core.$strict>>>;
374
+ maxOutputTokens: z.ZodDefault<z.ZodNumber>;
375
+ temperature: z.ZodDefault<z.ZodNumber>;
376
+ topP: z.ZodDefault<z.ZodNumber>;
377
+ }, z.core.$strict>, z.ZodObject<{
378
+ kind: z.ZodLiteral<"fixture">;
379
+ }, z.core.$strict>, z.ZodObject<{
380
+ kind: z.ZodLiteral<"hosted_harness">;
381
+ modelId: z.ZodString;
382
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
383
+ high: "high";
384
+ low: "low";
385
+ off: "off";
386
+ medium: "medium";
387
+ xhigh: "xhigh";
388
+ max: "max";
389
+ }>>;
390
+ profileRepositoryId: z.ZodString;
391
+ source: z.ZodObject<{
392
+ profileId: z.ZodString;
393
+ sourceRevision: z.ZodString;
394
+ harnessRelease: z.ZodObject<{
395
+ id: z.ZodString;
396
+ contentHash: z.ZodString;
397
+ }, z.core.$strict>;
398
+ catalogHash: z.ZodString;
399
+ definitionId: z.ZodString;
400
+ definitionHash: z.ZodString;
401
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
402
+ kind: z.ZodLiteral<"profile">;
403
+ }, z.core.$strict>, z.ZodObject<{
404
+ kind: z.ZodLiteral<"workflow">;
405
+ workflowId: z.ZodString;
406
+ }, z.core.$strict>, z.ZodObject<{
407
+ kind: z.ZodLiteral<"skill">;
408
+ skillPath: z.ZodString;
409
+ }, z.core.$strict>, z.ZodObject<{
410
+ kind: z.ZodLiteral<"agent_action">;
411
+ actionId: z.ZodString;
412
+ }, z.core.$strict>], "kind">;
413
+ environmentHash: z.ZodString;
414
+ }, z.core.$strict>;
415
+ modelConfigurationHash: z.ZodString;
416
+ packageHash: z.ZodString;
417
+ }, z.core.$strict>], "kind">;
418
+ population: z.ZodArray<z.ZodObject<{
419
+ receiptId: z.ZodString;
420
+ taskId: z.ZodString;
421
+ seed: z.ZodString;
422
+ fixtureId: z.ZodNullable<z.ZodString>;
423
+ }, z.core.$strict>>;
424
+ }, z.core.$strict>;
425
+ manifest: z.ZodObject<{
426
+ schemaVersion: z.ZodLiteral<"openpond.tasksetRunManifest.v1">;
427
+ id: z.ZodString;
428
+ tasksetRelease: z.ZodObject<{
429
+ id: z.ZodString;
430
+ contentHash: z.ZodString;
431
+ }, z.core.$strict>;
432
+ packageHash: z.ZodString;
433
+ execution: z.ZodDiscriminatedUnion<[z.ZodObject<{
434
+ kind: z.ZodLiteral<"taskset">;
435
+ environmentRelease: z.ZodObject<{
436
+ id: z.ZodString;
437
+ contentHash: z.ZodString;
438
+ }, z.core.$strict>;
439
+ verifierSetRelease: z.ZodObject<{
440
+ id: z.ZodString;
441
+ contentHash: z.ZodString;
442
+ }, z.core.$strict>;
443
+ policyHash: z.ZodString;
444
+ }, z.core.$strict>, z.ZodObject<{
445
+ kind: z.ZodLiteral<"harness">;
446
+ harnessRelease: z.ZodObject<{
447
+ id: z.ZodString;
448
+ contentHash: z.ZodString;
449
+ }, z.core.$strict>;
450
+ }, z.core.$strict>], "kind">;
451
+ profileEvaluation: z.ZodOptional<z.ZodObject<{
452
+ profileId: z.ZodString;
453
+ sourceRevision: z.ZodString;
454
+ harnessRelease: z.ZodObject<{
455
+ id: z.ZodString;
456
+ contentHash: z.ZodString;
457
+ }, z.core.$strict>;
458
+ catalogHash: z.ZodString;
459
+ definitionId: z.ZodString;
460
+ definitionHash: z.ZodString;
461
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
462
+ kind: z.ZodLiteral<"profile">;
463
+ }, z.core.$strict>, z.ZodObject<{
464
+ kind: z.ZodLiteral<"workflow">;
465
+ workflowId: z.ZodString;
466
+ }, z.core.$strict>, z.ZodObject<{
467
+ kind: z.ZodLiteral<"skill">;
468
+ skillPath: z.ZodString;
469
+ }, z.core.$strict>, z.ZodObject<{
470
+ kind: z.ZodLiteral<"agent_action">;
471
+ actionId: z.ZodString;
472
+ }, z.core.$strict>], "kind">;
473
+ environmentHash: z.ZodString;
474
+ }, z.core.$strict>>;
475
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
476
+ kind: z.ZodLiteral<"model">;
477
+ model: z.ZodObject<{
478
+ provider: z.ZodString;
479
+ model: z.ZodString;
480
+ revision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
481
+ artifactHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
482
+ tokenizerRevision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
483
+ chatTemplateHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
484
+ }, z.core.$strict>;
485
+ configurationHash: z.ZodString;
486
+ }, z.core.$strict>, z.ZodObject<{
487
+ kind: z.ZodLiteral<"fixture">;
488
+ }, z.core.$strict>], "kind">;
489
+ gradingRole: z.ZodLiteral<"evaluation">;
490
+ metricPolicy: z.ZodObject<{
491
+ schemaVersion: z.ZodLiteral<"openpond.tasksetMetricPolicy.v1">;
492
+ primaryMetric: z.ZodString;
493
+ aggregation: z.ZodEnum<{
494
+ mean_score: "mean_score";
495
+ pass_rate: "pass_rate";
496
+ weighted_mean: "weighted_mean";
497
+ custom: "custom";
498
+ }>;
499
+ missingReward: z.ZodEnum<{
500
+ zero: "zero";
501
+ exclude: "exclude";
502
+ }>;
503
+ taskWeights: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodNumber>>;
504
+ customAggregator: z.ZodNullable<z.ZodObject<{
505
+ module: z.ZodString;
506
+ exportName: z.ZodString;
507
+ contentHash: z.ZodString;
508
+ timeoutMs: z.ZodNumber;
509
+ networkPolicy: z.ZodLiteral<"none">;
510
+ }, z.core.$strict>>;
511
+ }, z.core.$strict>;
512
+ population: z.ZodArray<z.ZodObject<{
513
+ receiptId: z.ZodString;
514
+ taskId: z.ZodString;
515
+ seed: z.ZodString;
516
+ fixtureId: z.ZodNullable<z.ZodString>;
517
+ }, z.core.$strict>>;
518
+ runtimeTarget: z.ZodObject<{
519
+ adapterId: z.ZodString;
520
+ placement: z.ZodEnum<{
521
+ local: "local";
522
+ remote: "remote";
523
+ colocated: "colocated";
524
+ provider_native: "provider_native";
525
+ }>;
526
+ runtimeVersion: z.ZodString;
527
+ capabilityReceipt: z.ZodString;
528
+ }, z.core.$strict>;
529
+ limits: z.ZodObject<{
530
+ maxTurns: z.ZodNumber;
531
+ timeoutMs: z.ZodNumber;
532
+ maxOutputBytes: z.ZodNumber;
533
+ maximumSpendUsd: z.ZodNullable<z.ZodNumber>;
534
+ }, z.core.$strict>;
535
+ createdAt: z.ZodString;
536
+ metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
537
+ contentHash: z.ZodString;
538
+ }, z.core.$strict>;
539
+ maximumCostUsd: z.ZodNumber;
540
+ graders: z.ZodArray<z.ZodObject<{
541
+ id: z.ZodString;
542
+ version: z.ZodString;
543
+ contentHash: z.ZodString;
544
+ feedbackKey: z.ZodString;
545
+ release: z.ZodNullable<z.ZodObject<{
546
+ id: z.ZodString;
547
+ revision: z.ZodNumber;
548
+ contentHash: z.ZodString;
549
+ }, z.core.$strict>>;
550
+ mappings: z.ZodOptional<z.ZodArray<z.ZodObject<{
551
+ destination: z.ZodEnum<{
552
+ input: "input";
553
+ output: "output";
554
+ expectedOutput: "expectedOutput";
555
+ evaluatorContext: "evaluatorContext";
556
+ }>;
557
+ path: z.ZodArray<z.ZodString>;
558
+ source: z.ZodEnum<{
559
+ input: "input";
560
+ output: "output";
561
+ expectedOutput: "expectedOutput";
562
+ evaluatorContext: "evaluatorContext";
563
+ }>;
564
+ sourcePath: z.ZodArray<z.ZodString>;
565
+ }, z.core.$strict>>>;
566
+ }, z.core.$strict>>;
567
+ }, z.core.$strict>;
568
+ export declare function verifyExperimentDefinition(value: unknown): ExperimentDefinition;
569
+ export declare const experimentDefinitionRef: (value: ExperimentDefinition) => ExperimentDefinitionRef;
570
+ //# sourceMappingURL=experiment-contracts.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"experiment-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/experiment-contracts.ts"],"names":[],"mappings":"AAIA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAKxB,eAAO,MAAM,6BAA6B;;;;kBAE/B,CAAC;AACZ,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK3B,CAAC;AACZ,eAAO,MAAM,+BAA+B;;;;;;;;;;;;;;;;;;;;kBAAiI,CAAC;AAC9K,eAAO,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAW/B,CAAC;AASH,eAAO,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAkD,CAAC;AAC1F,MAAM,MAAM,oBAAoB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,0BAA0B,CAAC,CAAC;AAC9E,MAAM,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,6BAA6B,CAAC,CAAC;AACpF,eAAO,MAAM,qBAAqB;;;;;;;kBAAoF,CAAC;AACvH,eAAO,MAAM,yBAAyB;;;;;;kBAG3B,CAAC;AACZ,eAAO,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAGlC,CAAC;AACZ,MAAM,MAAM,0BAA0B,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,gCAAgC,CAAC,CAAC;AAE1F,eAAO,MAAM,8BAA8B;;;;;;;;;;;;;;kBAIhC,CAAC;AACZ,eAAO,MAAM,+BAA+B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAIjC,CAAC;AAEZ,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,oBAAoB,CAQ/E;AAED,eAAO,MAAM,uBAAuB,GAAI,OAAO,oBAAoB,KAAG,uBAEpE,CAAC"}
@@ -0,0 +1,43 @@
1
+ import { z } from "zod";
2
+ export declare const ExperimentFieldMappingSchema: z.ZodObject<{
3
+ destination: z.ZodEnum<{
4
+ input: "input";
5
+ output: "output";
6
+ expectedOutput: "expectedOutput";
7
+ evaluatorContext: "evaluatorContext";
8
+ }>;
9
+ path: z.ZodArray<z.ZodString>;
10
+ source: z.ZodEnum<{
11
+ input: "input";
12
+ output: "output";
13
+ expectedOutput: "expectedOutput";
14
+ evaluatorContext: "evaluatorContext";
15
+ }>;
16
+ sourcePath: z.ZodArray<z.ZodString>;
17
+ }, z.core.$strict>;
18
+ export declare const ExperimentFieldMappingsSchema: z.ZodArray<z.ZodObject<{
19
+ destination: z.ZodEnum<{
20
+ input: "input";
21
+ output: "output";
22
+ expectedOutput: "expectedOutput";
23
+ evaluatorContext: "evaluatorContext";
24
+ }>;
25
+ path: z.ZodArray<z.ZodString>;
26
+ source: z.ZodEnum<{
27
+ input: "input";
28
+ output: "output";
29
+ expectedOutput: "expectedOutput";
30
+ evaluatorContext: "evaluatorContext";
31
+ }>;
32
+ sourcePath: z.ZodArray<z.ZodString>;
33
+ }, z.core.$strict>>;
34
+ export type ExperimentFieldMapping = z.infer<typeof ExperimentFieldMappingSchema>;
35
+ export type ExperimentGradingFields = {
36
+ input: Record<string, unknown>;
37
+ output: Record<string, unknown>;
38
+ expectedOutput: Record<string, unknown> | null;
39
+ evaluatorContext: Record<string, unknown> | null;
40
+ };
41
+ /** Map only grader evidence. The caller must never use this projection as a policy input. */
42
+ export declare function mapExperimentGradingFields(fields: ExperimentGradingFields, raw: readonly ExperimentFieldMapping[]): ExperimentGradingFields;
43
+ //# sourceMappingURL=experiment-field-mappings.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"experiment-field-mappings.d.ts","sourceRoot":"","sources":["../../../../../src/experiment-field-mappings.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAKxB,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;kBAE9B,CAAC;AACZ,eAAO,MAAM,6BAA6B;;;;;;;;;;;;;;;mBAMxC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,MAAM,MAAM,uBAAuB,GAAG;IAAE,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAAC,cAAc,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;IAAC,gBAAgB,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI,CAAA;CAAE,CAAC;AAE5M,6FAA6F;AAC7F,wBAAgB,0BAA0B,CAAC,MAAM,EAAE,uBAAuB,EAAE,GAAG,EAAE,SAAS,sBAAsB,EAAE,GAAG,uBAAuB,CAqB3I"}