openpond-sdk 0.10.0 → 0.11.0
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.
- package/dist/advanced-refiner-evaluations.js +42 -0
- package/dist/advanced-refiner-evaluations.js.map +7 -0
- package/dist/candidate-evaluations.js +25 -0
- package/dist/candidate-evaluations.js.map +7 -0
- package/dist/{chunk-W5TXWVFS.js → chunk-2JKV5OJO.js} +88 -59
- package/dist/chunk-2JKV5OJO.js.map +7 -0
- package/dist/{chunk-2ZJRG3KL.js → chunk-2NSNBED7.js} +7 -3
- package/dist/chunk-2NSNBED7.js.map +7 -0
- package/dist/chunk-5RZAPZWI.js +445 -0
- package/dist/chunk-5RZAPZWI.js.map +7 -0
- package/dist/{chunk-JCPLAIZL.js → chunk-5U25KAJM.js} +106 -76
- package/dist/chunk-5U25KAJM.js.map +7 -0
- package/dist/chunk-7Q7I6FAT.js +139 -0
- package/dist/chunk-7Q7I6FAT.js.map +7 -0
- package/dist/chunk-A64VYVXN.js +64 -0
- package/dist/chunk-A64VYVXN.js.map +7 -0
- package/dist/{chunk-DT7FRNIZ.js → chunk-CSWVE3E6.js} +2 -2
- package/dist/chunk-DJWHN7NU.js +394 -0
- package/dist/chunk-DJWHN7NU.js.map +7 -0
- package/dist/chunk-GDJRGRHA.js +144 -0
- package/dist/chunk-GDJRGRHA.js.map +7 -0
- package/dist/chunk-HOAYWW33.js +587 -0
- package/dist/chunk-HOAYWW33.js.map +7 -0
- package/dist/{chunk-Y4NGWGTA.js → chunk-I2MBVOVX.js} +265 -264
- package/dist/chunk-I2MBVOVX.js.map +7 -0
- package/dist/{chunk-I3TQTXKB.js → chunk-IXBSFJCX.js} +4 -4
- package/dist/chunk-JMQ4HBWP.js +160 -0
- package/dist/chunk-JMQ4HBWP.js.map +7 -0
- package/dist/chunk-KQWCXW3A.js +294 -0
- package/dist/chunk-KQWCXW3A.js.map +7 -0
- package/dist/{chunk-I2YM3UY4.js → chunk-NHCGOFQW.js} +2 -2
- package/dist/{chunk-32H2EJRA.js → chunk-R3M6BIOK.js} +2 -2
- package/dist/chunk-SLS3GFPG.js +125 -0
- package/dist/chunk-SLS3GFPG.js.map +7 -0
- package/dist/chunk-SVU4RZ4K.js +70 -0
- package/dist/chunk-SVU4RZ4K.js.map +7 -0
- package/dist/chunk-T4APHDES.js +74 -0
- package/dist/chunk-T4APHDES.js.map +7 -0
- package/dist/chunk-VC5OV5AS.js +174 -0
- package/dist/chunk-VC5OV5AS.js.map +7 -0
- package/dist/{chunk-7B44TQZD.js → chunk-WX22WN33.js} +13 -7
- package/dist/chunk-WX22WN33.js.map +7 -0
- package/dist/chunk-XEUOJ5WF.js +31 -0
- package/dist/chunk-XEUOJ5WF.js.map +7 -0
- package/dist/{chunk-LGEF4S5I.js → chunk-XJDWHTH5.js} +390 -362
- package/dist/chunk-XJDWHTH5.js.map +7 -0
- package/dist/chunk-YCQVVFE5.js +37 -0
- package/dist/chunk-YCQVVFE5.js.map +7 -0
- package/dist/chunk-ZVJU5CJ5.js +267 -0
- package/dist/chunk-ZVJU5CJ5.js.map +7 -0
- package/dist/connected-evidence.js +81 -0
- package/dist/connected-evidence.js.map +7 -0
- package/dist/dataset-workspaces.js +44 -79
- package/dist/dataset-workspaces.js.map +4 -4
- package/dist/experiment-evaluation-schedules.js +48 -0
- package/dist/experiment-evaluation-schedules.js.map +7 -0
- package/dist/experiment-improvements.js +52 -0
- package/dist/experiment-improvements.js.map +7 -0
- package/dist/experiments.js +48 -90
- package/dist/experiments.js.map +3 -3
- package/dist/human-review.js +13 -0
- package/dist/human-review.js.map +7 -0
- package/dist/index.js +208 -17
- package/dist/index.js.map +2 -2
- package/dist/learning.js +1 -1
- package/dist/model-starter-attempts.js +2 -2
- package/dist/model-starter-catalog.js +7 -2
- package/dist/model-starter-catalog.js.map +1 -1
- package/dist/model-starters.js +7 -2
- package/dist/model-taskset-runs.js +3 -3
- package/dist/post-training.js +47 -0
- package/dist/post-training.js.map +7 -0
- package/dist/taskset-drafts.js +11 -4
- package/dist/taskset-drafts.js.map +2 -2
- package/dist/taskset-packages.js +26 -556
- package/dist/taskset-packages.js.map +4 -4
- package/dist/training-bundle.js +58 -17
- package/dist/training-bundle.js.map +2 -2
- package/dist/training-projects.js +1 -1
- package/dist/training.js +7 -2
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts +2673 -0
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts +1079 -0
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +845 -0
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/claude-code-runtime.d.ts +51 -0
- package/dist/types/packages/sdk/src/claude-code-runtime.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-dataset-publication.d.ts +15 -0
- package/dist/types/packages/sdk/src/connected-dataset-publication.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-dataset-source.d.ts +48 -0
- package/dist/types/packages/sdk/src/connected-dataset-source.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts +2460 -0
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts +215 -0
- package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence-http.d.ts +6 -0
- package/dist/types/packages/sdk/src/connected-evidence-http.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence.d.ts +6 -0
- package/dist/types/packages/sdk/src/connected-evidence.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts +690 -0
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +771 -120
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +278 -38
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +280 -40
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts +17 -0
- package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-client.d.ts +3513 -52
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +1017 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts +2780 -0
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +8 -0
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts +5724 -0
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts +13024 -0
- package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts +1361 -0
- package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-improvements.d.ts +4 -0
- package/dist/types/packages/sdk/src/experiment-improvements.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts +262 -0
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts +1388 -1
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +5 -2
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiments.d.ts +3 -0
- package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/grader-inspection.d.ts +307 -5
- package/dist/types/packages/sdk/src/grader-inspection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +238 -0
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/human-review-client.d.ts +1267 -0
- package/dist/types/packages/sdk/src/human-review-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/human-review.d.ts +3 -0
- package/dist/types/packages/sdk/src/human-review.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/index.d.ts +10 -0
- package/dist/types/packages/sdk/src/index.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +181 -0
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts +20 -4
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +79 -5
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review.d.ts +15 -13
- package/dist/types/packages/sdk/src/model-batch-review.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts +359 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +288 -19
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +234 -12
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts +78 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starters.d.ts +466 -13
- package/dist/types/packages/sdk/src/model-starters.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts +312 -4
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts +78 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +1282 -4
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +1285 -5
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/post-training.d.ts +3518 -0
- package/dist/types/packages/sdk/src/post-training.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts +1127 -0
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts +1369 -0
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +142 -22
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +111 -18
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +278 -38
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +661 -103
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +217 -20
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +12 -12
- package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +17 -17
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +140 -20
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +638 -42
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +249 -36
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +278 -38
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-drafts.d.ts +1 -0
- package/dist/types/packages/sdk/src/taskset-drafts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts +1322 -26
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +660 -12
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +77 -3
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +29 -11
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +430 -39
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-acceptance.d.ts +10 -10
- package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts +2 -2
- package/dist/types/packages/sdk/src/training-evaluation-results.d.ts +1 -1
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts +376 -42
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-preparations.d.ts +3930 -0
- package/dist/types/packages/sdk/src/training-preparations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-client.d.ts +776 -8
- package/dist/types/packages/sdk/src/training-project-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts +1161 -11
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training.d.ts +2655 -24
- package/dist/types/packages/sdk/src/training.d.ts.map +1 -1
- package/package.json +30 -2
- package/dist/chunk-2ZJRG3KL.js.map +0 -7
- package/dist/chunk-7B44TQZD.js.map +0 -7
- package/dist/chunk-JCPLAIZL.js.map +0 -7
- package/dist/chunk-LGEF4S5I.js.map +0 -7
- package/dist/chunk-PRRUBH2B.js +0 -314
- package/dist/chunk-PRRUBH2B.js.map +0 -7
- package/dist/chunk-W5TXWVFS.js.map +0 -7
- package/dist/chunk-Y4NGWGTA.js.map +0 -7
- /package/dist/{chunk-DT7FRNIZ.js.map → chunk-CSWVE3E6.js.map} +0 -0
- /package/dist/{chunk-I3TQTXKB.js.map → chunk-IXBSFJCX.js.map} +0 -0
- /package/dist/{chunk-I2YM3UY4.js.map → chunk-NHCGOFQW.js.map} +0 -0
- /package/dist/{chunk-32H2EJRA.js.map → chunk-R3M6BIOK.js.map} +0 -0
|
@@ -24,7 +24,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
24
24
|
};
|
|
25
25
|
environment: {
|
|
26
26
|
protocolVersion: "openpond.environment.v1";
|
|
27
|
-
kind: "
|
|
27
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
28
28
|
entrypoint: string;
|
|
29
29
|
stateful: boolean;
|
|
30
30
|
deterministicSeeds: boolean;
|
|
@@ -85,7 +85,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
85
85
|
hardGate: boolean;
|
|
86
86
|
rewardEligible: boolean;
|
|
87
87
|
privileged: boolean;
|
|
88
|
-
kind: "content" | "state" | "
|
|
88
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
89
89
|
config: Record<string, unknown>;
|
|
90
90
|
} | {
|
|
91
91
|
id: string;
|
|
@@ -103,7 +103,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
103
103
|
mediaType: string;
|
|
104
104
|
visibility: "policy" | "verifier" | "host_private";
|
|
105
105
|
};
|
|
106
|
-
calibrationStatus: "passed" | "
|
|
106
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
107
107
|
model?: {
|
|
108
108
|
providerId: string;
|
|
109
109
|
modelId: string;
|
|
@@ -147,6 +147,80 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
147
147
|
visibility: "policy" | "verifier" | "host_private";
|
|
148
148
|
};
|
|
149
149
|
reviewerRole: string;
|
|
150
|
+
form?: {
|
|
151
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
152
|
+
mode: "individual" | "pairwise";
|
|
153
|
+
instructions: string;
|
|
154
|
+
criteria: ({
|
|
155
|
+
id: string;
|
|
156
|
+
label: string;
|
|
157
|
+
instructions: string;
|
|
158
|
+
required: boolean;
|
|
159
|
+
allowAbstain: boolean;
|
|
160
|
+
kind: "score";
|
|
161
|
+
minimum: number;
|
|
162
|
+
maximum: number;
|
|
163
|
+
step: number;
|
|
164
|
+
requireRationale?: boolean | undefined;
|
|
165
|
+
anchors?: {
|
|
166
|
+
value: string | number | boolean;
|
|
167
|
+
description: string;
|
|
168
|
+
}[] | undefined;
|
|
169
|
+
} | {
|
|
170
|
+
id: string;
|
|
171
|
+
label: string;
|
|
172
|
+
instructions: string;
|
|
173
|
+
required: boolean;
|
|
174
|
+
allowAbstain: boolean;
|
|
175
|
+
kind: "category";
|
|
176
|
+
options: string[];
|
|
177
|
+
requireRationale?: boolean | undefined;
|
|
178
|
+
anchors?: {
|
|
179
|
+
value: string | number | boolean;
|
|
180
|
+
description: string;
|
|
181
|
+
}[] | undefined;
|
|
182
|
+
} | {
|
|
183
|
+
id: string;
|
|
184
|
+
label: string;
|
|
185
|
+
instructions: string;
|
|
186
|
+
required: boolean;
|
|
187
|
+
allowAbstain: boolean;
|
|
188
|
+
kind: "boolean";
|
|
189
|
+
requireRationale?: boolean | undefined;
|
|
190
|
+
anchors?: {
|
|
191
|
+
value: string | number | boolean;
|
|
192
|
+
description: string;
|
|
193
|
+
}[] | undefined;
|
|
194
|
+
} | {
|
|
195
|
+
id: string;
|
|
196
|
+
label: string;
|
|
197
|
+
instructions: string;
|
|
198
|
+
required: boolean;
|
|
199
|
+
allowAbstain: boolean;
|
|
200
|
+
kind: "text";
|
|
201
|
+
maximumLength: number;
|
|
202
|
+
requireRationale?: boolean | undefined;
|
|
203
|
+
anchors?: {
|
|
204
|
+
value: string | number | boolean;
|
|
205
|
+
description: string;
|
|
206
|
+
}[] | undefined;
|
|
207
|
+
minimumLength?: number | undefined;
|
|
208
|
+
} | {
|
|
209
|
+
id: string;
|
|
210
|
+
label: string;
|
|
211
|
+
instructions: string;
|
|
212
|
+
required: boolean;
|
|
213
|
+
allowAbstain: boolean;
|
|
214
|
+
kind: "preference";
|
|
215
|
+
allowTie: boolean;
|
|
216
|
+
requireRationale?: boolean | undefined;
|
|
217
|
+
anchors?: {
|
|
218
|
+
value: string | number | boolean;
|
|
219
|
+
description: string;
|
|
220
|
+
}[] | undefined;
|
|
221
|
+
requireStrength?: boolean | undefined;
|
|
222
|
+
})[];
|
|
223
|
+
} | undefined;
|
|
150
224
|
})[];
|
|
151
225
|
metadata: Record<string, unknown>;
|
|
152
226
|
contentHash: string;
|
|
@@ -179,7 +253,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
179
253
|
revision: number;
|
|
180
254
|
contract: {
|
|
181
255
|
protocolVersion: "openpond.environment.v1";
|
|
182
|
-
kind: "
|
|
256
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
183
257
|
entrypoint: string;
|
|
184
258
|
stateful: boolean;
|
|
185
259
|
deterministicSeeds: boolean;
|
|
@@ -230,7 +304,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
230
304
|
hardGate: boolean;
|
|
231
305
|
rewardEligible: boolean;
|
|
232
306
|
privileged: boolean;
|
|
233
|
-
kind: "content" | "state" | "
|
|
307
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
234
308
|
config: Record<string, unknown>;
|
|
235
309
|
} | {
|
|
236
310
|
id: string;
|
|
@@ -248,7 +322,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
248
322
|
mediaType: string;
|
|
249
323
|
visibility: "policy" | "verifier" | "host_private";
|
|
250
324
|
};
|
|
251
|
-
calibrationStatus: "passed" | "
|
|
325
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
252
326
|
model?: {
|
|
253
327
|
providerId: string;
|
|
254
328
|
modelId: string;
|
|
@@ -292,6 +366,80 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
292
366
|
visibility: "policy" | "verifier" | "host_private";
|
|
293
367
|
};
|
|
294
368
|
reviewerRole: string;
|
|
369
|
+
form?: {
|
|
370
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
371
|
+
mode: "individual" | "pairwise";
|
|
372
|
+
instructions: string;
|
|
373
|
+
criteria: ({
|
|
374
|
+
id: string;
|
|
375
|
+
label: string;
|
|
376
|
+
instructions: string;
|
|
377
|
+
required: boolean;
|
|
378
|
+
allowAbstain: boolean;
|
|
379
|
+
kind: "score";
|
|
380
|
+
minimum: number;
|
|
381
|
+
maximum: number;
|
|
382
|
+
step: number;
|
|
383
|
+
requireRationale?: boolean | undefined;
|
|
384
|
+
anchors?: {
|
|
385
|
+
value: string | number | boolean;
|
|
386
|
+
description: string;
|
|
387
|
+
}[] | undefined;
|
|
388
|
+
} | {
|
|
389
|
+
id: string;
|
|
390
|
+
label: string;
|
|
391
|
+
instructions: string;
|
|
392
|
+
required: boolean;
|
|
393
|
+
allowAbstain: boolean;
|
|
394
|
+
kind: "category";
|
|
395
|
+
options: string[];
|
|
396
|
+
requireRationale?: boolean | undefined;
|
|
397
|
+
anchors?: {
|
|
398
|
+
value: string | number | boolean;
|
|
399
|
+
description: string;
|
|
400
|
+
}[] | undefined;
|
|
401
|
+
} | {
|
|
402
|
+
id: string;
|
|
403
|
+
label: string;
|
|
404
|
+
instructions: string;
|
|
405
|
+
required: boolean;
|
|
406
|
+
allowAbstain: boolean;
|
|
407
|
+
kind: "boolean";
|
|
408
|
+
requireRationale?: boolean | undefined;
|
|
409
|
+
anchors?: {
|
|
410
|
+
value: string | number | boolean;
|
|
411
|
+
description: string;
|
|
412
|
+
}[] | undefined;
|
|
413
|
+
} | {
|
|
414
|
+
id: string;
|
|
415
|
+
label: string;
|
|
416
|
+
instructions: string;
|
|
417
|
+
required: boolean;
|
|
418
|
+
allowAbstain: boolean;
|
|
419
|
+
kind: "text";
|
|
420
|
+
maximumLength: number;
|
|
421
|
+
requireRationale?: boolean | undefined;
|
|
422
|
+
anchors?: {
|
|
423
|
+
value: string | number | boolean;
|
|
424
|
+
description: string;
|
|
425
|
+
}[] | undefined;
|
|
426
|
+
minimumLength?: number | undefined;
|
|
427
|
+
} | {
|
|
428
|
+
id: string;
|
|
429
|
+
label: string;
|
|
430
|
+
instructions: string;
|
|
431
|
+
required: boolean;
|
|
432
|
+
allowAbstain: boolean;
|
|
433
|
+
kind: "preference";
|
|
434
|
+
allowTie: boolean;
|
|
435
|
+
requireRationale?: boolean | undefined;
|
|
436
|
+
anchors?: {
|
|
437
|
+
value: string | number | boolean;
|
|
438
|
+
description: string;
|
|
439
|
+
}[] | undefined;
|
|
440
|
+
requireStrength?: boolean | undefined;
|
|
441
|
+
})[];
|
|
442
|
+
} | undefined;
|
|
295
443
|
})[];
|
|
296
444
|
isolation: {
|
|
297
445
|
processBoundary: "same_process" | "isolated_process" | "container";
|
|
@@ -347,7 +495,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
347
495
|
};
|
|
348
496
|
environment: {
|
|
349
497
|
protocolVersion: "openpond.environment.v1";
|
|
350
|
-
kind: "
|
|
498
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
351
499
|
entrypoint: string;
|
|
352
500
|
stateful: boolean;
|
|
353
501
|
deterministicSeeds: boolean;
|
|
@@ -417,7 +565,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
417
565
|
name: string;
|
|
418
566
|
description: string;
|
|
419
567
|
implementation: {
|
|
420
|
-
kind: "content" | "state" | "
|
|
568
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
421
569
|
config: Record<string, unknown>;
|
|
422
570
|
} | {
|
|
423
571
|
kind: "model_judge";
|
|
@@ -429,7 +577,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
429
577
|
mediaType: string;
|
|
430
578
|
visibility: "policy" | "verifier" | "host_private";
|
|
431
579
|
};
|
|
432
|
-
calibrationStatus: "passed" | "
|
|
580
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
433
581
|
model?: {
|
|
434
582
|
providerId: string;
|
|
435
583
|
modelId: string;
|
|
@@ -461,6 +609,80 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
461
609
|
visibility: "policy" | "verifier" | "host_private";
|
|
462
610
|
};
|
|
463
611
|
reviewerRole: string;
|
|
612
|
+
form?: {
|
|
613
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
614
|
+
mode: "individual" | "pairwise";
|
|
615
|
+
instructions: string;
|
|
616
|
+
criteria: ({
|
|
617
|
+
id: string;
|
|
618
|
+
label: string;
|
|
619
|
+
instructions: string;
|
|
620
|
+
required: boolean;
|
|
621
|
+
allowAbstain: boolean;
|
|
622
|
+
kind: "score";
|
|
623
|
+
minimum: number;
|
|
624
|
+
maximum: number;
|
|
625
|
+
step: number;
|
|
626
|
+
requireRationale?: boolean | undefined;
|
|
627
|
+
anchors?: {
|
|
628
|
+
value: string | number | boolean;
|
|
629
|
+
description: string;
|
|
630
|
+
}[] | undefined;
|
|
631
|
+
} | {
|
|
632
|
+
id: string;
|
|
633
|
+
label: string;
|
|
634
|
+
instructions: string;
|
|
635
|
+
required: boolean;
|
|
636
|
+
allowAbstain: boolean;
|
|
637
|
+
kind: "category";
|
|
638
|
+
options: string[];
|
|
639
|
+
requireRationale?: boolean | undefined;
|
|
640
|
+
anchors?: {
|
|
641
|
+
value: string | number | boolean;
|
|
642
|
+
description: string;
|
|
643
|
+
}[] | undefined;
|
|
644
|
+
} | {
|
|
645
|
+
id: string;
|
|
646
|
+
label: string;
|
|
647
|
+
instructions: string;
|
|
648
|
+
required: boolean;
|
|
649
|
+
allowAbstain: boolean;
|
|
650
|
+
kind: "boolean";
|
|
651
|
+
requireRationale?: boolean | undefined;
|
|
652
|
+
anchors?: {
|
|
653
|
+
value: string | number | boolean;
|
|
654
|
+
description: string;
|
|
655
|
+
}[] | undefined;
|
|
656
|
+
} | {
|
|
657
|
+
id: string;
|
|
658
|
+
label: string;
|
|
659
|
+
instructions: string;
|
|
660
|
+
required: boolean;
|
|
661
|
+
allowAbstain: boolean;
|
|
662
|
+
kind: "text";
|
|
663
|
+
maximumLength: number;
|
|
664
|
+
requireRationale?: boolean | undefined;
|
|
665
|
+
anchors?: {
|
|
666
|
+
value: string | number | boolean;
|
|
667
|
+
description: string;
|
|
668
|
+
}[] | undefined;
|
|
669
|
+
minimumLength?: number | undefined;
|
|
670
|
+
} | {
|
|
671
|
+
id: string;
|
|
672
|
+
label: string;
|
|
673
|
+
instructions: string;
|
|
674
|
+
required: boolean;
|
|
675
|
+
allowAbstain: boolean;
|
|
676
|
+
kind: "preference";
|
|
677
|
+
allowTie: boolean;
|
|
678
|
+
requireRationale?: boolean | undefined;
|
|
679
|
+
anchors?: {
|
|
680
|
+
value: string | number | boolean;
|
|
681
|
+
description: string;
|
|
682
|
+
}[] | undefined;
|
|
683
|
+
requireStrength?: boolean | undefined;
|
|
684
|
+
})[];
|
|
685
|
+
} | undefined;
|
|
464
686
|
} | {
|
|
465
687
|
kind: "learned_model";
|
|
466
688
|
modelVersion: {
|
|
@@ -551,7 +773,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
551
773
|
revision: number;
|
|
552
774
|
contract: {
|
|
553
775
|
protocolVersion: "openpond.environment.v1";
|
|
554
|
-
kind: "
|
|
776
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
555
777
|
entrypoint: string;
|
|
556
778
|
stateful: boolean;
|
|
557
779
|
deterministicSeeds: boolean;
|
|
@@ -602,7 +824,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
602
824
|
hardGate: boolean;
|
|
603
825
|
rewardEligible: boolean;
|
|
604
826
|
privileged: boolean;
|
|
605
|
-
kind: "content" | "state" | "
|
|
827
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
606
828
|
config: Record<string, unknown>;
|
|
607
829
|
} | {
|
|
608
830
|
id: string;
|
|
@@ -620,7 +842,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
620
842
|
mediaType: string;
|
|
621
843
|
visibility: "policy" | "verifier" | "host_private";
|
|
622
844
|
};
|
|
623
|
-
calibrationStatus: "passed" | "
|
|
845
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
624
846
|
model?: {
|
|
625
847
|
providerId: string;
|
|
626
848
|
modelId: string;
|
|
@@ -664,6 +886,80 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
664
886
|
visibility: "policy" | "verifier" | "host_private";
|
|
665
887
|
};
|
|
666
888
|
reviewerRole: string;
|
|
889
|
+
form?: {
|
|
890
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
891
|
+
mode: "individual" | "pairwise";
|
|
892
|
+
instructions: string;
|
|
893
|
+
criteria: ({
|
|
894
|
+
id: string;
|
|
895
|
+
label: string;
|
|
896
|
+
instructions: string;
|
|
897
|
+
required: boolean;
|
|
898
|
+
allowAbstain: boolean;
|
|
899
|
+
kind: "score";
|
|
900
|
+
minimum: number;
|
|
901
|
+
maximum: number;
|
|
902
|
+
step: number;
|
|
903
|
+
requireRationale?: boolean | undefined;
|
|
904
|
+
anchors?: {
|
|
905
|
+
value: string | number | boolean;
|
|
906
|
+
description: string;
|
|
907
|
+
}[] | undefined;
|
|
908
|
+
} | {
|
|
909
|
+
id: string;
|
|
910
|
+
label: string;
|
|
911
|
+
instructions: string;
|
|
912
|
+
required: boolean;
|
|
913
|
+
allowAbstain: boolean;
|
|
914
|
+
kind: "category";
|
|
915
|
+
options: string[];
|
|
916
|
+
requireRationale?: boolean | undefined;
|
|
917
|
+
anchors?: {
|
|
918
|
+
value: string | number | boolean;
|
|
919
|
+
description: string;
|
|
920
|
+
}[] | undefined;
|
|
921
|
+
} | {
|
|
922
|
+
id: string;
|
|
923
|
+
label: string;
|
|
924
|
+
instructions: string;
|
|
925
|
+
required: boolean;
|
|
926
|
+
allowAbstain: boolean;
|
|
927
|
+
kind: "boolean";
|
|
928
|
+
requireRationale?: boolean | undefined;
|
|
929
|
+
anchors?: {
|
|
930
|
+
value: string | number | boolean;
|
|
931
|
+
description: string;
|
|
932
|
+
}[] | undefined;
|
|
933
|
+
} | {
|
|
934
|
+
id: string;
|
|
935
|
+
label: string;
|
|
936
|
+
instructions: string;
|
|
937
|
+
required: boolean;
|
|
938
|
+
allowAbstain: boolean;
|
|
939
|
+
kind: "text";
|
|
940
|
+
maximumLength: number;
|
|
941
|
+
requireRationale?: boolean | undefined;
|
|
942
|
+
anchors?: {
|
|
943
|
+
value: string | number | boolean;
|
|
944
|
+
description: string;
|
|
945
|
+
}[] | undefined;
|
|
946
|
+
minimumLength?: number | undefined;
|
|
947
|
+
} | {
|
|
948
|
+
id: string;
|
|
949
|
+
label: string;
|
|
950
|
+
instructions: string;
|
|
951
|
+
required: boolean;
|
|
952
|
+
allowAbstain: boolean;
|
|
953
|
+
kind: "preference";
|
|
954
|
+
allowTie: boolean;
|
|
955
|
+
requireRationale?: boolean | undefined;
|
|
956
|
+
anchors?: {
|
|
957
|
+
value: string | number | boolean;
|
|
958
|
+
description: string;
|
|
959
|
+
}[] | undefined;
|
|
960
|
+
requireStrength?: boolean | undefined;
|
|
961
|
+
})[];
|
|
962
|
+
} | undefined;
|
|
667
963
|
})[];
|
|
668
964
|
isolation: {
|
|
669
965
|
processBoundary: "same_process" | "isolated_process" | "container";
|
|
@@ -862,9 +1158,9 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
862
1158
|
} | null;
|
|
863
1159
|
};
|
|
864
1160
|
evidenceValidity: "pending" | "valid" | "invalid";
|
|
865
|
-
taskAdmissibility: "pending" | "
|
|
1161
|
+
taskAdmissibility: "pending" | "approved" | "rejected";
|
|
866
1162
|
observedQuality: "passed" | "failed" | "unavailable" | "unscored";
|
|
867
|
-
targetApproval: "pending" | "
|
|
1163
|
+
targetApproval: "pending" | "approved" | "rejected" | "not_required";
|
|
868
1164
|
approvedTarget: Record<string, import("zod").JSONType> | null;
|
|
869
1165
|
grade: {
|
|
870
1166
|
schemaVersion: "openpond.rewardComposition.v1";
|
|
@@ -883,7 +1179,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
883
1179
|
revision: number;
|
|
884
1180
|
};
|
|
885
1181
|
role: "training" | "evaluation";
|
|
886
|
-
status: "
|
|
1182
|
+
status: "scored" | "pending" | "failed" | "unavailable";
|
|
887
1183
|
rawScore: number | null;
|
|
888
1184
|
normalizedScore: number | null;
|
|
889
1185
|
passed: boolean | null;
|
|
@@ -901,6 +1197,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
901
1197
|
visibleEvidenceRefs: string[];
|
|
902
1198
|
privilegedEvidenceRefs: string[];
|
|
903
1199
|
contentHash: string;
|
|
1200
|
+
status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
|
|
904
1201
|
modelJudgeReceipt?: {
|
|
905
1202
|
schemaVersion: "openpond.modelJudgeReceipt.v1";
|
|
906
1203
|
providerId: string;
|
|
@@ -916,12 +1213,12 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
916
1213
|
} | undefined;
|
|
917
1214
|
}[];
|
|
918
1215
|
training: {
|
|
919
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
1216
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
920
1217
|
score: number | null;
|
|
921
1218
|
passed: boolean | null;
|
|
922
1219
|
};
|
|
923
1220
|
evaluation: {
|
|
924
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
1221
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
925
1222
|
score: number | null;
|
|
926
1223
|
passed: boolean | null;
|
|
927
1224
|
};
|
|
@@ -944,7 +1241,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
944
1241
|
revision: number;
|
|
945
1242
|
};
|
|
946
1243
|
role: "training" | "evaluation";
|
|
947
|
-
status: "
|
|
1244
|
+
status: "scored" | "pending" | "failed" | "unavailable";
|
|
948
1245
|
rawScore: number | null;
|
|
949
1246
|
normalizedScore: number | null;
|
|
950
1247
|
passed: boolean | null;
|
|
@@ -962,6 +1259,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
962
1259
|
visibleEvidenceRefs: string[];
|
|
963
1260
|
privilegedEvidenceRefs: string[];
|
|
964
1261
|
contentHash: string;
|
|
1262
|
+
status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
|
|
965
1263
|
modelJudgeReceipt?: {
|
|
966
1264
|
schemaVersion: "openpond.modelJudgeReceipt.v1";
|
|
967
1265
|
providerId: string;
|
|
@@ -977,12 +1275,12 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
977
1275
|
} | undefined;
|
|
978
1276
|
}[];
|
|
979
1277
|
training: {
|
|
980
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
1278
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
981
1279
|
score: number | null;
|
|
982
1280
|
passed: boolean | null;
|
|
983
1281
|
};
|
|
984
1282
|
evaluation: {
|
|
985
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
1283
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
986
1284
|
score: number | null;
|
|
987
1285
|
passed: boolean | null;
|
|
988
1286
|
};
|
|
@@ -997,7 +1295,7 @@ export declare function compileTasksetDraftWorkspace(input: {
|
|
|
997
1295
|
id: string;
|
|
998
1296
|
revision: number;
|
|
999
1297
|
name: string;
|
|
1000
|
-
kind: "provider" | "
|
|
1298
|
+
kind: "provider" | "work" | "benchmark" | "direct" | "opentelemetry";
|
|
1001
1299
|
taskDefinition: {
|
|
1002
1300
|
id: string;
|
|
1003
1301
|
contentHash: string;
|
|
@@ -1071,7 +1369,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1071
1369
|
};
|
|
1072
1370
|
environment: {
|
|
1073
1371
|
protocolVersion: "openpond.environment.v1";
|
|
1074
|
-
kind: "
|
|
1372
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
1075
1373
|
entrypoint: string;
|
|
1076
1374
|
stateful: boolean;
|
|
1077
1375
|
deterministicSeeds: boolean;
|
|
@@ -1132,7 +1430,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1132
1430
|
hardGate: boolean;
|
|
1133
1431
|
rewardEligible: boolean;
|
|
1134
1432
|
privileged: boolean;
|
|
1135
|
-
kind: "content" | "state" | "
|
|
1433
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
1136
1434
|
config: Record<string, unknown>;
|
|
1137
1435
|
} | {
|
|
1138
1436
|
id: string;
|
|
@@ -1150,7 +1448,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1150
1448
|
mediaType: string;
|
|
1151
1449
|
visibility: "policy" | "verifier" | "host_private";
|
|
1152
1450
|
};
|
|
1153
|
-
calibrationStatus: "passed" | "
|
|
1451
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
1154
1452
|
model?: {
|
|
1155
1453
|
providerId: string;
|
|
1156
1454
|
modelId: string;
|
|
@@ -1194,6 +1492,80 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1194
1492
|
visibility: "policy" | "verifier" | "host_private";
|
|
1195
1493
|
};
|
|
1196
1494
|
reviewerRole: string;
|
|
1495
|
+
form?: {
|
|
1496
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
1497
|
+
mode: "individual" | "pairwise";
|
|
1498
|
+
instructions: string;
|
|
1499
|
+
criteria: ({
|
|
1500
|
+
id: string;
|
|
1501
|
+
label: string;
|
|
1502
|
+
instructions: string;
|
|
1503
|
+
required: boolean;
|
|
1504
|
+
allowAbstain: boolean;
|
|
1505
|
+
kind: "score";
|
|
1506
|
+
minimum: number;
|
|
1507
|
+
maximum: number;
|
|
1508
|
+
step: number;
|
|
1509
|
+
requireRationale?: boolean | undefined;
|
|
1510
|
+
anchors?: {
|
|
1511
|
+
value: string | number | boolean;
|
|
1512
|
+
description: string;
|
|
1513
|
+
}[] | undefined;
|
|
1514
|
+
} | {
|
|
1515
|
+
id: string;
|
|
1516
|
+
label: string;
|
|
1517
|
+
instructions: string;
|
|
1518
|
+
required: boolean;
|
|
1519
|
+
allowAbstain: boolean;
|
|
1520
|
+
kind: "category";
|
|
1521
|
+
options: string[];
|
|
1522
|
+
requireRationale?: boolean | undefined;
|
|
1523
|
+
anchors?: {
|
|
1524
|
+
value: string | number | boolean;
|
|
1525
|
+
description: string;
|
|
1526
|
+
}[] | undefined;
|
|
1527
|
+
} | {
|
|
1528
|
+
id: string;
|
|
1529
|
+
label: string;
|
|
1530
|
+
instructions: string;
|
|
1531
|
+
required: boolean;
|
|
1532
|
+
allowAbstain: boolean;
|
|
1533
|
+
kind: "boolean";
|
|
1534
|
+
requireRationale?: boolean | undefined;
|
|
1535
|
+
anchors?: {
|
|
1536
|
+
value: string | number | boolean;
|
|
1537
|
+
description: string;
|
|
1538
|
+
}[] | undefined;
|
|
1539
|
+
} | {
|
|
1540
|
+
id: string;
|
|
1541
|
+
label: string;
|
|
1542
|
+
instructions: string;
|
|
1543
|
+
required: boolean;
|
|
1544
|
+
allowAbstain: boolean;
|
|
1545
|
+
kind: "text";
|
|
1546
|
+
maximumLength: number;
|
|
1547
|
+
requireRationale?: boolean | undefined;
|
|
1548
|
+
anchors?: {
|
|
1549
|
+
value: string | number | boolean;
|
|
1550
|
+
description: string;
|
|
1551
|
+
}[] | undefined;
|
|
1552
|
+
minimumLength?: number | undefined;
|
|
1553
|
+
} | {
|
|
1554
|
+
id: string;
|
|
1555
|
+
label: string;
|
|
1556
|
+
instructions: string;
|
|
1557
|
+
required: boolean;
|
|
1558
|
+
allowAbstain: boolean;
|
|
1559
|
+
kind: "preference";
|
|
1560
|
+
allowTie: boolean;
|
|
1561
|
+
requireRationale?: boolean | undefined;
|
|
1562
|
+
anchors?: {
|
|
1563
|
+
value: string | number | boolean;
|
|
1564
|
+
description: string;
|
|
1565
|
+
}[] | undefined;
|
|
1566
|
+
requireStrength?: boolean | undefined;
|
|
1567
|
+
})[];
|
|
1568
|
+
} | undefined;
|
|
1197
1569
|
})[];
|
|
1198
1570
|
metadata: Record<string, unknown>;
|
|
1199
1571
|
contentHash: string;
|
|
@@ -1226,7 +1598,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1226
1598
|
revision: number;
|
|
1227
1599
|
contract: {
|
|
1228
1600
|
protocolVersion: "openpond.environment.v1";
|
|
1229
|
-
kind: "
|
|
1601
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
1230
1602
|
entrypoint: string;
|
|
1231
1603
|
stateful: boolean;
|
|
1232
1604
|
deterministicSeeds: boolean;
|
|
@@ -1277,7 +1649,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1277
1649
|
hardGate: boolean;
|
|
1278
1650
|
rewardEligible: boolean;
|
|
1279
1651
|
privileged: boolean;
|
|
1280
|
-
kind: "content" | "state" | "
|
|
1652
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
1281
1653
|
config: Record<string, unknown>;
|
|
1282
1654
|
} | {
|
|
1283
1655
|
id: string;
|
|
@@ -1295,7 +1667,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1295
1667
|
mediaType: string;
|
|
1296
1668
|
visibility: "policy" | "verifier" | "host_private";
|
|
1297
1669
|
};
|
|
1298
|
-
calibrationStatus: "passed" | "
|
|
1670
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
1299
1671
|
model?: {
|
|
1300
1672
|
providerId: string;
|
|
1301
1673
|
modelId: string;
|
|
@@ -1339,6 +1711,80 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1339
1711
|
visibility: "policy" | "verifier" | "host_private";
|
|
1340
1712
|
};
|
|
1341
1713
|
reviewerRole: string;
|
|
1714
|
+
form?: {
|
|
1715
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
1716
|
+
mode: "individual" | "pairwise";
|
|
1717
|
+
instructions: string;
|
|
1718
|
+
criteria: ({
|
|
1719
|
+
id: string;
|
|
1720
|
+
label: string;
|
|
1721
|
+
instructions: string;
|
|
1722
|
+
required: boolean;
|
|
1723
|
+
allowAbstain: boolean;
|
|
1724
|
+
kind: "score";
|
|
1725
|
+
minimum: number;
|
|
1726
|
+
maximum: number;
|
|
1727
|
+
step: number;
|
|
1728
|
+
requireRationale?: boolean | undefined;
|
|
1729
|
+
anchors?: {
|
|
1730
|
+
value: string | number | boolean;
|
|
1731
|
+
description: string;
|
|
1732
|
+
}[] | undefined;
|
|
1733
|
+
} | {
|
|
1734
|
+
id: string;
|
|
1735
|
+
label: string;
|
|
1736
|
+
instructions: string;
|
|
1737
|
+
required: boolean;
|
|
1738
|
+
allowAbstain: boolean;
|
|
1739
|
+
kind: "category";
|
|
1740
|
+
options: string[];
|
|
1741
|
+
requireRationale?: boolean | undefined;
|
|
1742
|
+
anchors?: {
|
|
1743
|
+
value: string | number | boolean;
|
|
1744
|
+
description: string;
|
|
1745
|
+
}[] | undefined;
|
|
1746
|
+
} | {
|
|
1747
|
+
id: string;
|
|
1748
|
+
label: string;
|
|
1749
|
+
instructions: string;
|
|
1750
|
+
required: boolean;
|
|
1751
|
+
allowAbstain: boolean;
|
|
1752
|
+
kind: "boolean";
|
|
1753
|
+
requireRationale?: boolean | undefined;
|
|
1754
|
+
anchors?: {
|
|
1755
|
+
value: string | number | boolean;
|
|
1756
|
+
description: string;
|
|
1757
|
+
}[] | undefined;
|
|
1758
|
+
} | {
|
|
1759
|
+
id: string;
|
|
1760
|
+
label: string;
|
|
1761
|
+
instructions: string;
|
|
1762
|
+
required: boolean;
|
|
1763
|
+
allowAbstain: boolean;
|
|
1764
|
+
kind: "text";
|
|
1765
|
+
maximumLength: number;
|
|
1766
|
+
requireRationale?: boolean | undefined;
|
|
1767
|
+
anchors?: {
|
|
1768
|
+
value: string | number | boolean;
|
|
1769
|
+
description: string;
|
|
1770
|
+
}[] | undefined;
|
|
1771
|
+
minimumLength?: number | undefined;
|
|
1772
|
+
} | {
|
|
1773
|
+
id: string;
|
|
1774
|
+
label: string;
|
|
1775
|
+
instructions: string;
|
|
1776
|
+
required: boolean;
|
|
1777
|
+
allowAbstain: boolean;
|
|
1778
|
+
kind: "preference";
|
|
1779
|
+
allowTie: boolean;
|
|
1780
|
+
requireRationale?: boolean | undefined;
|
|
1781
|
+
anchors?: {
|
|
1782
|
+
value: string | number | boolean;
|
|
1783
|
+
description: string;
|
|
1784
|
+
}[] | undefined;
|
|
1785
|
+
requireStrength?: boolean | undefined;
|
|
1786
|
+
})[];
|
|
1787
|
+
} | undefined;
|
|
1342
1788
|
})[];
|
|
1343
1789
|
isolation: {
|
|
1344
1790
|
processBoundary: "same_process" | "isolated_process" | "container";
|
|
@@ -1394,7 +1840,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1394
1840
|
};
|
|
1395
1841
|
environment: {
|
|
1396
1842
|
protocolVersion: "openpond.environment.v1";
|
|
1397
|
-
kind: "
|
|
1843
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
1398
1844
|
entrypoint: string;
|
|
1399
1845
|
stateful: boolean;
|
|
1400
1846
|
deterministicSeeds: boolean;
|
|
@@ -1464,7 +1910,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1464
1910
|
name: string;
|
|
1465
1911
|
description: string;
|
|
1466
1912
|
implementation: {
|
|
1467
|
-
kind: "content" | "state" | "
|
|
1913
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
1468
1914
|
config: Record<string, unknown>;
|
|
1469
1915
|
} | {
|
|
1470
1916
|
kind: "model_judge";
|
|
@@ -1476,7 +1922,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1476
1922
|
mediaType: string;
|
|
1477
1923
|
visibility: "policy" | "verifier" | "host_private";
|
|
1478
1924
|
};
|
|
1479
|
-
calibrationStatus: "passed" | "
|
|
1925
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
1480
1926
|
model?: {
|
|
1481
1927
|
providerId: string;
|
|
1482
1928
|
modelId: string;
|
|
@@ -1508,6 +1954,80 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1508
1954
|
visibility: "policy" | "verifier" | "host_private";
|
|
1509
1955
|
};
|
|
1510
1956
|
reviewerRole: string;
|
|
1957
|
+
form?: {
|
|
1958
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
1959
|
+
mode: "individual" | "pairwise";
|
|
1960
|
+
instructions: string;
|
|
1961
|
+
criteria: ({
|
|
1962
|
+
id: string;
|
|
1963
|
+
label: string;
|
|
1964
|
+
instructions: string;
|
|
1965
|
+
required: boolean;
|
|
1966
|
+
allowAbstain: boolean;
|
|
1967
|
+
kind: "score";
|
|
1968
|
+
minimum: number;
|
|
1969
|
+
maximum: number;
|
|
1970
|
+
step: number;
|
|
1971
|
+
requireRationale?: boolean | undefined;
|
|
1972
|
+
anchors?: {
|
|
1973
|
+
value: string | number | boolean;
|
|
1974
|
+
description: string;
|
|
1975
|
+
}[] | undefined;
|
|
1976
|
+
} | {
|
|
1977
|
+
id: string;
|
|
1978
|
+
label: string;
|
|
1979
|
+
instructions: string;
|
|
1980
|
+
required: boolean;
|
|
1981
|
+
allowAbstain: boolean;
|
|
1982
|
+
kind: "category";
|
|
1983
|
+
options: string[];
|
|
1984
|
+
requireRationale?: boolean | undefined;
|
|
1985
|
+
anchors?: {
|
|
1986
|
+
value: string | number | boolean;
|
|
1987
|
+
description: string;
|
|
1988
|
+
}[] | undefined;
|
|
1989
|
+
} | {
|
|
1990
|
+
id: string;
|
|
1991
|
+
label: string;
|
|
1992
|
+
instructions: string;
|
|
1993
|
+
required: boolean;
|
|
1994
|
+
allowAbstain: boolean;
|
|
1995
|
+
kind: "boolean";
|
|
1996
|
+
requireRationale?: boolean | undefined;
|
|
1997
|
+
anchors?: {
|
|
1998
|
+
value: string | number | boolean;
|
|
1999
|
+
description: string;
|
|
2000
|
+
}[] | undefined;
|
|
2001
|
+
} | {
|
|
2002
|
+
id: string;
|
|
2003
|
+
label: string;
|
|
2004
|
+
instructions: string;
|
|
2005
|
+
required: boolean;
|
|
2006
|
+
allowAbstain: boolean;
|
|
2007
|
+
kind: "text";
|
|
2008
|
+
maximumLength: number;
|
|
2009
|
+
requireRationale?: boolean | undefined;
|
|
2010
|
+
anchors?: {
|
|
2011
|
+
value: string | number | boolean;
|
|
2012
|
+
description: string;
|
|
2013
|
+
}[] | undefined;
|
|
2014
|
+
minimumLength?: number | undefined;
|
|
2015
|
+
} | {
|
|
2016
|
+
id: string;
|
|
2017
|
+
label: string;
|
|
2018
|
+
instructions: string;
|
|
2019
|
+
required: boolean;
|
|
2020
|
+
allowAbstain: boolean;
|
|
2021
|
+
kind: "preference";
|
|
2022
|
+
allowTie: boolean;
|
|
2023
|
+
requireRationale?: boolean | undefined;
|
|
2024
|
+
anchors?: {
|
|
2025
|
+
value: string | number | boolean;
|
|
2026
|
+
description: string;
|
|
2027
|
+
}[] | undefined;
|
|
2028
|
+
requireStrength?: boolean | undefined;
|
|
2029
|
+
})[];
|
|
2030
|
+
} | undefined;
|
|
1511
2031
|
} | {
|
|
1512
2032
|
kind: "learned_model";
|
|
1513
2033
|
modelVersion: {
|
|
@@ -1598,7 +2118,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1598
2118
|
revision: number;
|
|
1599
2119
|
contract: {
|
|
1600
2120
|
protocolVersion: "openpond.environment.v1";
|
|
1601
|
-
kind: "
|
|
2121
|
+
kind: "text" | "agent" | "work" | "custom_program";
|
|
1602
2122
|
entrypoint: string;
|
|
1603
2123
|
stateful: boolean;
|
|
1604
2124
|
deterministicSeeds: boolean;
|
|
@@ -1649,7 +2169,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1649
2169
|
hardGate: boolean;
|
|
1650
2170
|
rewardEligible: boolean;
|
|
1651
2171
|
privileged: boolean;
|
|
1652
|
-
kind: "content" | "state" | "
|
|
2172
|
+
kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
|
|
1653
2173
|
config: Record<string, unknown>;
|
|
1654
2174
|
} | {
|
|
1655
2175
|
id: string;
|
|
@@ -1667,7 +2187,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1667
2187
|
mediaType: string;
|
|
1668
2188
|
visibility: "policy" | "verifier" | "host_private";
|
|
1669
2189
|
};
|
|
1670
|
-
calibrationStatus: "passed" | "
|
|
2190
|
+
calibrationStatus: "passed" | "pending" | "failed";
|
|
1671
2191
|
model?: {
|
|
1672
2192
|
providerId: string;
|
|
1673
2193
|
modelId: string;
|
|
@@ -1711,6 +2231,80 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1711
2231
|
visibility: "policy" | "verifier" | "host_private";
|
|
1712
2232
|
};
|
|
1713
2233
|
reviewerRole: string;
|
|
2234
|
+
form?: {
|
|
2235
|
+
schemaVersion: "openpond.humanForm.v1";
|
|
2236
|
+
mode: "individual" | "pairwise";
|
|
2237
|
+
instructions: string;
|
|
2238
|
+
criteria: ({
|
|
2239
|
+
id: string;
|
|
2240
|
+
label: string;
|
|
2241
|
+
instructions: string;
|
|
2242
|
+
required: boolean;
|
|
2243
|
+
allowAbstain: boolean;
|
|
2244
|
+
kind: "score";
|
|
2245
|
+
minimum: number;
|
|
2246
|
+
maximum: number;
|
|
2247
|
+
step: number;
|
|
2248
|
+
requireRationale?: boolean | undefined;
|
|
2249
|
+
anchors?: {
|
|
2250
|
+
value: string | number | boolean;
|
|
2251
|
+
description: string;
|
|
2252
|
+
}[] | undefined;
|
|
2253
|
+
} | {
|
|
2254
|
+
id: string;
|
|
2255
|
+
label: string;
|
|
2256
|
+
instructions: string;
|
|
2257
|
+
required: boolean;
|
|
2258
|
+
allowAbstain: boolean;
|
|
2259
|
+
kind: "category";
|
|
2260
|
+
options: string[];
|
|
2261
|
+
requireRationale?: boolean | undefined;
|
|
2262
|
+
anchors?: {
|
|
2263
|
+
value: string | number | boolean;
|
|
2264
|
+
description: string;
|
|
2265
|
+
}[] | undefined;
|
|
2266
|
+
} | {
|
|
2267
|
+
id: string;
|
|
2268
|
+
label: string;
|
|
2269
|
+
instructions: string;
|
|
2270
|
+
required: boolean;
|
|
2271
|
+
allowAbstain: boolean;
|
|
2272
|
+
kind: "boolean";
|
|
2273
|
+
requireRationale?: boolean | undefined;
|
|
2274
|
+
anchors?: {
|
|
2275
|
+
value: string | number | boolean;
|
|
2276
|
+
description: string;
|
|
2277
|
+
}[] | undefined;
|
|
2278
|
+
} | {
|
|
2279
|
+
id: string;
|
|
2280
|
+
label: string;
|
|
2281
|
+
instructions: string;
|
|
2282
|
+
required: boolean;
|
|
2283
|
+
allowAbstain: boolean;
|
|
2284
|
+
kind: "text";
|
|
2285
|
+
maximumLength: number;
|
|
2286
|
+
requireRationale?: boolean | undefined;
|
|
2287
|
+
anchors?: {
|
|
2288
|
+
value: string | number | boolean;
|
|
2289
|
+
description: string;
|
|
2290
|
+
}[] | undefined;
|
|
2291
|
+
minimumLength?: number | undefined;
|
|
2292
|
+
} | {
|
|
2293
|
+
id: string;
|
|
2294
|
+
label: string;
|
|
2295
|
+
instructions: string;
|
|
2296
|
+
required: boolean;
|
|
2297
|
+
allowAbstain: boolean;
|
|
2298
|
+
kind: "preference";
|
|
2299
|
+
allowTie: boolean;
|
|
2300
|
+
requireRationale?: boolean | undefined;
|
|
2301
|
+
anchors?: {
|
|
2302
|
+
value: string | number | boolean;
|
|
2303
|
+
description: string;
|
|
2304
|
+
}[] | undefined;
|
|
2305
|
+
requireStrength?: boolean | undefined;
|
|
2306
|
+
})[];
|
|
2307
|
+
} | undefined;
|
|
1714
2308
|
})[];
|
|
1715
2309
|
isolation: {
|
|
1716
2310
|
processBoundary: "same_process" | "isolated_process" | "container";
|
|
@@ -1909,9 +2503,9 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1909
2503
|
} | null;
|
|
1910
2504
|
};
|
|
1911
2505
|
evidenceValidity: "pending" | "valid" | "invalid";
|
|
1912
|
-
taskAdmissibility: "pending" | "
|
|
2506
|
+
taskAdmissibility: "pending" | "approved" | "rejected";
|
|
1913
2507
|
observedQuality: "passed" | "failed" | "unavailable" | "unscored";
|
|
1914
|
-
targetApproval: "pending" | "
|
|
2508
|
+
targetApproval: "pending" | "approved" | "rejected" | "not_required";
|
|
1915
2509
|
approvedTarget: Record<string, import("zod").JSONType> | null;
|
|
1916
2510
|
grade: {
|
|
1917
2511
|
schemaVersion: "openpond.rewardComposition.v1";
|
|
@@ -1930,7 +2524,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1930
2524
|
revision: number;
|
|
1931
2525
|
};
|
|
1932
2526
|
role: "training" | "evaluation";
|
|
1933
|
-
status: "
|
|
2527
|
+
status: "scored" | "pending" | "failed" | "unavailable";
|
|
1934
2528
|
rawScore: number | null;
|
|
1935
2529
|
normalizedScore: number | null;
|
|
1936
2530
|
passed: boolean | null;
|
|
@@ -1948,6 +2542,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1948
2542
|
visibleEvidenceRefs: string[];
|
|
1949
2543
|
privilegedEvidenceRefs: string[];
|
|
1950
2544
|
contentHash: string;
|
|
2545
|
+
status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
|
|
1951
2546
|
modelJudgeReceipt?: {
|
|
1952
2547
|
schemaVersion: "openpond.modelJudgeReceipt.v1";
|
|
1953
2548
|
providerId: string;
|
|
@@ -1963,12 +2558,12 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1963
2558
|
} | undefined;
|
|
1964
2559
|
}[];
|
|
1965
2560
|
training: {
|
|
1966
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
2561
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
1967
2562
|
score: number | null;
|
|
1968
2563
|
passed: boolean | null;
|
|
1969
2564
|
};
|
|
1970
2565
|
evaluation: {
|
|
1971
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
2566
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
1972
2567
|
score: number | null;
|
|
1973
2568
|
passed: boolean | null;
|
|
1974
2569
|
};
|
|
@@ -1991,7 +2586,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
1991
2586
|
revision: number;
|
|
1992
2587
|
};
|
|
1993
2588
|
role: "training" | "evaluation";
|
|
1994
|
-
status: "
|
|
2589
|
+
status: "scored" | "pending" | "failed" | "unavailable";
|
|
1995
2590
|
rawScore: number | null;
|
|
1996
2591
|
normalizedScore: number | null;
|
|
1997
2592
|
passed: boolean | null;
|
|
@@ -2009,6 +2604,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
2009
2604
|
visibleEvidenceRefs: string[];
|
|
2010
2605
|
privilegedEvidenceRefs: string[];
|
|
2011
2606
|
contentHash: string;
|
|
2607
|
+
status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
|
|
2012
2608
|
modelJudgeReceipt?: {
|
|
2013
2609
|
schemaVersion: "openpond.modelJudgeReceipt.v1";
|
|
2014
2610
|
providerId: string;
|
|
@@ -2024,12 +2620,12 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
2024
2620
|
} | undefined;
|
|
2025
2621
|
}[];
|
|
2026
2622
|
training: {
|
|
2027
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
2623
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
2028
2624
|
score: number | null;
|
|
2029
2625
|
passed: boolean | null;
|
|
2030
2626
|
};
|
|
2031
2627
|
evaluation: {
|
|
2032
|
-
status: "scored" | "unscorable" | "not_configured";
|
|
2628
|
+
status: "scored" | "pending" | "unscorable" | "not_configured";
|
|
2033
2629
|
score: number | null;
|
|
2034
2630
|
passed: boolean | null;
|
|
2035
2631
|
};
|
|
@@ -2044,7 +2640,7 @@ export declare function compileModelTasksetDraftWorkspace(input: Parameters<type
|
|
|
2044
2640
|
id: string;
|
|
2045
2641
|
revision: number;
|
|
2046
2642
|
name: string;
|
|
2047
|
-
kind: "provider" | "
|
|
2643
|
+
kind: "provider" | "work" | "benchmark" | "direct" | "opentelemetry";
|
|
2048
2644
|
taskDefinition: {
|
|
2049
2645
|
id: string;
|
|
2050
2646
|
contentHash: string;
|