@warmdrift/kgauto-compiler 2.0.0-alpha.97 → 2.0.0-alpha.99
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/{chunk-77AHSX6Y.mjs → chunk-2YCTWV4S.mjs} +1 -1
- package/dist/{chunk-AFYBZRQV.mjs → chunk-G33BUB43.mjs} +96 -20
- package/dist/{chunk-IEEROFZW.mjs → chunk-L3VBUB4I.mjs} +1 -1
- package/dist/{chunk-HHIXHL4O.mjs → chunk-PY4CBZFF.mjs} +12 -2
- package/dist/{compile-C2IfAGV1.d.mts → compile-hh_l7COL.d.mts} +1 -1
- package/dist/{compile-BlMXv6QT.d.ts → compile-nNpq-6yu.d.ts} +1 -1
- package/dist/glassbox/index.d.mts +3 -3
- package/dist/glassbox/index.d.ts +3 -3
- package/dist/glassbox-routes/format.d.mts +2 -2
- package/dist/glassbox-routes/format.d.ts +2 -2
- package/dist/glassbox-routes/index.d.mts +4 -4
- package/dist/glassbox-routes/index.d.ts +4 -4
- package/dist/glassbox-routes/index.js +12 -2
- package/dist/glassbox-routes/index.mjs +2 -2
- package/dist/glassbox-routes/react/index.d.mts +2 -2
- package/dist/glassbox-routes/react/index.d.ts +2 -2
- package/dist/index.d.mts +74 -9
- package/dist/index.d.ts +74 -9
- package/dist/index.js +109 -23
- package/dist/index.mjs +6 -6
- package/dist/{ir-DyMJ84je.d.mts → ir-BZ2qOPMp.d.mts} +13 -0
- package/dist/{ir-DbvOFKF-.d.ts → ir-CpkidtAd.d.ts} +13 -0
- package/dist/key-health.js +1 -1
- package/dist/key-health.mjs +1 -1
- package/dist/probe.d.mts +2 -2
- package/dist/probe.d.ts +2 -2
- package/dist/probe.js +106 -20
- package/dist/probe.mjs +3 -3
- package/dist/profiles.d.mts +1 -1
- package/dist/profiles.d.ts +1 -1
- package/dist/profiles.js +12 -2
- package/dist/profiles.mjs +1 -1
- package/dist/{types-D441T-KC.d.mts → types-BkdfoONZ.d.mts} +1 -1
- package/dist/{types-5TqjBeZD.d.ts → types-Bn-zF0S4.d.ts} +1 -1
- package/dist/{types-_myk4bxn.d.mts → types-DJFlO7ux.d.mts} +1 -1
- package/dist/{types-BlrbNQfj.d.ts → types-uWypgaAU.d.ts} +1 -1
- package/package.json +1 -1
|
@@ -9,7 +9,7 @@ import {
|
|
|
9
9
|
getModelCompatibility,
|
|
10
10
|
isBrainQueryActiveFor,
|
|
11
11
|
policySetHas
|
|
12
|
-
} from "./chunk-
|
|
12
|
+
} from "./chunk-2YCTWV4S.mjs";
|
|
13
13
|
import {
|
|
14
14
|
bucketContext,
|
|
15
15
|
bucketHistory,
|
|
@@ -30,7 +30,7 @@ import {
|
|
|
30
30
|
latencyTierOf,
|
|
31
31
|
resolveModelAlias,
|
|
32
32
|
tryGetProfile
|
|
33
|
-
} from "./chunk-
|
|
33
|
+
} from "./chunk-PY4CBZFF.mjs";
|
|
34
34
|
|
|
35
35
|
// src/models-brain.ts
|
|
36
36
|
function isModelRow(x) {
|
|
@@ -1146,12 +1146,19 @@ function mapRows(rows) {
|
|
|
1146
1146
|
if (!isRawFailureRow(row)) continue;
|
|
1147
1147
|
const n = coerceCount(row.n);
|
|
1148
1148
|
const nFail = coerceCount(row.n_fail) ?? 0;
|
|
1149
|
-
|
|
1149
|
+
const nAttempted = coerceCount(row.n_attempted);
|
|
1150
|
+
const nUnavailable = coerceCount(row.n_unavailable) ?? 0;
|
|
1151
|
+
const nUnclassified = coerceCount(row.n_unclassified) ?? 0;
|
|
1152
|
+
if (n === null) continue;
|
|
1153
|
+
if (n <= 0 && nUnavailable <= 0) continue;
|
|
1150
1154
|
out.push({
|
|
1151
1155
|
archetype: row.intent_archetype,
|
|
1152
1156
|
model: row.model,
|
|
1153
1157
|
n,
|
|
1154
|
-
nFail
|
|
1158
|
+
nFail,
|
|
1159
|
+
nAttempted: nAttempted ?? n,
|
|
1160
|
+
nUnavailable,
|
|
1161
|
+
nUnclassified
|
|
1155
1162
|
});
|
|
1156
1163
|
}
|
|
1157
1164
|
return out;
|
|
@@ -1175,7 +1182,24 @@ var MEASURED_FAILURE_CFG = {
|
|
|
1175
1182
|
/** 95% one-sided-ish confidence (standard two-sided z at α=0.05). */
|
|
1176
1183
|
z: 1.96,
|
|
1177
1184
|
/** Must match the view's window. Documented here for the advisory text. */
|
|
1178
|
-
windowDays: 28
|
|
1185
|
+
windowDays: 28,
|
|
1186
|
+
/**
|
|
1187
|
+
* alpha.98 — the AVAILABILITY axis. Same statistic, separate vocabulary,
|
|
1188
|
+
* separate verdict, because a 429 and a parse-failure send an operator in
|
|
1189
|
+
* opposite directions (fund or throttle the account vs. distrust the
|
|
1190
|
+
* model's declared structuredOutput). Merging them into one "failure rate"
|
|
1191
|
+
* is the s75 normalization defect.
|
|
1192
|
+
*
|
|
1193
|
+
* Thresholds match the quality axis deliberately: "we are 95% confident
|
|
1194
|
+
* this model is unreachable more often than not" is the same unarguable
|
|
1195
|
+
* bar, and a de-rank is the same remedy. Kept as a distinct config block
|
|
1196
|
+
* so the two can be tuned apart when evidence says they should be.
|
|
1197
|
+
*/
|
|
1198
|
+
availability: {
|
|
1199
|
+
minSample: 5,
|
|
1200
|
+
lowerBoundThreshold: 0.5,
|
|
1201
|
+
z: 1.96
|
|
1202
|
+
}
|
|
1179
1203
|
};
|
|
1180
1204
|
function wilsonLowerBound(failures, n, z = MEASURED_FAILURE_CFG.z) {
|
|
1181
1205
|
if (n <= 0) return 0;
|
|
@@ -1192,15 +1216,53 @@ function mapMeasuredFailureRows(rows) {
|
|
|
1192
1216
|
}
|
|
1193
1217
|
function judgeMeasuredFailure(row, cfg = MEASURED_FAILURE_CFG) {
|
|
1194
1218
|
if (!row) return void 0;
|
|
1195
|
-
const
|
|
1196
|
-
if (!
|
|
1197
|
-
const
|
|
1219
|
+
const raw = "nFail" in row && typeof row.n === "number" ? row : mapRows([row])[0];
|
|
1220
|
+
if (!raw) return void 0;
|
|
1221
|
+
const normalized = {
|
|
1222
|
+
...raw,
|
|
1223
|
+
nAttempted: typeof raw.nAttempted === "number" ? raw.nAttempted : raw.n,
|
|
1224
|
+
nUnavailable: typeof raw.nUnavailable === "number" ? raw.nUnavailable : 0,
|
|
1225
|
+
nUnclassified: typeof raw.nUnclassified === "number" ? raw.nUnclassified : 0
|
|
1226
|
+
};
|
|
1227
|
+
const quality = judgeAxis(
|
|
1228
|
+
normalized.nFail,
|
|
1229
|
+
normalized.n,
|
|
1230
|
+
cfg.minSample,
|
|
1231
|
+
cfg.lowerBoundThreshold,
|
|
1232
|
+
cfg.z
|
|
1233
|
+
);
|
|
1234
|
+
const availability = judgeAxis(
|
|
1235
|
+
normalized.nUnavailable,
|
|
1236
|
+
normalized.nAttempted,
|
|
1237
|
+
cfg.availability.minSample,
|
|
1238
|
+
cfg.availability.lowerBoundThreshold,
|
|
1239
|
+
cfg.availability.z
|
|
1240
|
+
);
|
|
1241
|
+
if (normalized.n < cfg.minSample && normalized.nAttempted < cfg.availability.minSample) {
|
|
1242
|
+
return void 0;
|
|
1243
|
+
}
|
|
1244
|
+
const axis = quality.gated ? "quality" : availability.gated ? "availability" : void 0;
|
|
1245
|
+
const lead = axis === "availability" ? availability : quality;
|
|
1246
|
+
return {
|
|
1247
|
+
gated: quality.gated || availability.gated,
|
|
1248
|
+
axis,
|
|
1249
|
+
rate: lead.rate,
|
|
1250
|
+
lowerBound: lead.lowerBound,
|
|
1251
|
+
n: lead.n,
|
|
1252
|
+
nFail: lead.nFail,
|
|
1253
|
+
quality,
|
|
1254
|
+
availability,
|
|
1255
|
+
nUnclassified: normalized.nUnclassified
|
|
1256
|
+
};
|
|
1257
|
+
}
|
|
1258
|
+
function judgeAxis(failures, n, minSample, threshold, z) {
|
|
1259
|
+
const lowerBound = wilsonLowerBound(failures, n, z);
|
|
1198
1260
|
return {
|
|
1199
|
-
|
|
1200
|
-
rate: normalized.nFail / normalized.n,
|
|
1261
|
+
rate: n > 0 ? failures / n : 0,
|
|
1201
1262
|
lowerBound,
|
|
1202
|
-
n
|
|
1203
|
-
nFail:
|
|
1263
|
+
n,
|
|
1264
|
+
nFail: failures,
|
|
1265
|
+
gated: n >= minSample && lowerBound > threshold
|
|
1204
1266
|
};
|
|
1205
1267
|
}
|
|
1206
1268
|
var snapshots2 = /* @__PURE__ */ new Map();
|
|
@@ -2242,9 +2304,13 @@ function detectSingleModelArray(ir, policy) {
|
|
|
2242
2304
|
];
|
|
2243
2305
|
}
|
|
2244
2306
|
function suppressedRecommendationReason(ir, archetype, altProfile) {
|
|
2245
|
-
|
|
2246
|
-
|
|
2247
|
-
|
|
2307
|
+
const measured = getMeasuredFailureVerdict({
|
|
2308
|
+
appId: ir.appId,
|
|
2309
|
+
archetype,
|
|
2310
|
+
model: altProfile.id
|
|
2311
|
+
});
|
|
2312
|
+
if (measured?.quality.gated === true) return "measured-failure-gate";
|
|
2313
|
+
if (measured?.availability.gated === true) return "measured-unavailable-gate";
|
|
2248
2314
|
if (ir.constraints?.structuredOutput && effectiveConventions(altProfile).some(
|
|
2249
2315
|
(c) => c.archetype === archetype && c.structuredOutputHint === "avoid"
|
|
2250
2316
|
)) {
|
|
@@ -2755,7 +2821,7 @@ function lowerGoogle(ir, profile, hints) {
|
|
|
2755
2821
|
if (hints.forceTerseOutput) {
|
|
2756
2822
|
generationConfig.maxOutputTokens = 200;
|
|
2757
2823
|
}
|
|
2758
|
-
if (ir.constraints?.structuredOutput && profile.structuredOutput
|
|
2824
|
+
if (ir.constraints?.structuredOutput && profile.structuredOutput !== "none") {
|
|
2759
2825
|
generationConfig.responseMimeType = "application/json";
|
|
2760
2826
|
}
|
|
2761
2827
|
const contents = buildGoogleContents(ir.history ?? [], ir.currentTurn);
|
|
@@ -2814,7 +2880,9 @@ function toGoogleTools(tools) {
|
|
|
2814
2880
|
}
|
|
2815
2881
|
function lowerOpenAI(ir, profile, hints) {
|
|
2816
2882
|
const ordered = sortSections(ir.sections);
|
|
2817
|
-
const
|
|
2883
|
+
const sectionText = ordered.map((s) => s.text).join("\n\n");
|
|
2884
|
+
const jsonInstruction = ir.constraints?.structuredOutput ? "Respond with a single valid JSON value and nothing else. Do not wrap the JSON in a markdown code fence, and do not add any prose before or after it." : "";
|
|
2885
|
+
const systemText = [sectionText, jsonInstruction].filter(Boolean).join("\n\n");
|
|
2818
2886
|
const systemRole = profile.systemPromptMode === "as_developer" ? "developer" : "system";
|
|
2819
2887
|
const messages = systemText ? [{ role: systemRole, content: systemText }] : [];
|
|
2820
2888
|
for (const m of ir.history ?? []) {
|
|
@@ -2860,7 +2928,9 @@ function toOpenAITools(tools) {
|
|
|
2860
2928
|
}
|
|
2861
2929
|
function lowerDeepSeek(ir, profile, hints) {
|
|
2862
2930
|
const ordered = sortSections(ir.sections);
|
|
2863
|
-
const
|
|
2931
|
+
const sectionText = ordered.map((s) => s.text).join("\n\n");
|
|
2932
|
+
const jsonInstruction = ir.constraints?.structuredOutput ? "Respond with a single valid JSON value and nothing else. Do not wrap the JSON in a markdown code fence, and do not add any prose before or after it." : "";
|
|
2933
|
+
const systemText = [sectionText, jsonInstruction].filter(Boolean).join("\n\n");
|
|
2864
2934
|
const messages = systemText ? [{ role: "system", content: systemText }] : [];
|
|
2865
2935
|
for (const m of ir.history ?? []) {
|
|
2866
2936
|
if (m.role === "system") continue;
|
|
@@ -2880,6 +2950,9 @@ function lowerDeepSeek(ir, profile, hints) {
|
|
|
2880
2950
|
provider: "deepseek",
|
|
2881
2951
|
model: profile.id,
|
|
2882
2952
|
messages,
|
|
2953
|
+
// Safe only because `jsonInstruction` above guarantees the literal word
|
|
2954
|
+
// "json" reaches the prompt — DeepSeek 400s otherwise (measurement C).
|
|
2955
|
+
response_format: ir.constraints?.structuredOutput ? { type: "json_object" } : void 0,
|
|
2883
2956
|
// Every tool the IR carries. Never fewer — see the note above.
|
|
2884
2957
|
tools: ir.tools && ir.tools.length > 0 ? ir.tools.map((t) => ({
|
|
2885
2958
|
type: "function",
|
|
@@ -2899,7 +2972,9 @@ function lowerDeepSeek(ir, profile, hints) {
|
|
|
2899
2972
|
}
|
|
2900
2973
|
function buildOpenAICompatibleParts(ir) {
|
|
2901
2974
|
const ordered = sortSections(ir.sections);
|
|
2902
|
-
const
|
|
2975
|
+
const sectionText = ordered.map((s) => s.text).join("\n\n");
|
|
2976
|
+
const jsonInstruction = ir.constraints?.structuredOutput ? "Respond with a single valid JSON value and nothing else. Do not wrap the JSON in a markdown code fence, and do not add any prose before or after it." : "";
|
|
2977
|
+
const systemText = [sectionText, jsonInstruction].filter(Boolean).join("\n\n");
|
|
2903
2978
|
const messages = systemText ? [{ role: "system", content: systemText }] : [];
|
|
2904
2979
|
for (const m of ir.history ?? []) {
|
|
2905
2980
|
if (m.role === "system") continue;
|
|
@@ -3057,7 +3132,8 @@ function compile(ir, opts = {}) {
|
|
|
3057
3132
|
rate: verdict.rate,
|
|
3058
3133
|
lowerBound: verdict.lowerBound,
|
|
3059
3134
|
n: verdict.n,
|
|
3060
|
-
nFail: verdict.nFail
|
|
3135
|
+
nFail: verdict.nFail,
|
|
3136
|
+
axis: verdict.axis ?? "quality"
|
|
3061
3137
|
});
|
|
3062
3138
|
}
|
|
3063
3139
|
}
|
|
@@ -805,7 +805,12 @@ var PROFILES_RAW = [
|
|
|
805
805
|
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
806
806
|
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
807
807
|
parallelToolCalls: true,
|
|
808
|
-
structuredOutput: "
|
|
808
|
+
structuredOutput: "grammar",
|
|
809
|
+
// alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
|
|
810
|
+
// `response_format: {type:'json_schema'}` outright ("This response_format type is
|
|
811
|
+
// unavailable now"), so there is no schema-enforced mode for `'native'` to name.
|
|
812
|
+
// It supports `json_object` only, and only when the prompt contains the literal
|
|
813
|
+
// word "json" — the lowering emits the instruction that guarantees it.
|
|
809
814
|
systemPromptMode: "inline",
|
|
810
815
|
streaming: true,
|
|
811
816
|
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
@@ -899,7 +904,12 @@ var PROFILES_RAW = [
|
|
|
899
904
|
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
900
905
|
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
901
906
|
parallelToolCalls: true,
|
|
902
|
-
structuredOutput: "
|
|
907
|
+
structuredOutput: "grammar",
|
|
908
|
+
// alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
|
|
909
|
+
// `response_format: {type:'json_schema'}` outright ("This response_format type is
|
|
910
|
+
// unavailable now"), so there is no schema-enforced mode for `'native'` to name.
|
|
911
|
+
// It supports `json_object` only, and only when the prompt contains the literal
|
|
912
|
+
// word "json" — the lowering emits the instruction that guarantees it.
|
|
903
913
|
systemPromptMode: "inline",
|
|
904
914
|
streaming: true,
|
|
905
915
|
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-
|
|
3
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-DJFlO7ux.mjs';
|
|
2
|
+
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-DJFlO7ux.mjs';
|
|
3
|
+
import '../ir-BZ2qOPMp.mjs';
|
|
4
4
|
import '../dialect.mjs';
|
|
5
5
|
|
|
6
6
|
/**
|
package/dist/glassbox/index.d.ts
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-
|
|
3
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-uWypgaAU.js';
|
|
2
|
+
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-uWypgaAU.js';
|
|
3
|
+
import '../ir-CpkidtAd.js';
|
|
4
4
|
import '../dialect.js';
|
|
5
5
|
|
|
6
6
|
/**
|
|
@@ -1,7 +1,7 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-
|
|
3
|
-
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-
|
|
4
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-DJFlO7ux.mjs';
|
|
2
|
+
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-BkdfoONZ.mjs';
|
|
3
|
+
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-BkdfoONZ.mjs';
|
|
4
|
+
import '../ir-BZ2qOPMp.mjs';
|
|
5
5
|
import '../dialect.mjs';
|
|
6
6
|
|
|
7
7
|
/**
|
|
@@ -1,7 +1,7 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-
|
|
3
|
-
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-
|
|
4
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-uWypgaAU.js';
|
|
2
|
+
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-Bn-zF0S4.js';
|
|
3
|
+
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-Bn-zF0S4.js';
|
|
4
|
+
import '../ir-CpkidtAd.js';
|
|
5
5
|
import '../dialect.js';
|
|
6
6
|
|
|
7
7
|
/**
|
|
@@ -850,7 +850,12 @@ var PROFILES_RAW = [
|
|
|
850
850
|
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
851
851
|
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
852
852
|
parallelToolCalls: true,
|
|
853
|
-
structuredOutput: "
|
|
853
|
+
structuredOutput: "grammar",
|
|
854
|
+
// alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
|
|
855
|
+
// `response_format: {type:'json_schema'}` outright ("This response_format type is
|
|
856
|
+
// unavailable now"), so there is no schema-enforced mode for `'native'` to name.
|
|
857
|
+
// It supports `json_object` only, and only when the prompt contains the literal
|
|
858
|
+
// word "json" — the lowering emits the instruction that guarantees it.
|
|
854
859
|
systemPromptMode: "inline",
|
|
855
860
|
streaming: true,
|
|
856
861
|
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
@@ -944,7 +949,12 @@ var PROFILES_RAW = [
|
|
|
944
949
|
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
945
950
|
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
946
951
|
parallelToolCalls: true,
|
|
947
|
-
structuredOutput: "
|
|
952
|
+
structuredOutput: "grammar",
|
|
953
|
+
// alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
|
|
954
|
+
// `response_format: {type:'json_schema'}` outright ("This response_format type is
|
|
955
|
+
// unavailable now"), so there is no schema-enforced mode for `'native'` to name.
|
|
956
|
+
// It supports `json_object` only, and only when the prompt contains the literal
|
|
957
|
+
// word "json" — the lowering emits the instruction that guarantees it.
|
|
948
958
|
systemPromptMode: "inline",
|
|
949
959
|
streaming: true,
|
|
950
960
|
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import {
|
|
2
2
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
3
3
|
getDefaultFallbackChain
|
|
4
|
-
} from "../chunk-
|
|
4
|
+
} from "../chunk-2YCTWV4S.mjs";
|
|
5
5
|
import {
|
|
6
6
|
tryGetProfile
|
|
7
|
-
} from "../chunk-
|
|
7
|
+
} from "../chunk-PY4CBZFF.mjs";
|
|
8
8
|
import {
|
|
9
9
|
subscribe,
|
|
10
10
|
subscribeApp
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import * as react_jsx_runtime from 'react/jsx-runtime';
|
|
2
|
-
import { a as TraceDetail } from '../../types-
|
|
3
|
-
import '../../ir-
|
|
2
|
+
import { a as TraceDetail } from '../../types-BkdfoONZ.mjs';
|
|
3
|
+
import '../../ir-BZ2qOPMp.mjs';
|
|
4
4
|
import '../../dialect.mjs';
|
|
5
5
|
|
|
6
6
|
/**
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import * as react_jsx_runtime from 'react/jsx-runtime';
|
|
2
|
-
import { a as TraceDetail } from '../../types-
|
|
3
|
-
import '../../ir-
|
|
2
|
+
import { a as TraceDetail } from '../../types-Bn-zF0S4.js';
|
|
3
|
+
import '../../ir-CpkidtAd.js';
|
|
4
4
|
import '../../dialect.js';
|
|
5
5
|
|
|
6
6
|
/**
|
package/dist/index.d.mts
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { C as CompileOptions } from './compile-
|
|
2
|
-
import { N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, C as CompiledRequest, a as CompilePolicy, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, F as FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-
|
|
3
|
-
export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-
|
|
1
|
+
import { C as CompileOptions } from './compile-hh_l7COL.mjs';
|
|
2
|
+
import { N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, C as CompiledRequest, a as CompilePolicy, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, F as FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-BZ2qOPMp.mjs';
|
|
3
|
+
export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-BZ2qOPMp.mjs';
|
|
4
4
|
export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.mjs';
|
|
5
5
|
export { KEY_FINGERPRINT_DOMAIN, KEY_FINGERPRINT_LENGTH, KeyHealthConfig, KeyHealthProvider, KeyHealthResponseBody, KeyHealthResult, KeyHealthRoute, createKeyHealthRoute, keyFingerprint } from './key-health.mjs';
|
|
6
6
|
import { ModelProfile, ArchetypeConvention, ProbedLimit } from './profiles.mjs';
|
|
@@ -1241,7 +1241,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
|
|
|
1241
1241
|
* guard in `tests/version.test.ts` fails the suite (and therefore
|
|
1242
1242
|
* `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
|
|
1243
1243
|
*/
|
|
1244
|
-
declare const LIBRARY_VERSION = "2.0.0-alpha.
|
|
1244
|
+
declare const LIBRARY_VERSION = "2.0.0-alpha.99";
|
|
1245
1245
|
|
|
1246
1246
|
/**
|
|
1247
1247
|
* Oracle contract — how an app tells the brain whether a response was good.
|
|
@@ -3735,10 +3735,26 @@ declare function _testWaitForPromotionsRefresh(): Promise<void>;
|
|
|
3735
3735
|
interface SurfaceFailureRow {
|
|
3736
3736
|
archetype: string;
|
|
3737
3737
|
model: string;
|
|
3738
|
-
/**
|
|
3738
|
+
/**
|
|
3739
|
+
* QUALITY-axis denominator: attempts that produced something judgeable
|
|
3740
|
+
* (served + quality walkaways). Excludes calls that never reached an
|
|
3741
|
+
* answer — a 429 is not a trial of the model's quality. See migration 071.
|
|
3742
|
+
*/
|
|
3739
3743
|
n: number;
|
|
3740
3744
|
/** Attempts that failed on the quality axis. */
|
|
3741
3745
|
nFail: number;
|
|
3746
|
+
/**
|
|
3747
|
+
* alpha.98 — AVAILABILITY-axis denominator: every served attempt is a
|
|
3748
|
+
* trial of whether the model can be reached at all.
|
|
3749
|
+
*/
|
|
3750
|
+
nAttempted: number;
|
|
3751
|
+
/** Attempts that never reached an answer — 429, auth, timeout, transport. */
|
|
3752
|
+
nUnavailable: number;
|
|
3753
|
+
/**
|
|
3754
|
+
* Failures in NEITHER vocabulary. Carried so the taxonomy's blind spot is
|
|
3755
|
+
* a number rather than silence; not gated on.
|
|
3756
|
+
*/
|
|
3757
|
+
nUnclassified: number;
|
|
3742
3758
|
}
|
|
3743
3759
|
declare const MEASURED_FAILURE_CFG: {
|
|
3744
3760
|
/**
|
|
@@ -3760,6 +3776,23 @@ declare const MEASURED_FAILURE_CFG: {
|
|
|
3760
3776
|
readonly z: 1.96;
|
|
3761
3777
|
/** Must match the view's window. Documented here for the advisory text. */
|
|
3762
3778
|
readonly windowDays: 28;
|
|
3779
|
+
/**
|
|
3780
|
+
* alpha.98 — the AVAILABILITY axis. Same statistic, separate vocabulary,
|
|
3781
|
+
* separate verdict, because a 429 and a parse-failure send an operator in
|
|
3782
|
+
* opposite directions (fund or throttle the account vs. distrust the
|
|
3783
|
+
* model's declared structuredOutput). Merging them into one "failure rate"
|
|
3784
|
+
* is the s75 normalization defect.
|
|
3785
|
+
*
|
|
3786
|
+
* Thresholds match the quality axis deliberately: "we are 95% confident
|
|
3787
|
+
* this model is unreachable more often than not" is the same unarguable
|
|
3788
|
+
* bar, and a de-rank is the same remedy. Kept as a distinct config block
|
|
3789
|
+
* so the two can be tuned apart when evidence says they should be.
|
|
3790
|
+
*/
|
|
3791
|
+
readonly availability: {
|
|
3792
|
+
readonly minSample: 5;
|
|
3793
|
+
readonly lowerBoundThreshold: 0.5;
|
|
3794
|
+
readonly z: 1.96;
|
|
3795
|
+
};
|
|
3763
3796
|
};
|
|
3764
3797
|
/**
|
|
3765
3798
|
* Wilson score interval, lower bound. Preferred over the normal
|
|
@@ -3771,16 +3804,45 @@ declare const MEASURED_FAILURE_CFG: {
|
|
|
3771
3804
|
* Returns 0 for n <= 0.
|
|
3772
3805
|
*/
|
|
3773
3806
|
declare function wilsonLowerBound(failures: number, n: number, z?: number): number;
|
|
3807
|
+
/** One axis's numbers. Same statistic, different vocabulary. */
|
|
3808
|
+
interface FailureAxisVerdict {
|
|
3809
|
+
/** Observed failure rate in-window on this axis. */
|
|
3810
|
+
rate: number;
|
|
3811
|
+
/** 95% lower confidence bound on that rate — what the gate tests. */
|
|
3812
|
+
lowerBound: number;
|
|
3813
|
+
/** Attempts backing this axis. */
|
|
3814
|
+
n: number;
|
|
3815
|
+
/** Failures on this axis. */
|
|
3816
|
+
nFail: number;
|
|
3817
|
+
/** Whether this axis on its own clears the bar. */
|
|
3818
|
+
gated: boolean;
|
|
3819
|
+
}
|
|
3774
3820
|
interface MeasuredFailureVerdict {
|
|
3775
|
-
/** Whether
|
|
3821
|
+
/** Whether a measured-failure gate fires for this tuple, on EITHER axis. */
|
|
3776
3822
|
gated: boolean;
|
|
3777
|
-
/**
|
|
3823
|
+
/**
|
|
3824
|
+
* alpha.98 — which axis fired. Undefined when `gated` is false. Read this
|
|
3825
|
+
* before rendering `rate`/`n`: the same number means "answered badly" on
|
|
3826
|
+
* one axis and "could not be reached" on the other, and the fixes differ.
|
|
3827
|
+
* When both axes fire, `'quality'` wins the label — the model is reachable
|
|
3828
|
+
* enough to have been judged, so its answers are the actionable half.
|
|
3829
|
+
*/
|
|
3830
|
+
axis?: 'quality' | 'availability';
|
|
3831
|
+
/**
|
|
3832
|
+
* The numbers that JUSTIFY the gate — i.e. drawn from `axis` when gated,
|
|
3833
|
+
* and from the quality axis otherwise. Deliberately not always-quality:
|
|
3834
|
+
* a consumer printing "gated: 0% over 16" for a model that 429'd 42 times
|
|
3835
|
+
* would be reporting a true number as an explanation of the wrong thing.
|
|
3836
|
+
*/
|
|
3778
3837
|
rate: number;
|
|
3779
|
-
/** 95% lower confidence bound on that rate — what the gate tests. */
|
|
3780
3838
|
lowerBound: number;
|
|
3781
|
-
/** Attempts backing the verdict. */
|
|
3782
3839
|
n: number;
|
|
3783
3840
|
nFail: number;
|
|
3841
|
+
/** Both axes, always present, so a caller never has to infer one. */
|
|
3842
|
+
quality: FailureAxisVerdict;
|
|
3843
|
+
availability: FailureAxisVerdict;
|
|
3844
|
+
/** Failures in neither vocabulary. Not gated on; surfaced for visibility. */
|
|
3845
|
+
nUnclassified: number;
|
|
3784
3846
|
}
|
|
3785
3847
|
/**
|
|
3786
3848
|
* alpha.78 — public row normalizer (PB 2026-07-25 trap): the counts
|
|
@@ -3805,6 +3867,9 @@ declare function judgeMeasuredFailure(row: SurfaceFailureRow | {
|
|
|
3805
3867
|
model: string;
|
|
3806
3868
|
n: number | string;
|
|
3807
3869
|
n_fail: number | string;
|
|
3870
|
+
n_attempted?: number | string;
|
|
3871
|
+
n_unavailable?: number | string;
|
|
3872
|
+
n_unclassified?: number | string;
|
|
3808
3873
|
} | undefined, cfg?: typeof MEASURED_FAILURE_CFG): MeasuredFailureVerdict | undefined;
|
|
3809
3874
|
interface MeasuredFailureRuntime {
|
|
3810
3875
|
/** Endpoint base URL. The library appends `?app_id=<id>`. */
|