niceeval 0.19.0 → 0.19.1-canary.102
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/INDEX.md +1 -1
- package/dist/assertions/record/codec.cjs +19 -13
- package/dist/assertions/record/codec.cjs.map +1 -1
- package/dist/assertions/record/codec.d.cts +2 -0
- package/dist/assertions/record/codec.d.mts +2 -0
- package/dist/assertions/record/codec.d.ts +2 -0
- package/dist/assertions/record/codec.js +16 -14
- package/dist/inspection/assertion-projection.cjs +4 -1
- package/dist/inspection/assertion-projection.cjs.map +1 -1
- package/dist/inspection/cli/contribution.cjs +18 -14
- package/dist/inspection/cli/contribution.cjs.map +1 -1
- package/dist/inspection/cli/contribution.d.cts +7 -0
- package/dist/inspection/cli/contribution.d.mts +7 -0
- package/dist/inspection/cli/contribution.d.ts +7 -0
- package/dist/inspection/overview.cjs +28 -14
- package/dist/inspection/overview.cjs.map +1 -1
- package/dist/inspection/overview.d.cts +151 -18
- package/dist/inspection/overview.d.mts +151 -18
- package/dist/inspection/overview.d.ts +151 -18
- package/dist/inspection/overview.js +4 -2
- package/dist/inspection/project-result.d.cts +148 -148
- package/dist/inspection/project-result.d.mts +148 -148
- package/dist/inspection/project-result.d.ts +148 -148
- package/dist/inspection/protocol-values.cjs +1 -1
- package/dist/inspection/protocol-values.cjs.map +1 -1
- package/dist/inspection/protocol-values.d.cts +1 -1
- package/dist/inspection/protocol-values.d.mts +1 -1
- package/dist/inspection/protocol-values.d.ts +1 -1
- package/dist/inspection/protocol.d.cts +611 -479
- package/dist/inspection/protocol.d.mts +611 -479
- package/dist/inspection/protocol.d.ts +611 -479
- package/dist/inspection/public.cjs +3 -1
- package/dist/inspection/public.cjs.map +1 -1
- package/dist/inspection/public.d.cts +1 -0
- package/dist/inspection/public.d.mts +1 -0
- package/dist/inspection/public.d.ts +1 -0
- package/dist/inspection/public.js +18 -16
- package/dist/inspection/public.mjs +54 -52
- package/dist/inspection/results.cjs +11 -15
- package/dist/inspection/results.cjs.map +1 -1
- package/dist/inspection/results.d.cts +769 -324
- package/dist/inspection/results.d.mts +769 -324
- package/dist/inspection/results.d.ts +769 -324
- package/dist/inspection/results.js +4 -0
- package/dist/inspection/select.cjs +12 -72
- package/dist/inspection/select.cjs.map +1 -1
- package/dist/inspection/verdict.cjs +61 -0
- package/dist/inspection/verdict.cjs.map +1 -0
- package/dist/inspection/verdict.d.cts +3 -0
- package/dist/inspection/verdict.d.mts +3 -0
- package/dist/inspection/verdict.d.ts +3 -0
- package/dist/inspection/verdict.js +6 -0
- package/dist/inspection/verdict.js.map +1 -0
- package/dist/runner/discover.cjs +5 -4
- package/dist/runner/discover.cjs.map +1 -1
- package/dist/sandbox/index.cjs +4 -2
- package/dist/sandbox/index.cjs.map +1 -1
- package/dist/sandbox/index.d.cts +2 -0
- package/dist/sandbox/index.d.mts +2 -0
- package/dist/sandbox/index.d.ts +2 -0
- package/dist/sandbox/index.js +34 -32
- package/dist/sandbox/index.mjs +34 -32
- package/dist/sandbox/opensandbox/managed-process.cjs +224 -0
- package/dist/sandbox/opensandbox/managed-process.cjs.map +1 -0
- package/dist/sandbox/opensandbox/managed-process.d.cts +9 -0
- package/dist/sandbox/opensandbox/managed-process.d.mts +9 -0
- package/dist/sandbox/opensandbox/managed-process.d.ts +9 -0
- package/dist/sandbox/opensandbox/managed-process.js +6 -0
- package/dist/sandbox/opensandbox/managed-process.js.map +1 -0
- package/dist/sandbox/opensandbox/materialize.cjs +25 -0
- package/dist/sandbox/opensandbox/materialize.cjs.map +1 -0
- package/dist/sandbox/opensandbox/materialize.d.cts +4 -0
- package/dist/sandbox/opensandbox/materialize.d.mts +4 -0
- package/dist/sandbox/opensandbox/materialize.d.ts +4 -0
- package/dist/sandbox/opensandbox/materialize.js +6 -0
- package/dist/sandbox/opensandbox/materialize.js.map +1 -0
- package/dist/sandbox/opensandbox/process-bridge.cjs +89 -0
- package/dist/sandbox/opensandbox/process-bridge.cjs.map +1 -0
- package/dist/sandbox/opensandbox/process-bridge.d.cts +2 -0
- package/dist/sandbox/opensandbox/process-bridge.d.mts +2 -0
- package/dist/sandbox/opensandbox/process-bridge.d.ts +2 -0
- package/dist/sandbox/opensandbox/process-bridge.js +6 -0
- package/dist/sandbox/opensandbox/process-bridge.js.map +1 -0
- package/dist/sandbox/opensandbox/sandbox.cjs +253 -0
- package/dist/sandbox/opensandbox/sandbox.cjs.map +1 -0
- package/dist/sandbox/opensandbox/sandbox.d.cts +51 -0
- package/dist/sandbox/opensandbox/sandbox.d.mts +51 -0
- package/dist/sandbox/opensandbox/sandbox.d.ts +51 -0
- package/dist/sandbox/opensandbox/sandbox.js +6 -0
- package/dist/sandbox/opensandbox/sandbox.js.map +1 -0
- package/dist/sandbox/opensandbox.cjs +88 -0
- package/dist/sandbox/opensandbox.cjs.map +1 -0
- package/dist/sandbox/opensandbox.d.cts +17 -0
- package/dist/sandbox/opensandbox.d.mts +17 -0
- package/dist/sandbox/opensandbox.d.ts +17 -0
- package/dist/sandbox/opensandbox.js +6 -0
- package/dist/sandbox/opensandbox.js.map +1 -0
- package/dist/sandbox/runtime.cjs +1 -0
- package/dist/sandbox/runtime.cjs.map +1 -1
- package/dist/sandbox/runtime.d.cts +2 -0
- package/dist/sandbox/runtime.d.mts +2 -0
- package/dist/sandbox/runtime.d.ts +2 -0
- package/dist/sandbox/runtime.js +2 -0
- package/dist/sandbox/types.cjs.map +1 -1
- package/dist/sandbox/types.d.cts +1 -1
- package/dist/sandbox/types.d.mts +1 -1
- package/dist/sandbox/types.d.ts +1 -1
- package/dist/show/model.cjs +6 -5
- package/dist/show/model.cjs.map +1 -1
- package/dist/show/model.d.cts +1 -1
- package/dist/show/model.d.mts +1 -1
- package/dist/show/model.d.ts +1 -1
- package/dist/show/render.cjs +2 -2
- package/dist/show/render.cjs.map +1 -1
- package/dist/view/app-dist/.vite/manifest.json +2 -2
- package/dist/view/app-dist/assets/{index-CxrOUvHC.js → index-f9ZKc4km.js} +16 -16
- package/dist/view/app-dist/assets/{quality-cost-scatter-B6Aiss7x.js → quality-cost-scatter-C6lsrigr.js} +1 -1
- package/dist/view/app-dist/index.html +1 -1
- package/dist/view/function-runtime.mjs +126 -118
- package/dist/view/http-protocol.d.cts +1 -1
- package/dist/view/http-protocol.d.mts +1 -1
- package/dist/view/http-protocol.d.ts +1 -1
- package/dist/view/inspection-host.d.cts +625 -493
- package/dist/view/inspection-host.d.mts +625 -493
- package/dist/view/inspection-host.d.ts +625 -493
- package/docs-site/zh/explanation/evals.mdx +4 -0
- package/docs-site/zh/reference/cli.mdx +45 -4
- package/docs-site/zh/tutorials/compare-runs.mdx +1 -1
- package/docs-site/zh/tutorials/niceeval-taste.mdx +34 -1
- package/docs-site/zh/tutorials/sandbox-providers.mdx +35 -4
- package/docs-site/zh/tutorials/upgrade-0-19.mdx +1 -1
- package/docs-site/zh/tutorials/viewing-results.mdx +36 -3
- package/package.json +6 -1
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { Effect } from "effect";
|
|
2
2
|
import type { ViewGeneration } from "./revision.cjs";
|
|
3
3
|
export declare function inspectViewGeneration(generation: ViewGeneration, input: unknown): Effect.Effect<({
|
|
4
|
-
readonly protocol: "niceeval.query/
|
|
4
|
+
readonly protocol: "niceeval.query/v3";
|
|
5
5
|
readonly outcome: "success";
|
|
6
6
|
readonly behaviorVersion: "2.0.0";
|
|
7
7
|
readonly source: {
|
|
@@ -1056,9 +1056,6 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
1056
1056
|
readonly groupPath: readonly string[];
|
|
1057
1057
|
readonly members: readonly {
|
|
1058
1058
|
readonly runId: string;
|
|
1059
|
-
readonly plannedAttemptId: string;
|
|
1060
|
-
readonly evalId: string;
|
|
1061
|
-
readonly attemptOrdinal: number;
|
|
1062
1059
|
readonly labels: {
|
|
1063
1060
|
readonly [x: string]: {
|
|
1064
1061
|
readonly state: "available";
|
|
@@ -1069,158 +1066,161 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
1069
1066
|
readonly state: "unavailable";
|
|
1070
1067
|
};
|
|
1071
1068
|
};
|
|
1069
|
+
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
1070
|
+
readonly score: {
|
|
1071
|
+
readonly value: number | null;
|
|
1072
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1073
|
+
readonly issues: readonly ({
|
|
1074
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1075
|
+
readonly locator: string;
|
|
1076
|
+
} | {
|
|
1077
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1078
|
+
readonly runId: string;
|
|
1079
|
+
readonly plannedAttemptId: string;
|
|
1080
|
+
readonly evalId: string;
|
|
1081
|
+
readonly attemptOrdinal: number;
|
|
1082
|
+
} | {
|
|
1083
|
+
readonly code: "score-contribution-unavailable";
|
|
1084
|
+
readonly locator: string;
|
|
1085
|
+
readonly entryId: string;
|
|
1086
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1087
|
+
} | {
|
|
1088
|
+
readonly code: "score-partial";
|
|
1089
|
+
readonly locator: string | null;
|
|
1090
|
+
})[];
|
|
1091
|
+
readonly basis: "eval" | "planned-attempt";
|
|
1092
|
+
readonly total: number;
|
|
1093
|
+
readonly refs: readonly {
|
|
1094
|
+
readonly identity: {
|
|
1095
|
+
readonly kind: "attempt";
|
|
1096
|
+
readonly locator: string;
|
|
1097
|
+
};
|
|
1098
|
+
}[];
|
|
1099
|
+
readonly samples: number;
|
|
1100
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1101
|
+
readonly bounds?: {
|
|
1102
|
+
readonly min: number;
|
|
1103
|
+
readonly max: number;
|
|
1104
|
+
} | undefined;
|
|
1105
|
+
};
|
|
1106
|
+
readonly costUSD: {
|
|
1107
|
+
readonly value: number | null;
|
|
1108
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1109
|
+
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
1110
|
+
readonly issues: readonly ({
|
|
1111
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1112
|
+
readonly locator: string;
|
|
1113
|
+
} | {
|
|
1114
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1115
|
+
readonly runId: string;
|
|
1116
|
+
readonly plannedAttemptId: string;
|
|
1117
|
+
readonly evalId: string;
|
|
1118
|
+
readonly attemptOrdinal: number;
|
|
1119
|
+
} | {
|
|
1120
|
+
readonly code: "score-contribution-unavailable";
|
|
1121
|
+
readonly locator: string;
|
|
1122
|
+
readonly entryId: string;
|
|
1123
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1124
|
+
} | {
|
|
1125
|
+
readonly code: "score-partial";
|
|
1126
|
+
readonly locator: string | null;
|
|
1127
|
+
})[];
|
|
1128
|
+
readonly basis: "eval" | "planned-attempt";
|
|
1129
|
+
readonly total: number;
|
|
1130
|
+
readonly refs: readonly {
|
|
1131
|
+
readonly identity: {
|
|
1132
|
+
readonly kind: "attempt";
|
|
1133
|
+
readonly locator: string;
|
|
1134
|
+
};
|
|
1135
|
+
}[];
|
|
1136
|
+
readonly samples: number;
|
|
1137
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1138
|
+
readonly bounds?: {
|
|
1139
|
+
readonly min: number;
|
|
1140
|
+
readonly max: number;
|
|
1141
|
+
} | undefined;
|
|
1142
|
+
};
|
|
1143
|
+
readonly durationMs: {
|
|
1144
|
+
readonly value: number | null;
|
|
1145
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1146
|
+
readonly issues: readonly ({
|
|
1147
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1148
|
+
readonly locator: string;
|
|
1149
|
+
} | {
|
|
1150
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1151
|
+
readonly runId: string;
|
|
1152
|
+
readonly plannedAttemptId: string;
|
|
1153
|
+
readonly evalId: string;
|
|
1154
|
+
readonly attemptOrdinal: number;
|
|
1155
|
+
} | {
|
|
1156
|
+
readonly code: "score-contribution-unavailable";
|
|
1157
|
+
readonly locator: string;
|
|
1158
|
+
readonly entryId: string;
|
|
1159
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1160
|
+
} | {
|
|
1161
|
+
readonly code: "score-partial";
|
|
1162
|
+
readonly locator: string | null;
|
|
1163
|
+
})[];
|
|
1164
|
+
readonly basis: "eval" | "planned-attempt";
|
|
1165
|
+
readonly total: number;
|
|
1166
|
+
readonly refs: readonly {
|
|
1167
|
+
readonly identity: {
|
|
1168
|
+
readonly kind: "attempt";
|
|
1169
|
+
readonly locator: string;
|
|
1170
|
+
};
|
|
1171
|
+
}[];
|
|
1172
|
+
readonly samples: number;
|
|
1173
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1174
|
+
readonly bounds?: {
|
|
1175
|
+
readonly min: number;
|
|
1176
|
+
readonly max: number;
|
|
1177
|
+
} | undefined;
|
|
1178
|
+
};
|
|
1179
|
+
readonly tokens: {
|
|
1180
|
+
readonly value: number | null;
|
|
1181
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1182
|
+
readonly issues: readonly ({
|
|
1183
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1184
|
+
readonly locator: string;
|
|
1185
|
+
} | {
|
|
1186
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1187
|
+
readonly runId: string;
|
|
1188
|
+
readonly plannedAttemptId: string;
|
|
1189
|
+
readonly evalId: string;
|
|
1190
|
+
readonly attemptOrdinal: number;
|
|
1191
|
+
} | {
|
|
1192
|
+
readonly code: "score-contribution-unavailable";
|
|
1193
|
+
readonly locator: string;
|
|
1194
|
+
readonly entryId: string;
|
|
1195
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1196
|
+
} | {
|
|
1197
|
+
readonly code: "score-partial";
|
|
1198
|
+
readonly locator: string | null;
|
|
1199
|
+
})[];
|
|
1200
|
+
readonly basis: "eval" | "planned-attempt";
|
|
1201
|
+
readonly total: number;
|
|
1202
|
+
readonly refs: readonly {
|
|
1203
|
+
readonly identity: {
|
|
1204
|
+
readonly kind: "attempt";
|
|
1205
|
+
readonly locator: string;
|
|
1206
|
+
};
|
|
1207
|
+
}[];
|
|
1208
|
+
readonly samples: number;
|
|
1209
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1210
|
+
readonly bounds?: {
|
|
1211
|
+
readonly min: number;
|
|
1212
|
+
readonly max: number;
|
|
1213
|
+
} | undefined;
|
|
1214
|
+
};
|
|
1215
|
+
readonly plannedAttemptId: import("../record/model/identifiers.cjs").SlotId;
|
|
1216
|
+
readonly evalId: import("../record/model/identifiers.cjs").EvalId;
|
|
1217
|
+
readonly attemptOrdinal: number;
|
|
1072
1218
|
readonly publication: {
|
|
1073
1219
|
readonly state: "pending";
|
|
1074
1220
|
} | {
|
|
1075
1221
|
readonly state: "absent";
|
|
1076
1222
|
readonly reason: "early-exit-satisfied" | "budget-exhausted" | "stopped-by-failure" | "interrupted-before-publication" | "dispatch-failed";
|
|
1077
1223
|
} | {
|
|
1078
|
-
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
1079
|
-
readonly score: {
|
|
1080
|
-
readonly value: number | null;
|
|
1081
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1082
|
-
readonly issues: readonly ({
|
|
1083
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1084
|
-
readonly locator: string;
|
|
1085
|
-
} | {
|
|
1086
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1087
|
-
readonly runId: string;
|
|
1088
|
-
readonly plannedAttemptId: string;
|
|
1089
|
-
readonly evalId: string;
|
|
1090
|
-
readonly attemptOrdinal: number;
|
|
1091
|
-
} | {
|
|
1092
|
-
readonly code: "score-contribution-unavailable";
|
|
1093
|
-
readonly locator: string;
|
|
1094
|
-
readonly entryId: string;
|
|
1095
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1096
|
-
} | {
|
|
1097
|
-
readonly code: "score-partial";
|
|
1098
|
-
readonly locator: string | null;
|
|
1099
|
-
})[];
|
|
1100
|
-
readonly basis: "eval" | "planned-attempt";
|
|
1101
|
-
readonly total: number;
|
|
1102
|
-
readonly refs: readonly {
|
|
1103
|
-
readonly identity: {
|
|
1104
|
-
readonly kind: "attempt";
|
|
1105
|
-
readonly locator: string;
|
|
1106
|
-
};
|
|
1107
|
-
}[];
|
|
1108
|
-
readonly samples: number;
|
|
1109
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1110
|
-
readonly bounds?: {
|
|
1111
|
-
readonly min: number;
|
|
1112
|
-
readonly max: number;
|
|
1113
|
-
} | undefined;
|
|
1114
|
-
};
|
|
1115
|
-
readonly costUSD: {
|
|
1116
|
-
readonly value: number | null;
|
|
1117
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1118
|
-
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
1119
|
-
readonly issues: readonly ({
|
|
1120
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1121
|
-
readonly locator: string;
|
|
1122
|
-
} | {
|
|
1123
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1124
|
-
readonly runId: string;
|
|
1125
|
-
readonly plannedAttemptId: string;
|
|
1126
|
-
readonly evalId: string;
|
|
1127
|
-
readonly attemptOrdinal: number;
|
|
1128
|
-
} | {
|
|
1129
|
-
readonly code: "score-contribution-unavailable";
|
|
1130
|
-
readonly locator: string;
|
|
1131
|
-
readonly entryId: string;
|
|
1132
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1133
|
-
} | {
|
|
1134
|
-
readonly code: "score-partial";
|
|
1135
|
-
readonly locator: string | null;
|
|
1136
|
-
})[];
|
|
1137
|
-
readonly basis: "eval" | "planned-attempt";
|
|
1138
|
-
readonly total: number;
|
|
1139
|
-
readonly refs: readonly {
|
|
1140
|
-
readonly identity: {
|
|
1141
|
-
readonly kind: "attempt";
|
|
1142
|
-
readonly locator: string;
|
|
1143
|
-
};
|
|
1144
|
-
}[];
|
|
1145
|
-
readonly samples: number;
|
|
1146
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1147
|
-
readonly bounds?: {
|
|
1148
|
-
readonly min: number;
|
|
1149
|
-
readonly max: number;
|
|
1150
|
-
} | undefined;
|
|
1151
|
-
};
|
|
1152
|
-
readonly durationMs: {
|
|
1153
|
-
readonly value: number | null;
|
|
1154
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1155
|
-
readonly issues: readonly ({
|
|
1156
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1157
|
-
readonly locator: string;
|
|
1158
|
-
} | {
|
|
1159
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1160
|
-
readonly runId: string;
|
|
1161
|
-
readonly plannedAttemptId: string;
|
|
1162
|
-
readonly evalId: string;
|
|
1163
|
-
readonly attemptOrdinal: number;
|
|
1164
|
-
} | {
|
|
1165
|
-
readonly code: "score-contribution-unavailable";
|
|
1166
|
-
readonly locator: string;
|
|
1167
|
-
readonly entryId: string;
|
|
1168
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1169
|
-
} | {
|
|
1170
|
-
readonly code: "score-partial";
|
|
1171
|
-
readonly locator: string | null;
|
|
1172
|
-
})[];
|
|
1173
|
-
readonly basis: "eval" | "planned-attempt";
|
|
1174
|
-
readonly total: number;
|
|
1175
|
-
readonly refs: readonly {
|
|
1176
|
-
readonly identity: {
|
|
1177
|
-
readonly kind: "attempt";
|
|
1178
|
-
readonly locator: string;
|
|
1179
|
-
};
|
|
1180
|
-
}[];
|
|
1181
|
-
readonly samples: number;
|
|
1182
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1183
|
-
readonly bounds?: {
|
|
1184
|
-
readonly min: number;
|
|
1185
|
-
readonly max: number;
|
|
1186
|
-
} | undefined;
|
|
1187
|
-
};
|
|
1188
|
-
readonly tokens: {
|
|
1189
|
-
readonly value: number | null;
|
|
1190
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
1191
|
-
readonly issues: readonly ({
|
|
1192
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
1193
|
-
readonly locator: string;
|
|
1194
|
-
} | {
|
|
1195
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
1196
|
-
readonly runId: string;
|
|
1197
|
-
readonly plannedAttemptId: string;
|
|
1198
|
-
readonly evalId: string;
|
|
1199
|
-
readonly attemptOrdinal: number;
|
|
1200
|
-
} | {
|
|
1201
|
-
readonly code: "score-contribution-unavailable";
|
|
1202
|
-
readonly locator: string;
|
|
1203
|
-
readonly entryId: string;
|
|
1204
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
1205
|
-
} | {
|
|
1206
|
-
readonly code: "score-partial";
|
|
1207
|
-
readonly locator: string | null;
|
|
1208
|
-
})[];
|
|
1209
|
-
readonly basis: "eval" | "planned-attempt";
|
|
1210
|
-
readonly total: number;
|
|
1211
|
-
readonly refs: readonly {
|
|
1212
|
-
readonly identity: {
|
|
1213
|
-
readonly kind: "attempt";
|
|
1214
|
-
readonly locator: string;
|
|
1215
|
-
};
|
|
1216
|
-
}[];
|
|
1217
|
-
readonly samples: number;
|
|
1218
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
1219
|
-
readonly bounds?: {
|
|
1220
|
-
readonly min: number;
|
|
1221
|
-
readonly max: number;
|
|
1222
|
-
} | undefined;
|
|
1223
|
-
};
|
|
1224
1224
|
readonly originPlannedAttemptId: string;
|
|
1225
1225
|
readonly state: "published";
|
|
1226
1226
|
readonly action: "accepted" | "executed" | "carried";
|
|
@@ -1232,7 +1232,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
1232
1232
|
}[];
|
|
1233
1233
|
};
|
|
1234
1234
|
}) | ({
|
|
1235
|
-
readonly protocol: "niceeval.query/
|
|
1235
|
+
readonly protocol: "niceeval.query/v3";
|
|
1236
1236
|
readonly outcome: "success";
|
|
1237
1237
|
readonly behaviorVersion: "2.0.0";
|
|
1238
1238
|
readonly source: {
|
|
@@ -2258,9 +2258,6 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
2258
2258
|
readonly groupPath: readonly string[];
|
|
2259
2259
|
readonly members: readonly {
|
|
2260
2260
|
readonly runId: string;
|
|
2261
|
-
readonly plannedAttemptId: string;
|
|
2262
|
-
readonly evalId: string;
|
|
2263
|
-
readonly attemptOrdinal: number;
|
|
2264
2261
|
readonly labels: {
|
|
2265
2262
|
readonly [x: string]: {
|
|
2266
2263
|
readonly state: "available";
|
|
@@ -2271,158 +2268,161 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
2271
2268
|
readonly state: "unavailable";
|
|
2272
2269
|
};
|
|
2273
2270
|
};
|
|
2271
|
+
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
2272
|
+
readonly score: {
|
|
2273
|
+
readonly value: number | null;
|
|
2274
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2275
|
+
readonly issues: readonly ({
|
|
2276
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2277
|
+
readonly locator: string;
|
|
2278
|
+
} | {
|
|
2279
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2280
|
+
readonly runId: string;
|
|
2281
|
+
readonly plannedAttemptId: string;
|
|
2282
|
+
readonly evalId: string;
|
|
2283
|
+
readonly attemptOrdinal: number;
|
|
2284
|
+
} | {
|
|
2285
|
+
readonly code: "score-contribution-unavailable";
|
|
2286
|
+
readonly locator: string;
|
|
2287
|
+
readonly entryId: string;
|
|
2288
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2289
|
+
} | {
|
|
2290
|
+
readonly code: "score-partial";
|
|
2291
|
+
readonly locator: string | null;
|
|
2292
|
+
})[];
|
|
2293
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2294
|
+
readonly total: number;
|
|
2295
|
+
readonly refs: readonly {
|
|
2296
|
+
readonly identity: {
|
|
2297
|
+
readonly kind: "attempt";
|
|
2298
|
+
readonly locator: string;
|
|
2299
|
+
};
|
|
2300
|
+
}[];
|
|
2301
|
+
readonly samples: number;
|
|
2302
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2303
|
+
readonly bounds?: {
|
|
2304
|
+
readonly min: number;
|
|
2305
|
+
readonly max: number;
|
|
2306
|
+
} | undefined;
|
|
2307
|
+
};
|
|
2308
|
+
readonly costUSD: {
|
|
2309
|
+
readonly value: number | null;
|
|
2310
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2311
|
+
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
2312
|
+
readonly issues: readonly ({
|
|
2313
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2314
|
+
readonly locator: string;
|
|
2315
|
+
} | {
|
|
2316
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2317
|
+
readonly runId: string;
|
|
2318
|
+
readonly plannedAttemptId: string;
|
|
2319
|
+
readonly evalId: string;
|
|
2320
|
+
readonly attemptOrdinal: number;
|
|
2321
|
+
} | {
|
|
2322
|
+
readonly code: "score-contribution-unavailable";
|
|
2323
|
+
readonly locator: string;
|
|
2324
|
+
readonly entryId: string;
|
|
2325
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2326
|
+
} | {
|
|
2327
|
+
readonly code: "score-partial";
|
|
2328
|
+
readonly locator: string | null;
|
|
2329
|
+
})[];
|
|
2330
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2331
|
+
readonly total: number;
|
|
2332
|
+
readonly refs: readonly {
|
|
2333
|
+
readonly identity: {
|
|
2334
|
+
readonly kind: "attempt";
|
|
2335
|
+
readonly locator: string;
|
|
2336
|
+
};
|
|
2337
|
+
}[];
|
|
2338
|
+
readonly samples: number;
|
|
2339
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2340
|
+
readonly bounds?: {
|
|
2341
|
+
readonly min: number;
|
|
2342
|
+
readonly max: number;
|
|
2343
|
+
} | undefined;
|
|
2344
|
+
};
|
|
2345
|
+
readonly durationMs: {
|
|
2346
|
+
readonly value: number | null;
|
|
2347
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2348
|
+
readonly issues: readonly ({
|
|
2349
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2350
|
+
readonly locator: string;
|
|
2351
|
+
} | {
|
|
2352
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2353
|
+
readonly runId: string;
|
|
2354
|
+
readonly plannedAttemptId: string;
|
|
2355
|
+
readonly evalId: string;
|
|
2356
|
+
readonly attemptOrdinal: number;
|
|
2357
|
+
} | {
|
|
2358
|
+
readonly code: "score-contribution-unavailable";
|
|
2359
|
+
readonly locator: string;
|
|
2360
|
+
readonly entryId: string;
|
|
2361
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2362
|
+
} | {
|
|
2363
|
+
readonly code: "score-partial";
|
|
2364
|
+
readonly locator: string | null;
|
|
2365
|
+
})[];
|
|
2366
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2367
|
+
readonly total: number;
|
|
2368
|
+
readonly refs: readonly {
|
|
2369
|
+
readonly identity: {
|
|
2370
|
+
readonly kind: "attempt";
|
|
2371
|
+
readonly locator: string;
|
|
2372
|
+
};
|
|
2373
|
+
}[];
|
|
2374
|
+
readonly samples: number;
|
|
2375
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2376
|
+
readonly bounds?: {
|
|
2377
|
+
readonly min: number;
|
|
2378
|
+
readonly max: number;
|
|
2379
|
+
} | undefined;
|
|
2380
|
+
};
|
|
2381
|
+
readonly tokens: {
|
|
2382
|
+
readonly value: number | null;
|
|
2383
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2384
|
+
readonly issues: readonly ({
|
|
2385
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2386
|
+
readonly locator: string;
|
|
2387
|
+
} | {
|
|
2388
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2389
|
+
readonly runId: string;
|
|
2390
|
+
readonly plannedAttemptId: string;
|
|
2391
|
+
readonly evalId: string;
|
|
2392
|
+
readonly attemptOrdinal: number;
|
|
2393
|
+
} | {
|
|
2394
|
+
readonly code: "score-contribution-unavailable";
|
|
2395
|
+
readonly locator: string;
|
|
2396
|
+
readonly entryId: string;
|
|
2397
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2398
|
+
} | {
|
|
2399
|
+
readonly code: "score-partial";
|
|
2400
|
+
readonly locator: string | null;
|
|
2401
|
+
})[];
|
|
2402
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2403
|
+
readonly total: number;
|
|
2404
|
+
readonly refs: readonly {
|
|
2405
|
+
readonly identity: {
|
|
2406
|
+
readonly kind: "attempt";
|
|
2407
|
+
readonly locator: string;
|
|
2408
|
+
};
|
|
2409
|
+
}[];
|
|
2410
|
+
readonly samples: number;
|
|
2411
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2412
|
+
readonly bounds?: {
|
|
2413
|
+
readonly min: number;
|
|
2414
|
+
readonly max: number;
|
|
2415
|
+
} | undefined;
|
|
2416
|
+
};
|
|
2417
|
+
readonly plannedAttemptId: import("../record/model/identifiers.cjs").SlotId;
|
|
2418
|
+
readonly evalId: import("../record/model/identifiers.cjs").EvalId;
|
|
2419
|
+
readonly attemptOrdinal: number;
|
|
2274
2420
|
readonly publication: {
|
|
2275
2421
|
readonly state: "pending";
|
|
2276
2422
|
} | {
|
|
2277
2423
|
readonly state: "absent";
|
|
2278
2424
|
readonly reason: "early-exit-satisfied" | "budget-exhausted" | "stopped-by-failure" | "interrupted-before-publication" | "dispatch-failed";
|
|
2279
2425
|
} | {
|
|
2280
|
-
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
2281
|
-
readonly score: {
|
|
2282
|
-
readonly value: number | null;
|
|
2283
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2284
|
-
readonly issues: readonly ({
|
|
2285
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2286
|
-
readonly locator: string;
|
|
2287
|
-
} | {
|
|
2288
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2289
|
-
readonly runId: string;
|
|
2290
|
-
readonly plannedAttemptId: string;
|
|
2291
|
-
readonly evalId: string;
|
|
2292
|
-
readonly attemptOrdinal: number;
|
|
2293
|
-
} | {
|
|
2294
|
-
readonly code: "score-contribution-unavailable";
|
|
2295
|
-
readonly locator: string;
|
|
2296
|
-
readonly entryId: string;
|
|
2297
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2298
|
-
} | {
|
|
2299
|
-
readonly code: "score-partial";
|
|
2300
|
-
readonly locator: string | null;
|
|
2301
|
-
})[];
|
|
2302
|
-
readonly basis: "eval" | "planned-attempt";
|
|
2303
|
-
readonly total: number;
|
|
2304
|
-
readonly refs: readonly {
|
|
2305
|
-
readonly identity: {
|
|
2306
|
-
readonly kind: "attempt";
|
|
2307
|
-
readonly locator: string;
|
|
2308
|
-
};
|
|
2309
|
-
}[];
|
|
2310
|
-
readonly samples: number;
|
|
2311
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2312
|
-
readonly bounds?: {
|
|
2313
|
-
readonly min: number;
|
|
2314
|
-
readonly max: number;
|
|
2315
|
-
} | undefined;
|
|
2316
|
-
};
|
|
2317
|
-
readonly costUSD: {
|
|
2318
|
-
readonly value: number | null;
|
|
2319
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2320
|
-
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
2321
|
-
readonly issues: readonly ({
|
|
2322
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2323
|
-
readonly locator: string;
|
|
2324
|
-
} | {
|
|
2325
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2326
|
-
readonly runId: string;
|
|
2327
|
-
readonly plannedAttemptId: string;
|
|
2328
|
-
readonly evalId: string;
|
|
2329
|
-
readonly attemptOrdinal: number;
|
|
2330
|
-
} | {
|
|
2331
|
-
readonly code: "score-contribution-unavailable";
|
|
2332
|
-
readonly locator: string;
|
|
2333
|
-
readonly entryId: string;
|
|
2334
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2335
|
-
} | {
|
|
2336
|
-
readonly code: "score-partial";
|
|
2337
|
-
readonly locator: string | null;
|
|
2338
|
-
})[];
|
|
2339
|
-
readonly basis: "eval" | "planned-attempt";
|
|
2340
|
-
readonly total: number;
|
|
2341
|
-
readonly refs: readonly {
|
|
2342
|
-
readonly identity: {
|
|
2343
|
-
readonly kind: "attempt";
|
|
2344
|
-
readonly locator: string;
|
|
2345
|
-
};
|
|
2346
|
-
}[];
|
|
2347
|
-
readonly samples: number;
|
|
2348
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2349
|
-
readonly bounds?: {
|
|
2350
|
-
readonly min: number;
|
|
2351
|
-
readonly max: number;
|
|
2352
|
-
} | undefined;
|
|
2353
|
-
};
|
|
2354
|
-
readonly durationMs: {
|
|
2355
|
-
readonly value: number | null;
|
|
2356
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2357
|
-
readonly issues: readonly ({
|
|
2358
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2359
|
-
readonly locator: string;
|
|
2360
|
-
} | {
|
|
2361
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2362
|
-
readonly runId: string;
|
|
2363
|
-
readonly plannedAttemptId: string;
|
|
2364
|
-
readonly evalId: string;
|
|
2365
|
-
readonly attemptOrdinal: number;
|
|
2366
|
-
} | {
|
|
2367
|
-
readonly code: "score-contribution-unavailable";
|
|
2368
|
-
readonly locator: string;
|
|
2369
|
-
readonly entryId: string;
|
|
2370
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2371
|
-
} | {
|
|
2372
|
-
readonly code: "score-partial";
|
|
2373
|
-
readonly locator: string | null;
|
|
2374
|
-
})[];
|
|
2375
|
-
readonly basis: "eval" | "planned-attempt";
|
|
2376
|
-
readonly total: number;
|
|
2377
|
-
readonly refs: readonly {
|
|
2378
|
-
readonly identity: {
|
|
2379
|
-
readonly kind: "attempt";
|
|
2380
|
-
readonly locator: string;
|
|
2381
|
-
};
|
|
2382
|
-
}[];
|
|
2383
|
-
readonly samples: number;
|
|
2384
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2385
|
-
readonly bounds?: {
|
|
2386
|
-
readonly min: number;
|
|
2387
|
-
readonly max: number;
|
|
2388
|
-
} | undefined;
|
|
2389
|
-
};
|
|
2390
|
-
readonly tokens: {
|
|
2391
|
-
readonly value: number | null;
|
|
2392
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2393
|
-
readonly issues: readonly ({
|
|
2394
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2395
|
-
readonly locator: string;
|
|
2396
|
-
} | {
|
|
2397
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2398
|
-
readonly runId: string;
|
|
2399
|
-
readonly plannedAttemptId: string;
|
|
2400
|
-
readonly evalId: string;
|
|
2401
|
-
readonly attemptOrdinal: number;
|
|
2402
|
-
} | {
|
|
2403
|
-
readonly code: "score-contribution-unavailable";
|
|
2404
|
-
readonly locator: string;
|
|
2405
|
-
readonly entryId: string;
|
|
2406
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2407
|
-
} | {
|
|
2408
|
-
readonly code: "score-partial";
|
|
2409
|
-
readonly locator: string | null;
|
|
2410
|
-
})[];
|
|
2411
|
-
readonly basis: "eval" | "planned-attempt";
|
|
2412
|
-
readonly total: number;
|
|
2413
|
-
readonly refs: readonly {
|
|
2414
|
-
readonly identity: {
|
|
2415
|
-
readonly kind: "attempt";
|
|
2416
|
-
readonly locator: string;
|
|
2417
|
-
};
|
|
2418
|
-
}[];
|
|
2419
|
-
readonly samples: number;
|
|
2420
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2421
|
-
readonly bounds?: {
|
|
2422
|
-
readonly min: number;
|
|
2423
|
-
readonly max: number;
|
|
2424
|
-
} | undefined;
|
|
2425
|
-
};
|
|
2426
2426
|
readonly originPlannedAttemptId: string;
|
|
2427
2427
|
readonly state: "published";
|
|
2428
2428
|
readonly action: "accepted" | "executed" | "carried";
|
|
@@ -2434,7 +2434,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
2434
2434
|
}[];
|
|
2435
2435
|
};
|
|
2436
2436
|
}) | ({
|
|
2437
|
-
readonly protocol: "niceeval.query/
|
|
2437
|
+
readonly protocol: "niceeval.query/v3";
|
|
2438
2438
|
readonly outcome: "success";
|
|
2439
2439
|
readonly behaviorVersion: "2.0.0";
|
|
2440
2440
|
readonly source: {
|
|
@@ -2474,7 +2474,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
2474
2474
|
}[];
|
|
2475
2475
|
readonly continuation: string | undefined;
|
|
2476
2476
|
}) | ({
|
|
2477
|
-
readonly protocol: "niceeval.query/
|
|
2477
|
+
readonly protocol: "niceeval.query/v3";
|
|
2478
2478
|
readonly outcome: "success";
|
|
2479
2479
|
readonly behaviorVersion: "2.0.0";
|
|
2480
2480
|
readonly source: {
|
|
@@ -2543,36 +2543,9 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
2543
2543
|
readonly missing: number;
|
|
2544
2544
|
};
|
|
2545
2545
|
readonly plannedAttempts: readonly {
|
|
2546
|
-
readonly publication: {
|
|
2547
|
-
readonly state: "pending";
|
|
2548
|
-
} | {
|
|
2549
|
-
readonly state: "absent";
|
|
2550
|
-
readonly reason: "early-exit-satisfied" | "budget-exhausted" | "stopped-by-failure" | "interrupted-before-publication" | "dispatch-failed";
|
|
2551
|
-
} | {
|
|
2552
|
-
readonly originPlannedAttemptId: string;
|
|
2553
|
-
readonly state: "published";
|
|
2554
|
-
readonly action: "accepted" | "executed" | "carried";
|
|
2555
|
-
readonly attemptId: string;
|
|
2556
|
-
readonly attemptLocator: string;
|
|
2557
|
-
readonly originRunId: string;
|
|
2558
|
-
};
|
|
2559
2546
|
readonly state: "accepted" | "pending" | "interrupted" | "executed" | "carried" | "not-dispatched";
|
|
2560
|
-
readonly locator: string | null;
|
|
2561
2547
|
readonly relation: "reference" | "origin" | null;
|
|
2562
2548
|
readonly outcome: "errored" | "cancelled" | "completed" | "interrupted" | null;
|
|
2563
|
-
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
2564
|
-
readonly score: {
|
|
2565
|
-
readonly state: "not-scored";
|
|
2566
|
-
} | {
|
|
2567
|
-
readonly state: "complete";
|
|
2568
|
-
readonly earned: number;
|
|
2569
|
-
readonly possible: number;
|
|
2570
|
-
} | {
|
|
2571
|
-
readonly state: "unavailable";
|
|
2572
|
-
readonly earned: number;
|
|
2573
|
-
readonly possible: number;
|
|
2574
|
-
readonly unavailable: number;
|
|
2575
|
-
} | null;
|
|
2576
2549
|
readonly coverage: {
|
|
2577
2550
|
readonly state: "complete" | "partial" | "unavailable" | "not-recorded" | "invalid";
|
|
2578
2551
|
readonly facts: readonly {
|
|
@@ -2830,10 +2803,169 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
2830
2803
|
readonly reason: string;
|
|
2831
2804
|
};
|
|
2832
2805
|
})[];
|
|
2806
|
+
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
2807
|
+
readonly score: {
|
|
2808
|
+
readonly value: number | null;
|
|
2809
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2810
|
+
readonly issues: readonly ({
|
|
2811
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2812
|
+
readonly locator: string;
|
|
2813
|
+
} | {
|
|
2814
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2815
|
+
readonly runId: string;
|
|
2816
|
+
readonly plannedAttemptId: string;
|
|
2817
|
+
readonly evalId: string;
|
|
2818
|
+
readonly attemptOrdinal: number;
|
|
2819
|
+
} | {
|
|
2820
|
+
readonly code: "score-contribution-unavailable";
|
|
2821
|
+
readonly locator: string;
|
|
2822
|
+
readonly entryId: string;
|
|
2823
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2824
|
+
} | {
|
|
2825
|
+
readonly code: "score-partial";
|
|
2826
|
+
readonly locator: string | null;
|
|
2827
|
+
})[];
|
|
2828
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2829
|
+
readonly total: number;
|
|
2830
|
+
readonly refs: readonly {
|
|
2831
|
+
readonly identity: {
|
|
2832
|
+
readonly kind: "attempt";
|
|
2833
|
+
readonly locator: string;
|
|
2834
|
+
};
|
|
2835
|
+
}[];
|
|
2836
|
+
readonly samples: number;
|
|
2837
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2838
|
+
readonly bounds?: {
|
|
2839
|
+
readonly min: number;
|
|
2840
|
+
readonly max: number;
|
|
2841
|
+
} | undefined;
|
|
2842
|
+
};
|
|
2843
|
+
readonly costUSD: {
|
|
2844
|
+
readonly value: number | null;
|
|
2845
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2846
|
+
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
2847
|
+
readonly issues: readonly ({
|
|
2848
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2849
|
+
readonly locator: string;
|
|
2850
|
+
} | {
|
|
2851
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2852
|
+
readonly runId: string;
|
|
2853
|
+
readonly plannedAttemptId: string;
|
|
2854
|
+
readonly evalId: string;
|
|
2855
|
+
readonly attemptOrdinal: number;
|
|
2856
|
+
} | {
|
|
2857
|
+
readonly code: "score-contribution-unavailable";
|
|
2858
|
+
readonly locator: string;
|
|
2859
|
+
readonly entryId: string;
|
|
2860
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2861
|
+
} | {
|
|
2862
|
+
readonly code: "score-partial";
|
|
2863
|
+
readonly locator: string | null;
|
|
2864
|
+
})[];
|
|
2865
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2866
|
+
readonly total: number;
|
|
2867
|
+
readonly refs: readonly {
|
|
2868
|
+
readonly identity: {
|
|
2869
|
+
readonly kind: "attempt";
|
|
2870
|
+
readonly locator: string;
|
|
2871
|
+
};
|
|
2872
|
+
}[];
|
|
2873
|
+
readonly samples: number;
|
|
2874
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2875
|
+
readonly bounds?: {
|
|
2876
|
+
readonly min: number;
|
|
2877
|
+
readonly max: number;
|
|
2878
|
+
} | undefined;
|
|
2879
|
+
};
|
|
2880
|
+
readonly durationMs: {
|
|
2881
|
+
readonly value: number | null;
|
|
2882
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2883
|
+
readonly issues: readonly ({
|
|
2884
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2885
|
+
readonly locator: string;
|
|
2886
|
+
} | {
|
|
2887
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2888
|
+
readonly runId: string;
|
|
2889
|
+
readonly plannedAttemptId: string;
|
|
2890
|
+
readonly evalId: string;
|
|
2891
|
+
readonly attemptOrdinal: number;
|
|
2892
|
+
} | {
|
|
2893
|
+
readonly code: "score-contribution-unavailable";
|
|
2894
|
+
readonly locator: string;
|
|
2895
|
+
readonly entryId: string;
|
|
2896
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2897
|
+
} | {
|
|
2898
|
+
readonly code: "score-partial";
|
|
2899
|
+
readonly locator: string | null;
|
|
2900
|
+
})[];
|
|
2901
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2902
|
+
readonly total: number;
|
|
2903
|
+
readonly refs: readonly {
|
|
2904
|
+
readonly identity: {
|
|
2905
|
+
readonly kind: "attempt";
|
|
2906
|
+
readonly locator: string;
|
|
2907
|
+
};
|
|
2908
|
+
}[];
|
|
2909
|
+
readonly samples: number;
|
|
2910
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2911
|
+
readonly bounds?: {
|
|
2912
|
+
readonly min: number;
|
|
2913
|
+
readonly max: number;
|
|
2914
|
+
} | undefined;
|
|
2915
|
+
};
|
|
2916
|
+
readonly tokens: {
|
|
2917
|
+
readonly value: number | null;
|
|
2918
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
2919
|
+
readonly issues: readonly ({
|
|
2920
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
2921
|
+
readonly locator: string;
|
|
2922
|
+
} | {
|
|
2923
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
2924
|
+
readonly runId: string;
|
|
2925
|
+
readonly plannedAttemptId: string;
|
|
2926
|
+
readonly evalId: string;
|
|
2927
|
+
readonly attemptOrdinal: number;
|
|
2928
|
+
} | {
|
|
2929
|
+
readonly code: "score-contribution-unavailable";
|
|
2930
|
+
readonly locator: string;
|
|
2931
|
+
readonly entryId: string;
|
|
2932
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
2933
|
+
} | {
|
|
2934
|
+
readonly code: "score-partial";
|
|
2935
|
+
readonly locator: string | null;
|
|
2936
|
+
})[];
|
|
2937
|
+
readonly basis: "eval" | "planned-attempt";
|
|
2938
|
+
readonly total: number;
|
|
2939
|
+
readonly refs: readonly {
|
|
2940
|
+
readonly identity: {
|
|
2941
|
+
readonly kind: "attempt";
|
|
2942
|
+
readonly locator: string;
|
|
2943
|
+
};
|
|
2944
|
+
}[];
|
|
2945
|
+
readonly samples: number;
|
|
2946
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
2947
|
+
readonly bounds?: {
|
|
2948
|
+
readonly min: number;
|
|
2949
|
+
readonly max: number;
|
|
2950
|
+
} | undefined;
|
|
2951
|
+
};
|
|
2952
|
+
readonly plannedAttemptId: import("../record/model/identifiers.cjs").SlotId;
|
|
2833
2953
|
readonly evalId: import("../record/model/identifiers.cjs").EvalId;
|
|
2834
2954
|
readonly attemptOrdinal: number;
|
|
2955
|
+
readonly publication: {
|
|
2956
|
+
readonly state: "pending";
|
|
2957
|
+
} | {
|
|
2958
|
+
readonly state: "absent";
|
|
2959
|
+
readonly reason: "early-exit-satisfied" | "budget-exhausted" | "stopped-by-failure" | "interrupted-before-publication" | "dispatch-failed";
|
|
2960
|
+
} | {
|
|
2961
|
+
readonly originPlannedAttemptId: string;
|
|
2962
|
+
readonly state: "published";
|
|
2963
|
+
readonly action: "accepted" | "executed" | "carried";
|
|
2964
|
+
readonly attemptId: string;
|
|
2965
|
+
readonly attemptLocator: string;
|
|
2966
|
+
readonly originRunId: string;
|
|
2967
|
+
};
|
|
2835
2968
|
readonly executionIdentityDigest: import("../record/model/identifiers.cjs").ExecutionIdentityDigest;
|
|
2836
|
-
readonly plannedAttemptId: import("../record/model/identifiers.cjs").SlotId;
|
|
2837
2969
|
}[];
|
|
2838
2970
|
readonly coverage: {
|
|
2839
2971
|
readonly state: "complete" | "partial" | "unavailable" | "not-recorded" | "invalid";
|
|
@@ -3218,7 +3350,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
3218
3350
|
}[];
|
|
3219
3351
|
};
|
|
3220
3352
|
}) | ({
|
|
3221
|
-
readonly protocol: "niceeval.query/
|
|
3353
|
+
readonly protocol: "niceeval.query/v3";
|
|
3222
3354
|
readonly outcome: "success";
|
|
3223
3355
|
readonly behaviorVersion: "2.0.0";
|
|
3224
3356
|
readonly source: {
|
|
@@ -3435,7 +3567,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
3435
3567
|
})[];
|
|
3436
3568
|
};
|
|
3437
3569
|
}) | ({
|
|
3438
|
-
readonly protocol: "niceeval.query/
|
|
3570
|
+
readonly protocol: "niceeval.query/v3";
|
|
3439
3571
|
readonly outcome: "success";
|
|
3440
3572
|
readonly behaviorVersion: "2.0.0";
|
|
3441
3573
|
readonly source: {
|
|
@@ -3510,7 +3642,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
3510
3642
|
};
|
|
3511
3643
|
};
|
|
3512
3644
|
}) | ({
|
|
3513
|
-
readonly protocol: "niceeval.query/
|
|
3645
|
+
readonly protocol: "niceeval.query/v3";
|
|
3514
3646
|
readonly outcome: "success";
|
|
3515
3647
|
readonly behaviorVersion: "2.0.0";
|
|
3516
3648
|
readonly source: {
|
|
@@ -3991,7 +4123,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
3991
4123
|
};
|
|
3992
4124
|
};
|
|
3993
4125
|
}) | ({
|
|
3994
|
-
readonly protocol: "niceeval.query/
|
|
4126
|
+
readonly protocol: "niceeval.query/v3";
|
|
3995
4127
|
readonly outcome: "success";
|
|
3996
4128
|
readonly behaviorVersion: "2.0.0";
|
|
3997
4129
|
readonly source: {
|
|
@@ -4458,7 +4590,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
4458
4590
|
};
|
|
4459
4591
|
};
|
|
4460
4592
|
}) | ({
|
|
4461
|
-
readonly protocol: "niceeval.query/
|
|
4593
|
+
readonly protocol: "niceeval.query/v3";
|
|
4462
4594
|
readonly outcome: "success";
|
|
4463
4595
|
readonly behaviorVersion: "2.0.0";
|
|
4464
4596
|
readonly source: {
|
|
@@ -4515,7 +4647,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
4515
4647
|
readonly omittedActivityCount: number;
|
|
4516
4648
|
};
|
|
4517
4649
|
}) | ({
|
|
4518
|
-
readonly protocol: "niceeval.query/
|
|
4650
|
+
readonly protocol: "niceeval.query/v3";
|
|
4519
4651
|
readonly outcome: "success";
|
|
4520
4652
|
readonly behaviorVersion: "2.0.0";
|
|
4521
4653
|
readonly source: {
|
|
@@ -4998,7 +5130,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
4998
5130
|
readonly omittedCallCount?: number | undefined;
|
|
4999
5131
|
};
|
|
5000
5132
|
}) | ({
|
|
5001
|
-
readonly protocol: "niceeval.query/
|
|
5133
|
+
readonly protocol: "niceeval.query/v3";
|
|
5002
5134
|
readonly outcome: "success";
|
|
5003
5135
|
readonly behaviorVersion: "2.0.0";
|
|
5004
5136
|
readonly source: {
|
|
@@ -5091,7 +5223,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
5091
5223
|
}[];
|
|
5092
5224
|
};
|
|
5093
5225
|
}) | ({
|
|
5094
|
-
readonly protocol: "niceeval.query/
|
|
5226
|
+
readonly protocol: "niceeval.query/v3";
|
|
5095
5227
|
readonly outcome: "success";
|
|
5096
5228
|
readonly behaviorVersion: "2.0.0";
|
|
5097
5229
|
readonly source: {
|
|
@@ -6153,9 +6285,6 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
6153
6285
|
readonly groupPath: readonly string[];
|
|
6154
6286
|
readonly members: readonly {
|
|
6155
6287
|
readonly runId: string;
|
|
6156
|
-
readonly plannedAttemptId: string;
|
|
6157
|
-
readonly evalId: string;
|
|
6158
|
-
readonly attemptOrdinal: number;
|
|
6159
6288
|
readonly labels: {
|
|
6160
6289
|
readonly [x: string]: {
|
|
6161
6290
|
readonly state: "available";
|
|
@@ -6166,158 +6295,161 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
6166
6295
|
readonly state: "unavailable";
|
|
6167
6296
|
};
|
|
6168
6297
|
};
|
|
6298
|
+
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
6299
|
+
readonly score: {
|
|
6300
|
+
readonly value: number | null;
|
|
6301
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6302
|
+
readonly issues: readonly ({
|
|
6303
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6304
|
+
readonly locator: string;
|
|
6305
|
+
} | {
|
|
6306
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6307
|
+
readonly runId: string;
|
|
6308
|
+
readonly plannedAttemptId: string;
|
|
6309
|
+
readonly evalId: string;
|
|
6310
|
+
readonly attemptOrdinal: number;
|
|
6311
|
+
} | {
|
|
6312
|
+
readonly code: "score-contribution-unavailable";
|
|
6313
|
+
readonly locator: string;
|
|
6314
|
+
readonly entryId: string;
|
|
6315
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6316
|
+
} | {
|
|
6317
|
+
readonly code: "score-partial";
|
|
6318
|
+
readonly locator: string | null;
|
|
6319
|
+
})[];
|
|
6320
|
+
readonly basis: "eval" | "planned-attempt";
|
|
6321
|
+
readonly total: number;
|
|
6322
|
+
readonly refs: readonly {
|
|
6323
|
+
readonly identity: {
|
|
6324
|
+
readonly kind: "attempt";
|
|
6325
|
+
readonly locator: string;
|
|
6326
|
+
};
|
|
6327
|
+
}[];
|
|
6328
|
+
readonly samples: number;
|
|
6329
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6330
|
+
readonly bounds?: {
|
|
6331
|
+
readonly min: number;
|
|
6332
|
+
readonly max: number;
|
|
6333
|
+
} | undefined;
|
|
6334
|
+
};
|
|
6335
|
+
readonly costUSD: {
|
|
6336
|
+
readonly value: number | null;
|
|
6337
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6338
|
+
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
6339
|
+
readonly issues: readonly ({
|
|
6340
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6341
|
+
readonly locator: string;
|
|
6342
|
+
} | {
|
|
6343
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6344
|
+
readonly runId: string;
|
|
6345
|
+
readonly plannedAttemptId: string;
|
|
6346
|
+
readonly evalId: string;
|
|
6347
|
+
readonly attemptOrdinal: number;
|
|
6348
|
+
} | {
|
|
6349
|
+
readonly code: "score-contribution-unavailable";
|
|
6350
|
+
readonly locator: string;
|
|
6351
|
+
readonly entryId: string;
|
|
6352
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6353
|
+
} | {
|
|
6354
|
+
readonly code: "score-partial";
|
|
6355
|
+
readonly locator: string | null;
|
|
6356
|
+
})[];
|
|
6357
|
+
readonly basis: "eval" | "planned-attempt";
|
|
6358
|
+
readonly total: number;
|
|
6359
|
+
readonly refs: readonly {
|
|
6360
|
+
readonly identity: {
|
|
6361
|
+
readonly kind: "attempt";
|
|
6362
|
+
readonly locator: string;
|
|
6363
|
+
};
|
|
6364
|
+
}[];
|
|
6365
|
+
readonly samples: number;
|
|
6366
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6367
|
+
readonly bounds?: {
|
|
6368
|
+
readonly min: number;
|
|
6369
|
+
readonly max: number;
|
|
6370
|
+
} | undefined;
|
|
6371
|
+
};
|
|
6372
|
+
readonly durationMs: {
|
|
6373
|
+
readonly value: number | null;
|
|
6374
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6375
|
+
readonly issues: readonly ({
|
|
6376
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6377
|
+
readonly locator: string;
|
|
6378
|
+
} | {
|
|
6379
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6380
|
+
readonly runId: string;
|
|
6381
|
+
readonly plannedAttemptId: string;
|
|
6382
|
+
readonly evalId: string;
|
|
6383
|
+
readonly attemptOrdinal: number;
|
|
6384
|
+
} | {
|
|
6385
|
+
readonly code: "score-contribution-unavailable";
|
|
6386
|
+
readonly locator: string;
|
|
6387
|
+
readonly entryId: string;
|
|
6388
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6389
|
+
} | {
|
|
6390
|
+
readonly code: "score-partial";
|
|
6391
|
+
readonly locator: string | null;
|
|
6392
|
+
})[];
|
|
6393
|
+
readonly basis: "eval" | "planned-attempt";
|
|
6394
|
+
readonly total: number;
|
|
6395
|
+
readonly refs: readonly {
|
|
6396
|
+
readonly identity: {
|
|
6397
|
+
readonly kind: "attempt";
|
|
6398
|
+
readonly locator: string;
|
|
6399
|
+
};
|
|
6400
|
+
}[];
|
|
6401
|
+
readonly samples: number;
|
|
6402
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6403
|
+
readonly bounds?: {
|
|
6404
|
+
readonly min: number;
|
|
6405
|
+
readonly max: number;
|
|
6406
|
+
} | undefined;
|
|
6407
|
+
};
|
|
6408
|
+
readonly tokens: {
|
|
6409
|
+
readonly value: number | null;
|
|
6410
|
+
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6411
|
+
readonly issues: readonly ({
|
|
6412
|
+
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6413
|
+
readonly locator: string;
|
|
6414
|
+
} | {
|
|
6415
|
+
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6416
|
+
readonly runId: string;
|
|
6417
|
+
readonly plannedAttemptId: string;
|
|
6418
|
+
readonly evalId: string;
|
|
6419
|
+
readonly attemptOrdinal: number;
|
|
6420
|
+
} | {
|
|
6421
|
+
readonly code: "score-contribution-unavailable";
|
|
6422
|
+
readonly locator: string;
|
|
6423
|
+
readonly entryId: string;
|
|
6424
|
+
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6425
|
+
} | {
|
|
6426
|
+
readonly code: "score-partial";
|
|
6427
|
+
readonly locator: string | null;
|
|
6428
|
+
})[];
|
|
6429
|
+
readonly basis: "eval" | "planned-attempt";
|
|
6430
|
+
readonly total: number;
|
|
6431
|
+
readonly refs: readonly {
|
|
6432
|
+
readonly identity: {
|
|
6433
|
+
readonly kind: "attempt";
|
|
6434
|
+
readonly locator: string;
|
|
6435
|
+
};
|
|
6436
|
+
}[];
|
|
6437
|
+
readonly samples: number;
|
|
6438
|
+
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6439
|
+
readonly bounds?: {
|
|
6440
|
+
readonly min: number;
|
|
6441
|
+
readonly max: number;
|
|
6442
|
+
} | undefined;
|
|
6443
|
+
};
|
|
6444
|
+
readonly plannedAttemptId: import("../record/model/identifiers.cjs").SlotId;
|
|
6445
|
+
readonly evalId: import("../record/model/identifiers.cjs").EvalId;
|
|
6446
|
+
readonly attemptOrdinal: number;
|
|
6169
6447
|
readonly publication: {
|
|
6170
6448
|
readonly state: "pending";
|
|
6171
6449
|
} | {
|
|
6172
6450
|
readonly state: "absent";
|
|
6173
6451
|
readonly reason: "early-exit-satisfied" | "budget-exhausted" | "stopped-by-failure" | "interrupted-before-publication" | "dispatch-failed";
|
|
6174
6452
|
} | {
|
|
6175
|
-
readonly verdict: "passed" | "failed" | "errored" | "skipped" | null;
|
|
6176
|
-
readonly score: {
|
|
6177
|
-
readonly value: number | null;
|
|
6178
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6179
|
-
readonly issues: readonly ({
|
|
6180
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6181
|
-
readonly locator: string;
|
|
6182
|
-
} | {
|
|
6183
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6184
|
-
readonly runId: string;
|
|
6185
|
-
readonly plannedAttemptId: string;
|
|
6186
|
-
readonly evalId: string;
|
|
6187
|
-
readonly attemptOrdinal: number;
|
|
6188
|
-
} | {
|
|
6189
|
-
readonly code: "score-contribution-unavailable";
|
|
6190
|
-
readonly locator: string;
|
|
6191
|
-
readonly entryId: string;
|
|
6192
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6193
|
-
} | {
|
|
6194
|
-
readonly code: "score-partial";
|
|
6195
|
-
readonly locator: string | null;
|
|
6196
|
-
})[];
|
|
6197
|
-
readonly basis: "eval" | "planned-attempt";
|
|
6198
|
-
readonly total: number;
|
|
6199
|
-
readonly refs: readonly {
|
|
6200
|
-
readonly identity: {
|
|
6201
|
-
readonly kind: "attempt";
|
|
6202
|
-
readonly locator: string;
|
|
6203
|
-
};
|
|
6204
|
-
}[];
|
|
6205
|
-
readonly samples: number;
|
|
6206
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6207
|
-
readonly bounds?: {
|
|
6208
|
-
readonly min: number;
|
|
6209
|
-
readonly max: number;
|
|
6210
|
-
} | undefined;
|
|
6211
|
-
};
|
|
6212
|
-
readonly costUSD: {
|
|
6213
|
-
readonly value: number | null;
|
|
6214
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6215
|
-
readonly source: "reported" | "estimated" | "mixed" | null;
|
|
6216
|
-
readonly issues: readonly ({
|
|
6217
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6218
|
-
readonly locator: string;
|
|
6219
|
-
} | {
|
|
6220
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6221
|
-
readonly runId: string;
|
|
6222
|
-
readonly plannedAttemptId: string;
|
|
6223
|
-
readonly evalId: string;
|
|
6224
|
-
readonly attemptOrdinal: number;
|
|
6225
|
-
} | {
|
|
6226
|
-
readonly code: "score-contribution-unavailable";
|
|
6227
|
-
readonly locator: string;
|
|
6228
|
-
readonly entryId: string;
|
|
6229
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6230
|
-
} | {
|
|
6231
|
-
readonly code: "score-partial";
|
|
6232
|
-
readonly locator: string | null;
|
|
6233
|
-
})[];
|
|
6234
|
-
readonly basis: "eval" | "planned-attempt";
|
|
6235
|
-
readonly total: number;
|
|
6236
|
-
readonly refs: readonly {
|
|
6237
|
-
readonly identity: {
|
|
6238
|
-
readonly kind: "attempt";
|
|
6239
|
-
readonly locator: string;
|
|
6240
|
-
};
|
|
6241
|
-
}[];
|
|
6242
|
-
readonly samples: number;
|
|
6243
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6244
|
-
readonly bounds?: {
|
|
6245
|
-
readonly min: number;
|
|
6246
|
-
readonly max: number;
|
|
6247
|
-
} | undefined;
|
|
6248
|
-
};
|
|
6249
|
-
readonly durationMs: {
|
|
6250
|
-
readonly value: number | null;
|
|
6251
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6252
|
-
readonly issues: readonly ({
|
|
6253
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6254
|
-
readonly locator: string;
|
|
6255
|
-
} | {
|
|
6256
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6257
|
-
readonly runId: string;
|
|
6258
|
-
readonly plannedAttemptId: string;
|
|
6259
|
-
readonly evalId: string;
|
|
6260
|
-
readonly attemptOrdinal: number;
|
|
6261
|
-
} | {
|
|
6262
|
-
readonly code: "score-contribution-unavailable";
|
|
6263
|
-
readonly locator: string;
|
|
6264
|
-
readonly entryId: string;
|
|
6265
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6266
|
-
} | {
|
|
6267
|
-
readonly code: "score-partial";
|
|
6268
|
-
readonly locator: string | null;
|
|
6269
|
-
})[];
|
|
6270
|
-
readonly basis: "eval" | "planned-attempt";
|
|
6271
|
-
readonly total: number;
|
|
6272
|
-
readonly refs: readonly {
|
|
6273
|
-
readonly identity: {
|
|
6274
|
-
readonly kind: "attempt";
|
|
6275
|
-
readonly locator: string;
|
|
6276
|
-
};
|
|
6277
|
-
}[];
|
|
6278
|
-
readonly samples: number;
|
|
6279
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6280
|
-
readonly bounds?: {
|
|
6281
|
-
readonly min: number;
|
|
6282
|
-
readonly max: number;
|
|
6283
|
-
} | undefined;
|
|
6284
|
-
};
|
|
6285
|
-
readonly tokens: {
|
|
6286
|
-
readonly value: number | null;
|
|
6287
|
-
readonly state: "partial" | "failed" | "available" | "unavailable" | "empty" | "unsupported";
|
|
6288
|
-
readonly issues: readonly ({
|
|
6289
|
-
readonly code: "assertions-not-recorded" | "assertions-revision-unsupported" | "assertions-current-invalid";
|
|
6290
|
-
readonly locator: string;
|
|
6291
|
-
} | {
|
|
6292
|
-
readonly code: "attempt-origin-missing" | "attempt-not-observed";
|
|
6293
|
-
readonly runId: string;
|
|
6294
|
-
readonly plannedAttemptId: string;
|
|
6295
|
-
readonly evalId: string;
|
|
6296
|
-
readonly attemptOrdinal: number;
|
|
6297
|
-
} | {
|
|
6298
|
-
readonly code: "score-contribution-unavailable";
|
|
6299
|
-
readonly locator: string;
|
|
6300
|
-
readonly entryId: string;
|
|
6301
|
-
readonly reason: "source-unavailable" | "not-applicable" | "evaluation-errored";
|
|
6302
|
-
} | {
|
|
6303
|
-
readonly code: "score-partial";
|
|
6304
|
-
readonly locator: string | null;
|
|
6305
|
-
})[];
|
|
6306
|
-
readonly basis: "eval" | "planned-attempt";
|
|
6307
|
-
readonly total: number;
|
|
6308
|
-
readonly refs: readonly {
|
|
6309
|
-
readonly identity: {
|
|
6310
|
-
readonly kind: "attempt";
|
|
6311
|
-
readonly locator: string;
|
|
6312
|
-
};
|
|
6313
|
-
}[];
|
|
6314
|
-
readonly samples: number;
|
|
6315
|
-
readonly unit?: "ms" | "tokens" | "USD" | "points" | undefined;
|
|
6316
|
-
readonly bounds?: {
|
|
6317
|
-
readonly min: number;
|
|
6318
|
-
readonly max: number;
|
|
6319
|
-
} | undefined;
|
|
6320
|
-
};
|
|
6321
6453
|
readonly originPlannedAttemptId: string;
|
|
6322
6454
|
readonly state: "published";
|
|
6323
6455
|
readonly action: "accepted" | "executed" | "carried";
|
|
@@ -6357,7 +6489,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
6357
6489
|
})[];
|
|
6358
6490
|
};
|
|
6359
6491
|
}) | ({
|
|
6360
|
-
readonly protocol: "niceeval.query/
|
|
6492
|
+
readonly protocol: "niceeval.query/v3";
|
|
6361
6493
|
readonly outcome: "success";
|
|
6362
6494
|
readonly behaviorVersion: "2.0.0";
|
|
6363
6495
|
readonly source: {
|
|
@@ -7294,7 +7426,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
7294
7426
|
};
|
|
7295
7427
|
};
|
|
7296
7428
|
}) | ({
|
|
7297
|
-
readonly protocol: "niceeval.query/
|
|
7429
|
+
readonly protocol: "niceeval.query/v3";
|
|
7298
7430
|
readonly outcome: "success";
|
|
7299
7431
|
readonly behaviorVersion: "2.0.0";
|
|
7300
7432
|
readonly source: {
|
|
@@ -7332,7 +7464,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
7332
7464
|
readonly state: "not-recorded";
|
|
7333
7465
|
};
|
|
7334
7466
|
}) | ({
|
|
7335
|
-
readonly protocol: "niceeval.query/
|
|
7467
|
+
readonly protocol: "niceeval.query/v3";
|
|
7336
7468
|
readonly outcome: "success";
|
|
7337
7469
|
readonly behaviorVersion: "2.0.0";
|
|
7338
7470
|
readonly source: {
|
|
@@ -7408,7 +7540,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
7408
7540
|
readonly contentsTruncated: boolean;
|
|
7409
7541
|
};
|
|
7410
7542
|
}) | ({
|
|
7411
|
-
readonly protocol: "niceeval.query/
|
|
7543
|
+
readonly protocol: "niceeval.query/v3";
|
|
7412
7544
|
readonly outcome: "success";
|
|
7413
7545
|
readonly behaviorVersion: "2.0.0";
|
|
7414
7546
|
readonly source: {
|
|
@@ -7446,7 +7578,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
7446
7578
|
readonly nextOffset: number | null;
|
|
7447
7579
|
};
|
|
7448
7580
|
}) | ({
|
|
7449
|
-
readonly protocol: "niceeval.query/
|
|
7581
|
+
readonly protocol: "niceeval.query/v3";
|
|
7450
7582
|
readonly source: {
|
|
7451
7583
|
readonly kind: "external-record" | "project-record";
|
|
7452
7584
|
readonly publicationCutoffIdentity: string;
|
|
@@ -7553,7 +7685,7 @@ export declare function inspectViewGeneration(generation: ViewGeneration, input:
|
|
|
7553
7685
|
}[];
|
|
7554
7686
|
};
|
|
7555
7687
|
}) | Readonly<{
|
|
7556
|
-
protocol: "niceeval.query/
|
|
7688
|
+
protocol: "niceeval.query/v3";
|
|
7557
7689
|
outcome: "failure";
|
|
7558
7690
|
operation: "run.list" | "run.get" | "project.get" | "overview.get" | "experiment.get" | "attempt.get" | "attempt.assertion.detail" | "attempt.assertion.image" | "attempt.trace" | "attempt.trace.detail" | "attempt.timing" | "attempt.usage" | "attempt.diff" | "attempt.sources" | "attempt.artifacts" | "attempt.artifact" | "runs.compare";
|
|
7559
7691
|
failure: Readonly<{
|