@warmdrift/kgauto-compiler 2.0.0-alpha.94 → 2.0.0-alpha.95
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-OAJZLZBN.mjs → chunk-BJD75WCL.mjs} +1 -1
- package/dist/{chunk-57XCZGN5.mjs → chunk-D3BVYXNA.mjs} +100 -0
- package/dist/{chunk-TN6UIX56.mjs → chunk-IRXC7WEQ.mjs} +13 -4
- package/dist/{chunk-VXSSZFES.mjs → chunk-JFM5A556.mjs} +1 -1
- package/dist/glassbox-routes/index.js +100 -0
- package/dist/glassbox-routes/index.mjs +2 -2
- package/dist/index.d.mts +12 -3
- package/dist/index.d.ts +12 -3
- package/dist/index.js +114 -3
- package/dist/index.mjs +5 -4
- package/dist/key-health.js +1 -1
- package/dist/key-health.mjs +1 -1
- package/dist/probe.js +105 -1
- package/dist/probe.mjs +3 -3
- package/dist/profiles.d.mts +41 -1
- package/dist/profiles.d.ts +41 -1
- package/dist/profiles.js +100 -0
- package/dist/profiles.mjs +1 -1
- package/package.json +2 -2
|
@@ -54,6 +54,26 @@ var PROFILES_RAW = [
|
|
|
54
54
|
// on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
|
|
55
55
|
// `thinking` field, so kgauto never sends the param. (L-049/L-081.)
|
|
56
56
|
id: "claude-fable-5",
|
|
57
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 32.
|
|
58
|
+
// This model declared no cliffs, which is why the offline check had
|
|
59
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
60
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
61
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
62
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
63
|
+
probedLimits: [
|
|
64
|
+
{
|
|
65
|
+
metric: "tool_count",
|
|
66
|
+
testedAt: 20,
|
|
67
|
+
demanded: 3,
|
|
68
|
+
served: [3, 3, 3],
|
|
69
|
+
outcome: "held",
|
|
70
|
+
trials: 3,
|
|
71
|
+
lastProbed: "2026-08-28",
|
|
72
|
+
probeRunId: 32,
|
|
73
|
+
route: "direct",
|
|
74
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
75
|
+
}
|
|
76
|
+
],
|
|
57
77
|
verifiedAgainstDocs: "2026-06-21",
|
|
58
78
|
provider: "anthropic",
|
|
59
79
|
status: "current",
|
|
@@ -174,6 +194,26 @@ var PROFILES_RAW = [
|
|
|
174
194
|
// failure, where a template-clone of PRICE (the field that actually varies,
|
|
175
195
|
// there by 2.5-3.75x) went unverified for 65 days.
|
|
176
196
|
id: "claude-opus-5",
|
|
197
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 33.
|
|
198
|
+
// This model declared no cliffs, which is why the offline check had
|
|
199
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
200
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
201
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
202
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
203
|
+
probedLimits: [
|
|
204
|
+
{
|
|
205
|
+
metric: "tool_count",
|
|
206
|
+
testedAt: 20,
|
|
207
|
+
demanded: 3,
|
|
208
|
+
served: [3, 3, 3],
|
|
209
|
+
outcome: "held",
|
|
210
|
+
trials: 3,
|
|
211
|
+
lastProbed: "2026-08-28",
|
|
212
|
+
probeRunId: 33,
|
|
213
|
+
route: "direct",
|
|
214
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
215
|
+
}
|
|
216
|
+
],
|
|
177
217
|
verifiedAgainstDocs: "2026-07-25",
|
|
178
218
|
provider: "anthropic",
|
|
179
219
|
status: "current",
|
|
@@ -1052,6 +1092,26 @@ var PROFILES_RAW = [
|
|
|
1052
1092
|
},
|
|
1053
1093
|
{
|
|
1054
1094
|
id: "gpt-5.4-mini",
|
|
1095
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 34.
|
|
1096
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1097
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1098
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1099
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1100
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1101
|
+
probedLimits: [
|
|
1102
|
+
{
|
|
1103
|
+
metric: "tool_count",
|
|
1104
|
+
testedAt: 20,
|
|
1105
|
+
demanded: 3,
|
|
1106
|
+
served: [3, 3, 3],
|
|
1107
|
+
outcome: "held",
|
|
1108
|
+
trials: 3,
|
|
1109
|
+
lastProbed: "2026-08-28",
|
|
1110
|
+
probeRunId: 34,
|
|
1111
|
+
route: "direct",
|
|
1112
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1113
|
+
}
|
|
1114
|
+
],
|
|
1055
1115
|
verifiedAgainstDocs: "2026-05-17",
|
|
1056
1116
|
provider: "openai",
|
|
1057
1117
|
status: "current",
|
|
@@ -1095,6 +1155,26 @@ var PROFILES_RAW = [
|
|
|
1095
1155
|
},
|
|
1096
1156
|
{
|
|
1097
1157
|
id: "gpt-5.4-nano",
|
|
1158
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 35.
|
|
1159
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1160
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1161
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1162
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1163
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1164
|
+
probedLimits: [
|
|
1165
|
+
{
|
|
1166
|
+
metric: "tool_count",
|
|
1167
|
+
testedAt: 20,
|
|
1168
|
+
demanded: 3,
|
|
1169
|
+
served: [3, 3, 3],
|
|
1170
|
+
outcome: "held",
|
|
1171
|
+
trials: 3,
|
|
1172
|
+
lastProbed: "2026-08-28",
|
|
1173
|
+
probeRunId: 35,
|
|
1174
|
+
route: "direct",
|
|
1175
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1176
|
+
}
|
|
1177
|
+
],
|
|
1098
1178
|
verifiedAgainstDocs: "2026-05-17",
|
|
1099
1179
|
provider: "openai",
|
|
1100
1180
|
status: "current",
|
|
@@ -1710,6 +1790,26 @@ var PROFILES_RAW = [
|
|
|
1710
1790
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1711
1791
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1712
1792
|
id: "claude-sonnet-5",
|
|
1793
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 36.
|
|
1794
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1795
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1796
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1797
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1798
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1799
|
+
probedLimits: [
|
|
1800
|
+
{
|
|
1801
|
+
metric: "tool_count",
|
|
1802
|
+
testedAt: 20,
|
|
1803
|
+
demanded: 3,
|
|
1804
|
+
served: [3, 3, 3],
|
|
1805
|
+
outcome: "held",
|
|
1806
|
+
trials: 3,
|
|
1807
|
+
lastProbed: "2026-08-28",
|
|
1808
|
+
probeRunId: 36,
|
|
1809
|
+
route: "direct",
|
|
1810
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1811
|
+
}
|
|
1812
|
+
],
|
|
1713
1813
|
verifiedAgainstDocs: "2026-08-14",
|
|
1714
1814
|
provider: "anthropic",
|
|
1715
1815
|
status: "current",
|
|
@@ -9,7 +9,7 @@ import {
|
|
|
9
9
|
getModelCompatibility,
|
|
10
10
|
isBrainQueryActiveFor,
|
|
11
11
|
policySetHas
|
|
12
|
-
} from "./chunk-
|
|
12
|
+
} from "./chunk-BJD75WCL.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-D3BVYXNA.mjs";
|
|
34
34
|
|
|
35
35
|
// src/models-brain.ts
|
|
36
36
|
function isModelRow(x) {
|
|
@@ -88,7 +88,11 @@ function rowToProfile(row) {
|
|
|
88
88
|
// latency_tier → undefined (latencyTierOf derives from tags; not
|
|
89
89
|
// safety-critical). archetype_conventions already array-validated above.
|
|
90
90
|
latencyTier: normalizeLatencyTier(row.latency_tier),
|
|
91
|
-
archetypeConventions: row.archetype_conventions ?? void 0
|
|
91
|
+
archetypeConventions: row.archetype_conventions ?? void 0,
|
|
92
|
+
// alpha.95 — measured negatives. NOT safety-critical (nothing routes on
|
|
93
|
+
// them), so a type-mismatch degrades to undefined rather than nulling
|
|
94
|
+
// the whole profile the way a corrupt `cliffs` does.
|
|
95
|
+
probedLimits: Array.isArray(row.probed_limits) ? row.probed_limits : void 0
|
|
92
96
|
};
|
|
93
97
|
} catch {
|
|
94
98
|
return null;
|
|
@@ -128,7 +132,12 @@ function profileToRow(profile, opts = {}) {
|
|
|
128
132
|
// so a reseed from bundled profiles makes the latency lever + schema
|
|
129
133
|
// conventions live warm. Closes the silently-dropped-field gap.
|
|
130
134
|
latency_tier: profile.latencyTier ?? null,
|
|
131
|
-
archetype_conventions: profile.archetypeConventions ?? null
|
|
135
|
+
archetype_conventions: profile.archetypeConventions ?? null,
|
|
136
|
+
// alpha.95 — round-trip the measured negatives. The s56 rule: a new
|
|
137
|
+
// ModelProfile field must land in migration + BOTH transformers + the
|
|
138
|
+
// config-endpoint cols list + the parity test, or it is theater for every
|
|
139
|
+
// brain-connected consumer.
|
|
140
|
+
probed_limits: profile.probedLimits ?? null
|
|
132
141
|
};
|
|
133
142
|
if (opts.verifiedAgainstDocs !== void 0) {
|
|
134
143
|
row.verified_against_docs = opts.verifiedAgainstDocs;
|
|
@@ -99,6 +99,26 @@ var PROFILES_RAW = [
|
|
|
99
99
|
// on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
|
|
100
100
|
// `thinking` field, so kgauto never sends the param. (L-049/L-081.)
|
|
101
101
|
id: "claude-fable-5",
|
|
102
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 32.
|
|
103
|
+
// This model declared no cliffs, which is why the offline check had
|
|
104
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
105
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
106
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
107
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
108
|
+
probedLimits: [
|
|
109
|
+
{
|
|
110
|
+
metric: "tool_count",
|
|
111
|
+
testedAt: 20,
|
|
112
|
+
demanded: 3,
|
|
113
|
+
served: [3, 3, 3],
|
|
114
|
+
outcome: "held",
|
|
115
|
+
trials: 3,
|
|
116
|
+
lastProbed: "2026-08-28",
|
|
117
|
+
probeRunId: 32,
|
|
118
|
+
route: "direct",
|
|
119
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
120
|
+
}
|
|
121
|
+
],
|
|
102
122
|
verifiedAgainstDocs: "2026-06-21",
|
|
103
123
|
provider: "anthropic",
|
|
104
124
|
status: "current",
|
|
@@ -219,6 +239,26 @@ var PROFILES_RAW = [
|
|
|
219
239
|
// failure, where a template-clone of PRICE (the field that actually varies,
|
|
220
240
|
// there by 2.5-3.75x) went unverified for 65 days.
|
|
221
241
|
id: "claude-opus-5",
|
|
242
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 33.
|
|
243
|
+
// This model declared no cliffs, which is why the offline check had
|
|
244
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
245
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
246
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
247
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
248
|
+
probedLimits: [
|
|
249
|
+
{
|
|
250
|
+
metric: "tool_count",
|
|
251
|
+
testedAt: 20,
|
|
252
|
+
demanded: 3,
|
|
253
|
+
served: [3, 3, 3],
|
|
254
|
+
outcome: "held",
|
|
255
|
+
trials: 3,
|
|
256
|
+
lastProbed: "2026-08-28",
|
|
257
|
+
probeRunId: 33,
|
|
258
|
+
route: "direct",
|
|
259
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
260
|
+
}
|
|
261
|
+
],
|
|
222
262
|
verifiedAgainstDocs: "2026-07-25",
|
|
223
263
|
provider: "anthropic",
|
|
224
264
|
status: "current",
|
|
@@ -1097,6 +1137,26 @@ var PROFILES_RAW = [
|
|
|
1097
1137
|
},
|
|
1098
1138
|
{
|
|
1099
1139
|
id: "gpt-5.4-mini",
|
|
1140
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 34.
|
|
1141
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1142
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1143
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1144
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1145
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1146
|
+
probedLimits: [
|
|
1147
|
+
{
|
|
1148
|
+
metric: "tool_count",
|
|
1149
|
+
testedAt: 20,
|
|
1150
|
+
demanded: 3,
|
|
1151
|
+
served: [3, 3, 3],
|
|
1152
|
+
outcome: "held",
|
|
1153
|
+
trials: 3,
|
|
1154
|
+
lastProbed: "2026-08-28",
|
|
1155
|
+
probeRunId: 34,
|
|
1156
|
+
route: "direct",
|
|
1157
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1158
|
+
}
|
|
1159
|
+
],
|
|
1100
1160
|
verifiedAgainstDocs: "2026-05-17",
|
|
1101
1161
|
provider: "openai",
|
|
1102
1162
|
status: "current",
|
|
@@ -1140,6 +1200,26 @@ var PROFILES_RAW = [
|
|
|
1140
1200
|
},
|
|
1141
1201
|
{
|
|
1142
1202
|
id: "gpt-5.4-nano",
|
|
1203
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 35.
|
|
1204
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1205
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1206
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1207
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1208
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1209
|
+
probedLimits: [
|
|
1210
|
+
{
|
|
1211
|
+
metric: "tool_count",
|
|
1212
|
+
testedAt: 20,
|
|
1213
|
+
demanded: 3,
|
|
1214
|
+
served: [3, 3, 3],
|
|
1215
|
+
outcome: "held",
|
|
1216
|
+
trials: 3,
|
|
1217
|
+
lastProbed: "2026-08-28",
|
|
1218
|
+
probeRunId: 35,
|
|
1219
|
+
route: "direct",
|
|
1220
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1221
|
+
}
|
|
1222
|
+
],
|
|
1143
1223
|
verifiedAgainstDocs: "2026-05-17",
|
|
1144
1224
|
provider: "openai",
|
|
1145
1225
|
status: "current",
|
|
@@ -1755,6 +1835,26 @@ var PROFILES_RAW = [
|
|
|
1755
1835
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1756
1836
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1757
1837
|
id: "claude-sonnet-5",
|
|
1838
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 36.
|
|
1839
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1840
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1841
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1842
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1843
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1844
|
+
probedLimits: [
|
|
1845
|
+
{
|
|
1846
|
+
metric: "tool_count",
|
|
1847
|
+
testedAt: 20,
|
|
1848
|
+
demanded: 3,
|
|
1849
|
+
served: [3, 3, 3],
|
|
1850
|
+
outcome: "held",
|
|
1851
|
+
trials: 3,
|
|
1852
|
+
lastProbed: "2026-08-28",
|
|
1853
|
+
probeRunId: 36,
|
|
1854
|
+
route: "direct",
|
|
1855
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1856
|
+
}
|
|
1857
|
+
],
|
|
1758
1858
|
verifiedAgainstDocs: "2026-08-14",
|
|
1759
1859
|
provider: "anthropic",
|
|
1760
1860
|
status: "current",
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import {
|
|
2
2
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
3
3
|
getDefaultFallbackChain
|
|
4
|
-
} from "../chunk-
|
|
4
|
+
} from "../chunk-BJD75WCL.mjs";
|
|
5
5
|
import {
|
|
6
6
|
tryGetProfile
|
|
7
|
-
} from "../chunk-
|
|
7
|
+
} from "../chunk-D3BVYXNA.mjs";
|
|
8
8
|
import {
|
|
9
9
|
subscribe,
|
|
10
10
|
subscribeApp
|
package/dist/index.d.mts
CHANGED
|
@@ -3,7 +3,7 @@ import { N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, C as Com
|
|
|
3
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-DyMJ84je.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
|
-
import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
|
|
6
|
+
import { ModelProfile, ArchetypeConvention, ProbedLimit } from './profiles.mjs';
|
|
7
7
|
export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, _testClearConsumerProfiles, allProfiles, bestEffortProfile, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile } from './profiles.mjs';
|
|
8
8
|
import { IntentArchetypeName, OutputMode } from './dialect.mjs';
|
|
9
9
|
export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.mjs';
|
|
@@ -76,6 +76,13 @@ interface ExecuteErr {
|
|
|
76
76
|
}
|
|
77
77
|
type ExecuteResult = ExecuteOk | ExecuteErr;
|
|
78
78
|
declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promise<ExecuteResult>;
|
|
79
|
+
/**
|
|
80
|
+
* Provider wordings that mean "this model will never serve THIS caller" —
|
|
81
|
+
* delistings and retirements that arrive on statuses other than 404. Kept
|
|
82
|
+
* deliberately narrow: a match reroutes the call to the next chain target,
|
|
83
|
+
* so a false positive here would mask a real validation error.
|
|
84
|
+
*/
|
|
85
|
+
declare function isModelUnavailableMessage(message: string | undefined): boolean;
|
|
79
86
|
|
|
80
87
|
/**
|
|
81
88
|
* call() — the high-level orchestrator that owns the network round-trip.
|
|
@@ -1234,7 +1241,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
|
|
|
1234
1241
|
* guard in `tests/version.test.ts` fails the suite (and therefore
|
|
1235
1242
|
* `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
|
|
1236
1243
|
*/
|
|
1237
|
-
declare const LIBRARY_VERSION = "2.0.0-alpha.
|
|
1244
|
+
declare const LIBRARY_VERSION = "2.0.0-alpha.95";
|
|
1238
1245
|
|
|
1239
1246
|
/**
|
|
1240
1247
|
* Oracle contract — how an app tells the brain whether a response was good.
|
|
@@ -3219,6 +3226,8 @@ interface ModelBrainRow {
|
|
|
3219
3226
|
latency_tier?: string | null;
|
|
3220
3227
|
/** alpha.49 — per-archetype prompt-shape conventions (migration 028 column). */
|
|
3221
3228
|
archetype_conventions?: ArchetypeConvention[] | null;
|
|
3229
|
+
/** alpha.95 — measured negatives from --discover (migration 070 column). */
|
|
3230
|
+
probed_limits?: ProbedLimit[] | null;
|
|
3222
3231
|
}
|
|
3223
3232
|
interface ProfileToRowOptions {
|
|
3224
3233
|
/** e.g. `'2.0.0-alpha.12'` — leave undefined to omit from row. */
|
|
@@ -4055,4 +4064,4 @@ declare function planDecomposition(args: PlanDecompositionArgs): DecompositionPl
|
|
|
4055
4064
|
*/
|
|
4056
4065
|
declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
|
|
4057
4066
|
|
|
4058
|
-
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_JUDGE_MODEL, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
|
4067
|
+
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_JUDGE_MODEL, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, ProbedLimit, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isModelUnavailableMessage, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
package/dist/index.d.ts
CHANGED
|
@@ -3,7 +3,7 @@ import { N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, C as Com
|
|
|
3
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-DbvOFKF-.js';
|
|
4
4
|
export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.js';
|
|
5
5
|
export { KEY_FINGERPRINT_DOMAIN, KEY_FINGERPRINT_LENGTH, KeyHealthConfig, KeyHealthProvider, KeyHealthResponseBody, KeyHealthResult, KeyHealthRoute, createKeyHealthRoute, keyFingerprint } from './key-health.js';
|
|
6
|
-
import { ModelProfile, ArchetypeConvention } from './profiles.js';
|
|
6
|
+
import { ModelProfile, ArchetypeConvention, ProbedLimit } from './profiles.js';
|
|
7
7
|
export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, _testClearConsumerProfiles, allProfiles, bestEffortProfile, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile } from './profiles.js';
|
|
8
8
|
import { IntentArchetypeName, OutputMode } from './dialect.js';
|
|
9
9
|
export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.js';
|
|
@@ -76,6 +76,13 @@ interface ExecuteErr {
|
|
|
76
76
|
}
|
|
77
77
|
type ExecuteResult = ExecuteOk | ExecuteErr;
|
|
78
78
|
declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promise<ExecuteResult>;
|
|
79
|
+
/**
|
|
80
|
+
* Provider wordings that mean "this model will never serve THIS caller" —
|
|
81
|
+
* delistings and retirements that arrive on statuses other than 404. Kept
|
|
82
|
+
* deliberately narrow: a match reroutes the call to the next chain target,
|
|
83
|
+
* so a false positive here would mask a real validation error.
|
|
84
|
+
*/
|
|
85
|
+
declare function isModelUnavailableMessage(message: string | undefined): boolean;
|
|
79
86
|
|
|
80
87
|
/**
|
|
81
88
|
* call() — the high-level orchestrator that owns the network round-trip.
|
|
@@ -1234,7 +1241,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
|
|
|
1234
1241
|
* guard in `tests/version.test.ts` fails the suite (and therefore
|
|
1235
1242
|
* `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
|
|
1236
1243
|
*/
|
|
1237
|
-
declare const LIBRARY_VERSION = "2.0.0-alpha.
|
|
1244
|
+
declare const LIBRARY_VERSION = "2.0.0-alpha.95";
|
|
1238
1245
|
|
|
1239
1246
|
/**
|
|
1240
1247
|
* Oracle contract — how an app tells the brain whether a response was good.
|
|
@@ -3219,6 +3226,8 @@ interface ModelBrainRow {
|
|
|
3219
3226
|
latency_tier?: string | null;
|
|
3220
3227
|
/** alpha.49 — per-archetype prompt-shape conventions (migration 028 column). */
|
|
3221
3228
|
archetype_conventions?: ArchetypeConvention[] | null;
|
|
3229
|
+
/** alpha.95 — measured negatives from --discover (migration 070 column). */
|
|
3230
|
+
probed_limits?: ProbedLimit[] | null;
|
|
3222
3231
|
}
|
|
3223
3232
|
interface ProfileToRowOptions {
|
|
3224
3233
|
/** e.g. `'2.0.0-alpha.12'` — leave undefined to omit from row. */
|
|
@@ -4055,4 +4064,4 @@ declare function planDecomposition(args: PlanDecompositionArgs): DecompositionPl
|
|
|
4055
4064
|
*/
|
|
4056
4065
|
declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
|
|
4057
4066
|
|
|
4058
|
-
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_JUDGE_MODEL, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
|
4067
|
+
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_JUDGE_MODEL, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, ProbedLimit, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isModelUnavailableMessage, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
package/dist/index.js
CHANGED
|
@@ -142,6 +142,7 @@ __export(index_exports, {
|
|
|
142
142
|
isMeasuredFailureBrainActive: () => isMeasuredFailureBrainActive,
|
|
143
143
|
isMeasuredFailureGateEnabledFromEnv: () => isMeasuredFailureGateEnabledFromEnv,
|
|
144
144
|
isModelReachable: () => isModelReachable,
|
|
145
|
+
isModelUnavailableMessage: () => isModelUnavailableMessage,
|
|
145
146
|
isPromotionsBrainActive: () => isPromotionsBrainActive,
|
|
146
147
|
isProviderReachable: () => isProviderReachable,
|
|
147
148
|
judgeMeasuredFailure: () => judgeMeasuredFailure,
|
|
@@ -418,6 +419,26 @@ var PROFILES_RAW = [
|
|
|
418
419
|
// on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
|
|
419
420
|
// `thinking` field, so kgauto never sends the param. (L-049/L-081.)
|
|
420
421
|
id: "claude-fable-5",
|
|
422
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 32.
|
|
423
|
+
// This model declared no cliffs, which is why the offline check had
|
|
424
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
425
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
426
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
427
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
428
|
+
probedLimits: [
|
|
429
|
+
{
|
|
430
|
+
metric: "tool_count",
|
|
431
|
+
testedAt: 20,
|
|
432
|
+
demanded: 3,
|
|
433
|
+
served: [3, 3, 3],
|
|
434
|
+
outcome: "held",
|
|
435
|
+
trials: 3,
|
|
436
|
+
lastProbed: "2026-08-28",
|
|
437
|
+
probeRunId: 32,
|
|
438
|
+
route: "direct",
|
|
439
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
440
|
+
}
|
|
441
|
+
],
|
|
421
442
|
verifiedAgainstDocs: "2026-06-21",
|
|
422
443
|
provider: "anthropic",
|
|
423
444
|
status: "current",
|
|
@@ -538,6 +559,26 @@ var PROFILES_RAW = [
|
|
|
538
559
|
// failure, where a template-clone of PRICE (the field that actually varies,
|
|
539
560
|
// there by 2.5-3.75x) went unverified for 65 days.
|
|
540
561
|
id: "claude-opus-5",
|
|
562
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 33.
|
|
563
|
+
// This model declared no cliffs, which is why the offline check had
|
|
564
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
565
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
566
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
567
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
568
|
+
probedLimits: [
|
|
569
|
+
{
|
|
570
|
+
metric: "tool_count",
|
|
571
|
+
testedAt: 20,
|
|
572
|
+
demanded: 3,
|
|
573
|
+
served: [3, 3, 3],
|
|
574
|
+
outcome: "held",
|
|
575
|
+
trials: 3,
|
|
576
|
+
lastProbed: "2026-08-28",
|
|
577
|
+
probeRunId: 33,
|
|
578
|
+
route: "direct",
|
|
579
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
580
|
+
}
|
|
581
|
+
],
|
|
541
582
|
verifiedAgainstDocs: "2026-07-25",
|
|
542
583
|
provider: "anthropic",
|
|
543
584
|
status: "current",
|
|
@@ -1416,6 +1457,26 @@ var PROFILES_RAW = [
|
|
|
1416
1457
|
},
|
|
1417
1458
|
{
|
|
1418
1459
|
id: "gpt-5.4-mini",
|
|
1460
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 34.
|
|
1461
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1462
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1463
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1464
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1465
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1466
|
+
probedLimits: [
|
|
1467
|
+
{
|
|
1468
|
+
metric: "tool_count",
|
|
1469
|
+
testedAt: 20,
|
|
1470
|
+
demanded: 3,
|
|
1471
|
+
served: [3, 3, 3],
|
|
1472
|
+
outcome: "held",
|
|
1473
|
+
trials: 3,
|
|
1474
|
+
lastProbed: "2026-08-28",
|
|
1475
|
+
probeRunId: 34,
|
|
1476
|
+
route: "direct",
|
|
1477
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1478
|
+
}
|
|
1479
|
+
],
|
|
1419
1480
|
verifiedAgainstDocs: "2026-05-17",
|
|
1420
1481
|
provider: "openai",
|
|
1421
1482
|
status: "current",
|
|
@@ -1459,6 +1520,26 @@ var PROFILES_RAW = [
|
|
|
1459
1520
|
},
|
|
1460
1521
|
{
|
|
1461
1522
|
id: "gpt-5.4-nano",
|
|
1523
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 35.
|
|
1524
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1525
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1526
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1527
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1528
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1529
|
+
probedLimits: [
|
|
1530
|
+
{
|
|
1531
|
+
metric: "tool_count",
|
|
1532
|
+
testedAt: 20,
|
|
1533
|
+
demanded: 3,
|
|
1534
|
+
served: [3, 3, 3],
|
|
1535
|
+
outcome: "held",
|
|
1536
|
+
trials: 3,
|
|
1537
|
+
lastProbed: "2026-08-28",
|
|
1538
|
+
probeRunId: 35,
|
|
1539
|
+
route: "direct",
|
|
1540
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1541
|
+
}
|
|
1542
|
+
],
|
|
1462
1543
|
verifiedAgainstDocs: "2026-05-17",
|
|
1463
1544
|
provider: "openai",
|
|
1464
1545
|
status: "current",
|
|
@@ -2074,6 +2155,26 @@ var PROFILES_RAW = [
|
|
|
2074
2155
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
2075
2156
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
2076
2157
|
id: "claude-sonnet-5",
|
|
2158
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 36.
|
|
2159
|
+
// This model declared no cliffs, which is why the offline check had
|
|
2160
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
2161
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
2162
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
2163
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
2164
|
+
probedLimits: [
|
|
2165
|
+
{
|
|
2166
|
+
metric: "tool_count",
|
|
2167
|
+
testedAt: 20,
|
|
2168
|
+
demanded: 3,
|
|
2169
|
+
served: [3, 3, 3],
|
|
2170
|
+
outcome: "held",
|
|
2171
|
+
trials: 3,
|
|
2172
|
+
lastProbed: "2026-08-28",
|
|
2173
|
+
probeRunId: 36,
|
|
2174
|
+
route: "direct",
|
|
2175
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
2176
|
+
}
|
|
2177
|
+
],
|
|
2077
2178
|
verifiedAgainstDocs: "2026-08-14",
|
|
2078
2179
|
provider: "anthropic",
|
|
2079
2180
|
status: "current",
|
|
@@ -2419,7 +2520,11 @@ function rowToProfile(row) {
|
|
|
2419
2520
|
// latency_tier → undefined (latencyTierOf derives from tags; not
|
|
2420
2521
|
// safety-critical). archetype_conventions already array-validated above.
|
|
2421
2522
|
latencyTier: normalizeLatencyTier(row.latency_tier),
|
|
2422
|
-
archetypeConventions: row.archetype_conventions ?? void 0
|
|
2523
|
+
archetypeConventions: row.archetype_conventions ?? void 0,
|
|
2524
|
+
// alpha.95 — measured negatives. NOT safety-critical (nothing routes on
|
|
2525
|
+
// them), so a type-mismatch degrades to undefined rather than nulling
|
|
2526
|
+
// the whole profile the way a corrupt `cliffs` does.
|
|
2527
|
+
probedLimits: Array.isArray(row.probed_limits) ? row.probed_limits : void 0
|
|
2423
2528
|
};
|
|
2424
2529
|
} catch {
|
|
2425
2530
|
return null;
|
|
@@ -2459,7 +2564,12 @@ function profileToRow(profile, opts = {}) {
|
|
|
2459
2564
|
// so a reseed from bundled profiles makes the latency lever + schema
|
|
2460
2565
|
// conventions live warm. Closes the silently-dropped-field gap.
|
|
2461
2566
|
latency_tier: profile.latencyTier ?? null,
|
|
2462
|
-
archetype_conventions: profile.archetypeConventions ?? null
|
|
2567
|
+
archetype_conventions: profile.archetypeConventions ?? null,
|
|
2568
|
+
// alpha.95 — round-trip the measured negatives. The s56 rule: a new
|
|
2569
|
+
// ModelProfile field must land in migration + BOTH transformers + the
|
|
2570
|
+
// config-endpoint cols list + the parity test, or it is theater for every
|
|
2571
|
+
// brain-connected consumer.
|
|
2572
|
+
probed_limits: profile.probedLimits ?? null
|
|
2463
2573
|
};
|
|
2464
2574
|
if (opts.verifiedAgainstDocs !== void 0) {
|
|
2465
2575
|
row.verified_against_docs = opts.verifiedAgainstDocs;
|
|
@@ -6464,7 +6574,7 @@ function validateFinalFit(ir, profile, tokens) {
|
|
|
6464
6574
|
}
|
|
6465
6575
|
|
|
6466
6576
|
// src/version.ts
|
|
6467
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
6577
|
+
var LIBRARY_VERSION = "2.0.0-alpha.95";
|
|
6468
6578
|
|
|
6469
6579
|
// src/pricing-brain.ts
|
|
6470
6580
|
function isPricingRow(x) {
|
|
@@ -11403,6 +11513,7 @@ function compile2(ir, opts) {
|
|
|
11403
11513
|
isMeasuredFailureBrainActive,
|
|
11404
11514
|
isMeasuredFailureGateEnabledFromEnv,
|
|
11405
11515
|
isModelReachable,
|
|
11516
|
+
isModelUnavailableMessage,
|
|
11406
11517
|
isPromotionsBrainActive,
|
|
11407
11518
|
isProviderReachable,
|
|
11408
11519
|
judgeMeasuredFailure,
|
package/dist/index.mjs
CHANGED
|
@@ -7,7 +7,7 @@ import {
|
|
|
7
7
|
LIBRARY_VERSION,
|
|
8
8
|
createKeyHealthRoute,
|
|
9
9
|
keyFingerprint
|
|
10
|
-
} from "./chunk-
|
|
10
|
+
} from "./chunk-JFM5A556.mjs";
|
|
11
11
|
import {
|
|
12
12
|
ARCHETYPE_FAMILY_FITS,
|
|
13
13
|
BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE,
|
|
@@ -63,7 +63,7 @@ import {
|
|
|
63
63
|
runAdvisor,
|
|
64
64
|
setTokenizer,
|
|
65
65
|
wilsonLowerBound
|
|
66
|
-
} from "./chunk-
|
|
66
|
+
} from "./chunk-IRXC7WEQ.mjs";
|
|
67
67
|
import {
|
|
68
68
|
ABSOLUTE_FLOOR,
|
|
69
69
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
@@ -100,7 +100,7 @@ import {
|
|
|
100
100
|
policySetHas,
|
|
101
101
|
readBrainReadEnv,
|
|
102
102
|
resolveProviderKey
|
|
103
|
-
} from "./chunk-
|
|
103
|
+
} from "./chunk-BJD75WCL.mjs";
|
|
104
104
|
import {
|
|
105
105
|
ALL_ARCHETYPES,
|
|
106
106
|
DIALECT_VERSION,
|
|
@@ -127,7 +127,7 @@ import {
|
|
|
127
127
|
registerProfiles,
|
|
128
128
|
resolveModelAlias,
|
|
129
129
|
tryGetProfile
|
|
130
|
-
} from "./chunk-
|
|
130
|
+
} from "./chunk-D3BVYXNA.mjs";
|
|
131
131
|
import {
|
|
132
132
|
emitAdvisoryFired,
|
|
133
133
|
emitCompileDone,
|
|
@@ -4397,6 +4397,7 @@ export {
|
|
|
4397
4397
|
isMeasuredFailureBrainActive,
|
|
4398
4398
|
isMeasuredFailureGateEnabledFromEnv,
|
|
4399
4399
|
isModelReachable,
|
|
4400
|
+
isModelUnavailableMessage,
|
|
4400
4401
|
isPromotionsBrainActive,
|
|
4401
4402
|
isProviderReachable,
|
|
4402
4403
|
judgeMeasuredFailure,
|
package/dist/key-health.js
CHANGED
|
@@ -28,7 +28,7 @@ __export(key_health_exports, {
|
|
|
28
28
|
module.exports = __toCommonJS(key_health_exports);
|
|
29
29
|
|
|
30
30
|
// src/version.ts
|
|
31
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
31
|
+
var LIBRARY_VERSION = "2.0.0-alpha.95";
|
|
32
32
|
|
|
33
33
|
// src/key-health.ts
|
|
34
34
|
var JSON_HEADERS = { "Content-Type": "application/json" };
|
package/dist/key-health.mjs
CHANGED
package/dist/probe.js
CHANGED
|
@@ -161,6 +161,26 @@ var PROFILES_RAW = [
|
|
|
161
161
|
// on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
|
|
162
162
|
// `thinking` field, so kgauto never sends the param. (L-049/L-081.)
|
|
163
163
|
id: "claude-fable-5",
|
|
164
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 32.
|
|
165
|
+
// This model declared no cliffs, which is why the offline check had
|
|
166
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
167
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
168
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
169
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
170
|
+
probedLimits: [
|
|
171
|
+
{
|
|
172
|
+
metric: "tool_count",
|
|
173
|
+
testedAt: 20,
|
|
174
|
+
demanded: 3,
|
|
175
|
+
served: [3, 3, 3],
|
|
176
|
+
outcome: "held",
|
|
177
|
+
trials: 3,
|
|
178
|
+
lastProbed: "2026-08-28",
|
|
179
|
+
probeRunId: 32,
|
|
180
|
+
route: "direct",
|
|
181
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
182
|
+
}
|
|
183
|
+
],
|
|
164
184
|
verifiedAgainstDocs: "2026-06-21",
|
|
165
185
|
provider: "anthropic",
|
|
166
186
|
status: "current",
|
|
@@ -281,6 +301,26 @@ var PROFILES_RAW = [
|
|
|
281
301
|
// failure, where a template-clone of PRICE (the field that actually varies,
|
|
282
302
|
// there by 2.5-3.75x) went unverified for 65 days.
|
|
283
303
|
id: "claude-opus-5",
|
|
304
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 33.
|
|
305
|
+
// This model declared no cliffs, which is why the offline check had
|
|
306
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
307
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
308
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
309
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
310
|
+
probedLimits: [
|
|
311
|
+
{
|
|
312
|
+
metric: "tool_count",
|
|
313
|
+
testedAt: 20,
|
|
314
|
+
demanded: 3,
|
|
315
|
+
served: [3, 3, 3],
|
|
316
|
+
outcome: "held",
|
|
317
|
+
trials: 3,
|
|
318
|
+
lastProbed: "2026-08-28",
|
|
319
|
+
probeRunId: 33,
|
|
320
|
+
route: "direct",
|
|
321
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
322
|
+
}
|
|
323
|
+
],
|
|
284
324
|
verifiedAgainstDocs: "2026-07-25",
|
|
285
325
|
provider: "anthropic",
|
|
286
326
|
status: "current",
|
|
@@ -1159,6 +1199,26 @@ var PROFILES_RAW = [
|
|
|
1159
1199
|
},
|
|
1160
1200
|
{
|
|
1161
1201
|
id: "gpt-5.4-mini",
|
|
1202
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 34.
|
|
1203
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1204
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1205
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1206
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1207
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1208
|
+
probedLimits: [
|
|
1209
|
+
{
|
|
1210
|
+
metric: "tool_count",
|
|
1211
|
+
testedAt: 20,
|
|
1212
|
+
demanded: 3,
|
|
1213
|
+
served: [3, 3, 3],
|
|
1214
|
+
outcome: "held",
|
|
1215
|
+
trials: 3,
|
|
1216
|
+
lastProbed: "2026-08-28",
|
|
1217
|
+
probeRunId: 34,
|
|
1218
|
+
route: "direct",
|
|
1219
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1220
|
+
}
|
|
1221
|
+
],
|
|
1162
1222
|
verifiedAgainstDocs: "2026-05-17",
|
|
1163
1223
|
provider: "openai",
|
|
1164
1224
|
status: "current",
|
|
@@ -1202,6 +1262,26 @@ var PROFILES_RAW = [
|
|
|
1202
1262
|
},
|
|
1203
1263
|
{
|
|
1204
1264
|
id: "gpt-5.4-nano",
|
|
1265
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 35.
|
|
1266
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1267
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1268
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1269
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1270
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1271
|
+
probedLimits: [
|
|
1272
|
+
{
|
|
1273
|
+
metric: "tool_count",
|
|
1274
|
+
testedAt: 20,
|
|
1275
|
+
demanded: 3,
|
|
1276
|
+
served: [3, 3, 3],
|
|
1277
|
+
outcome: "held",
|
|
1278
|
+
trials: 3,
|
|
1279
|
+
lastProbed: "2026-08-28",
|
|
1280
|
+
probeRunId: 35,
|
|
1281
|
+
route: "direct",
|
|
1282
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1283
|
+
}
|
|
1284
|
+
],
|
|
1205
1285
|
verifiedAgainstDocs: "2026-05-17",
|
|
1206
1286
|
provider: "openai",
|
|
1207
1287
|
status: "current",
|
|
@@ -1817,6 +1897,26 @@ var PROFILES_RAW = [
|
|
|
1817
1897
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1818
1898
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1819
1899
|
id: "claude-sonnet-5",
|
|
1900
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 36.
|
|
1901
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1902
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1903
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1904
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1905
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1906
|
+
probedLimits: [
|
|
1907
|
+
{
|
|
1908
|
+
metric: "tool_count",
|
|
1909
|
+
testedAt: 20,
|
|
1910
|
+
demanded: 3,
|
|
1911
|
+
served: [3, 3, 3],
|
|
1912
|
+
outcome: "held",
|
|
1913
|
+
trials: 3,
|
|
1914
|
+
lastProbed: "2026-08-28",
|
|
1915
|
+
probeRunId: 36,
|
|
1916
|
+
route: "direct",
|
|
1917
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1918
|
+
}
|
|
1919
|
+
],
|
|
1820
1920
|
verifiedAgainstDocs: "2026-08-14",
|
|
1821
1921
|
provider: "anthropic",
|
|
1822
1922
|
status: "current",
|
|
@@ -2153,7 +2253,11 @@ function rowToProfile(row) {
|
|
|
2153
2253
|
// latency_tier → undefined (latencyTierOf derives from tags; not
|
|
2154
2254
|
// safety-critical). archetype_conventions already array-validated above.
|
|
2155
2255
|
latencyTier: normalizeLatencyTier(row.latency_tier),
|
|
2156
|
-
archetypeConventions: row.archetype_conventions ?? void 0
|
|
2256
|
+
archetypeConventions: row.archetype_conventions ?? void 0,
|
|
2257
|
+
// alpha.95 — measured negatives. NOT safety-critical (nothing routes on
|
|
2258
|
+
// them), so a type-mismatch degrades to undefined rather than nulling
|
|
2259
|
+
// the whole profile the way a corrupt `cliffs` does.
|
|
2260
|
+
probedLimits: Array.isArray(row.probed_limits) ? row.probed_limits : void 0
|
|
2157
2261
|
};
|
|
2158
2262
|
} catch {
|
|
2159
2263
|
return null;
|
package/dist/probe.mjs
CHANGED
|
@@ -1,12 +1,12 @@
|
|
|
1
1
|
import {
|
|
2
2
|
compile
|
|
3
|
-
} from "./chunk-
|
|
4
|
-
import "./chunk-
|
|
3
|
+
} from "./chunk-IRXC7WEQ.mjs";
|
|
4
|
+
import "./chunk-BJD75WCL.mjs";
|
|
5
5
|
import "./chunk-FEMPY6EC.mjs";
|
|
6
6
|
import {
|
|
7
7
|
deriveCliffEvidenceClass,
|
|
8
8
|
getProfile
|
|
9
|
-
} from "./chunk-
|
|
9
|
+
} from "./chunk-D3BVYXNA.mjs";
|
|
10
10
|
|
|
11
11
|
// src/probe.ts
|
|
12
12
|
function compileSuppressing(ir, opts) {
|
package/dist/profiles.d.mts
CHANGED
|
@@ -61,6 +61,39 @@ interface CliffEvidence {
|
|
|
61
61
|
/** FK into kgauto_guard_probe_runs. */
|
|
62
62
|
probeRunId?: number;
|
|
63
63
|
}
|
|
64
|
+
/**
|
|
65
|
+
* alpha.95 — a MEASURED NEGATIVE: a limit we deliberately probed and which
|
|
66
|
+
* held. It has no home in `CliffRule`, because a CliffRule requires an action
|
|
67
|
+
* and there is nothing to do when nothing broke — which is exactly why the
|
|
68
|
+
* fact was invisible until now. Of 15 `current` bundled models, 5 declared
|
|
69
|
+
* zero cliffs (claude-opus-5, claude-sonnet-5, claude-fable-5, gpt-5.4-mini,
|
|
70
|
+
* gpt-5.4-nano — the models a stranger hardcodes), so the offline check had
|
|
71
|
+
* nothing measured to say about precisely them.
|
|
72
|
+
*
|
|
73
|
+
* `claim` is written to be quotable verbatim and MUST state its own
|
|
74
|
+
* denominator: "served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials"
|
|
75
|
+
* is a fact; "handles 20 tools" is an overclaim of the same run.
|
|
76
|
+
*/
|
|
77
|
+
interface ProbedLimit {
|
|
78
|
+
metric: CliffRule['metric'];
|
|
79
|
+
/** The point actually exercised (e.g. tool-ladder size). */
|
|
80
|
+
testedAt: number;
|
|
81
|
+
/** What the fixture demanded at that point — the denominator. */
|
|
82
|
+
demanded?: number;
|
|
83
|
+
/** Distinct tools served, per trial. */
|
|
84
|
+
served?: number[];
|
|
85
|
+
/** `held` = probed and did not break. `broke` = a cliff we shipped blind. */
|
|
86
|
+
outcome: 'held' | 'broke';
|
|
87
|
+
trials: number;
|
|
88
|
+
/** ISO date of the run. */
|
|
89
|
+
lastProbed: string;
|
|
90
|
+
/** FK into kgauto_guard_probe_runs. */
|
|
91
|
+
probeRunId?: number;
|
|
92
|
+
/** `direct`, or a gateway route — a verdict is a per-route fact. */
|
|
93
|
+
route?: string;
|
|
94
|
+
/** One sentence, quotable, denominator included. */
|
|
95
|
+
claim: string;
|
|
96
|
+
}
|
|
64
97
|
/**
|
|
65
98
|
* alpha.93 — evidence class for a guard, derived when the rule predates the
|
|
66
99
|
* explicit field. Explicit `evidence.class` always wins; the legacy
|
|
@@ -291,6 +324,13 @@ interface ModelProfile {
|
|
|
291
324
|
* See `ArchetypeConvention` for field semantics.
|
|
292
325
|
*/
|
|
293
326
|
archetypeConventions?: ArchetypeConvention[];
|
|
327
|
+
/**
|
|
328
|
+
* alpha.95 — measured negatives from the guard-probe suite's `--discover`
|
|
329
|
+
* mode. Lives in the BUNDLE deliberately: `kgauto-check` runs offline and
|
|
330
|
+
* never reads the brain, so evidence that lands brain-only is invisible to
|
|
331
|
+
* the artifact first contact leads with.
|
|
332
|
+
*/
|
|
333
|
+
probedLimits?: ProbedLimit[];
|
|
294
334
|
/**
|
|
295
335
|
* alpha.87 — true ONLY on profiles synthesized by
|
|
296
336
|
* `onUnprofiledModel: 'best-effort'`. Lets compile() (and any consumer)
|
|
@@ -377,4 +417,4 @@ declare function allProfiles(): readonly ModelProfile[];
|
|
|
377
417
|
declare function allProfilesRaw(): readonly ModelProfile[];
|
|
378
418
|
declare function profilesByProvider(provider: Provider): readonly ModelProfile[];
|
|
379
419
|
|
|
380
|
-
export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffEvidence, type CliffRule, LATENCY_TIER_MS, type LatencyTier, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, _testClearConsumerProfiles, allProfiles, allProfilesRaw, bestEffortProfile, deriveCliffEvidenceClass, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile };
|
|
420
|
+
export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffEvidence, type CliffRule, LATENCY_TIER_MS, type LatencyTier, type LoweringSpec, type ModelProfile, type ProbedLimit, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, _testClearConsumerProfiles, allProfiles, allProfilesRaw, bestEffortProfile, deriveCliffEvidenceClass, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile };
|
package/dist/profiles.d.ts
CHANGED
|
@@ -61,6 +61,39 @@ interface CliffEvidence {
|
|
|
61
61
|
/** FK into kgauto_guard_probe_runs. */
|
|
62
62
|
probeRunId?: number;
|
|
63
63
|
}
|
|
64
|
+
/**
|
|
65
|
+
* alpha.95 — a MEASURED NEGATIVE: a limit we deliberately probed and which
|
|
66
|
+
* held. It has no home in `CliffRule`, because a CliffRule requires an action
|
|
67
|
+
* and there is nothing to do when nothing broke — which is exactly why the
|
|
68
|
+
* fact was invisible until now. Of 15 `current` bundled models, 5 declared
|
|
69
|
+
* zero cliffs (claude-opus-5, claude-sonnet-5, claude-fable-5, gpt-5.4-mini,
|
|
70
|
+
* gpt-5.4-nano — the models a stranger hardcodes), so the offline check had
|
|
71
|
+
* nothing measured to say about precisely them.
|
|
72
|
+
*
|
|
73
|
+
* `claim` is written to be quotable verbatim and MUST state its own
|
|
74
|
+
* denominator: "served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials"
|
|
75
|
+
* is a fact; "handles 20 tools" is an overclaim of the same run.
|
|
76
|
+
*/
|
|
77
|
+
interface ProbedLimit {
|
|
78
|
+
metric: CliffRule['metric'];
|
|
79
|
+
/** The point actually exercised (e.g. tool-ladder size). */
|
|
80
|
+
testedAt: number;
|
|
81
|
+
/** What the fixture demanded at that point — the denominator. */
|
|
82
|
+
demanded?: number;
|
|
83
|
+
/** Distinct tools served, per trial. */
|
|
84
|
+
served?: number[];
|
|
85
|
+
/** `held` = probed and did not break. `broke` = a cliff we shipped blind. */
|
|
86
|
+
outcome: 'held' | 'broke';
|
|
87
|
+
trials: number;
|
|
88
|
+
/** ISO date of the run. */
|
|
89
|
+
lastProbed: string;
|
|
90
|
+
/** FK into kgauto_guard_probe_runs. */
|
|
91
|
+
probeRunId?: number;
|
|
92
|
+
/** `direct`, or a gateway route — a verdict is a per-route fact. */
|
|
93
|
+
route?: string;
|
|
94
|
+
/** One sentence, quotable, denominator included. */
|
|
95
|
+
claim: string;
|
|
96
|
+
}
|
|
64
97
|
/**
|
|
65
98
|
* alpha.93 — evidence class for a guard, derived when the rule predates the
|
|
66
99
|
* explicit field. Explicit `evidence.class` always wins; the legacy
|
|
@@ -291,6 +324,13 @@ interface ModelProfile {
|
|
|
291
324
|
* See `ArchetypeConvention` for field semantics.
|
|
292
325
|
*/
|
|
293
326
|
archetypeConventions?: ArchetypeConvention[];
|
|
327
|
+
/**
|
|
328
|
+
* alpha.95 — measured negatives from the guard-probe suite's `--discover`
|
|
329
|
+
* mode. Lives in the BUNDLE deliberately: `kgauto-check` runs offline and
|
|
330
|
+
* never reads the brain, so evidence that lands brain-only is invisible to
|
|
331
|
+
* the artifact first contact leads with.
|
|
332
|
+
*/
|
|
333
|
+
probedLimits?: ProbedLimit[];
|
|
294
334
|
/**
|
|
295
335
|
* alpha.87 — true ONLY on profiles synthesized by
|
|
296
336
|
* `onUnprofiledModel: 'best-effort'`. Lets compile() (and any consumer)
|
|
@@ -377,4 +417,4 @@ declare function allProfiles(): readonly ModelProfile[];
|
|
|
377
417
|
declare function allProfilesRaw(): readonly ModelProfile[];
|
|
378
418
|
declare function profilesByProvider(provider: Provider): readonly ModelProfile[];
|
|
379
419
|
|
|
380
|
-
export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffEvidence, type CliffRule, LATENCY_TIER_MS, type LatencyTier, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, _testClearConsumerProfiles, allProfiles, allProfilesRaw, bestEffortProfile, deriveCliffEvidenceClass, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile };
|
|
420
|
+
export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffEvidence, type CliffRule, LATENCY_TIER_MS, type LatencyTier, type LoweringSpec, type ModelProfile, type ProbedLimit, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, _testClearConsumerProfiles, allProfiles, allProfilesRaw, bestEffortProfile, deriveCliffEvidenceClass, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile };
|
package/dist/profiles.js
CHANGED
|
@@ -92,6 +92,26 @@ var PROFILES_RAW = [
|
|
|
92
92
|
// on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
|
|
93
93
|
// `thinking` field, so kgauto never sends the param. (L-049/L-081.)
|
|
94
94
|
id: "claude-fable-5",
|
|
95
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 32.
|
|
96
|
+
// This model declared no cliffs, which is why the offline check had
|
|
97
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
98
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
99
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
100
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
101
|
+
probedLimits: [
|
|
102
|
+
{
|
|
103
|
+
metric: "tool_count",
|
|
104
|
+
testedAt: 20,
|
|
105
|
+
demanded: 3,
|
|
106
|
+
served: [3, 3, 3],
|
|
107
|
+
outcome: "held",
|
|
108
|
+
trials: 3,
|
|
109
|
+
lastProbed: "2026-08-28",
|
|
110
|
+
probeRunId: 32,
|
|
111
|
+
route: "direct",
|
|
112
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
113
|
+
}
|
|
114
|
+
],
|
|
95
115
|
verifiedAgainstDocs: "2026-06-21",
|
|
96
116
|
provider: "anthropic",
|
|
97
117
|
status: "current",
|
|
@@ -212,6 +232,26 @@ var PROFILES_RAW = [
|
|
|
212
232
|
// failure, where a template-clone of PRICE (the field that actually varies,
|
|
213
233
|
// there by 2.5-3.75x) went unverified for 65 days.
|
|
214
234
|
id: "claude-opus-5",
|
|
235
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 33.
|
|
236
|
+
// This model declared no cliffs, which is why the offline check had
|
|
237
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
238
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
239
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
240
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
241
|
+
probedLimits: [
|
|
242
|
+
{
|
|
243
|
+
metric: "tool_count",
|
|
244
|
+
testedAt: 20,
|
|
245
|
+
demanded: 3,
|
|
246
|
+
served: [3, 3, 3],
|
|
247
|
+
outcome: "held",
|
|
248
|
+
trials: 3,
|
|
249
|
+
lastProbed: "2026-08-28",
|
|
250
|
+
probeRunId: 33,
|
|
251
|
+
route: "direct",
|
|
252
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
253
|
+
}
|
|
254
|
+
],
|
|
215
255
|
verifiedAgainstDocs: "2026-07-25",
|
|
216
256
|
provider: "anthropic",
|
|
217
257
|
status: "current",
|
|
@@ -1090,6 +1130,26 @@ var PROFILES_RAW = [
|
|
|
1090
1130
|
},
|
|
1091
1131
|
{
|
|
1092
1132
|
id: "gpt-5.4-mini",
|
|
1133
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 34.
|
|
1134
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1135
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1136
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1137
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1138
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1139
|
+
probedLimits: [
|
|
1140
|
+
{
|
|
1141
|
+
metric: "tool_count",
|
|
1142
|
+
testedAt: 20,
|
|
1143
|
+
demanded: 3,
|
|
1144
|
+
served: [3, 3, 3],
|
|
1145
|
+
outcome: "held",
|
|
1146
|
+
trials: 3,
|
|
1147
|
+
lastProbed: "2026-08-28",
|
|
1148
|
+
probeRunId: 34,
|
|
1149
|
+
route: "direct",
|
|
1150
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1151
|
+
}
|
|
1152
|
+
],
|
|
1093
1153
|
verifiedAgainstDocs: "2026-05-17",
|
|
1094
1154
|
provider: "openai",
|
|
1095
1155
|
status: "current",
|
|
@@ -1133,6 +1193,26 @@ var PROFILES_RAW = [
|
|
|
1133
1193
|
},
|
|
1134
1194
|
{
|
|
1135
1195
|
id: "gpt-5.4-nano",
|
|
1196
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 35.
|
|
1197
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1198
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1199
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1200
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1201
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1202
|
+
probedLimits: [
|
|
1203
|
+
{
|
|
1204
|
+
metric: "tool_count",
|
|
1205
|
+
testedAt: 20,
|
|
1206
|
+
demanded: 3,
|
|
1207
|
+
served: [3, 3, 3],
|
|
1208
|
+
outcome: "held",
|
|
1209
|
+
trials: 3,
|
|
1210
|
+
lastProbed: "2026-08-28",
|
|
1211
|
+
probeRunId: 35,
|
|
1212
|
+
route: "direct",
|
|
1213
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1214
|
+
}
|
|
1215
|
+
],
|
|
1136
1216
|
verifiedAgainstDocs: "2026-05-17",
|
|
1137
1217
|
provider: "openai",
|
|
1138
1218
|
status: "current",
|
|
@@ -1748,6 +1828,26 @@ var PROFILES_RAW = [
|
|
|
1748
1828
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1749
1829
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1750
1830
|
id: "claude-sonnet-5",
|
|
1831
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 36.
|
|
1832
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1833
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1834
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1835
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1836
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1837
|
+
probedLimits: [
|
|
1838
|
+
{
|
|
1839
|
+
metric: "tool_count",
|
|
1840
|
+
testedAt: 20,
|
|
1841
|
+
demanded: 3,
|
|
1842
|
+
served: [3, 3, 3],
|
|
1843
|
+
outcome: "held",
|
|
1844
|
+
trials: 3,
|
|
1845
|
+
lastProbed: "2026-08-28",
|
|
1846
|
+
probeRunId: 36,
|
|
1847
|
+
route: "direct",
|
|
1848
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1849
|
+
}
|
|
1850
|
+
],
|
|
1751
1851
|
verifiedAgainstDocs: "2026-08-14",
|
|
1752
1852
|
provider: "anthropic",
|
|
1753
1853
|
status: "current",
|
package/dist/profiles.mjs
CHANGED
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@warmdrift/kgauto-compiler",
|
|
3
|
-
"version": "2.0.0-alpha.
|
|
3
|
+
"version": "2.0.0-alpha.95",
|
|
4
4
|
"description": "Prompt compiler with executable provider knowledge for multi-model AI apps: normalized multi-provider transport with fallback chains, compile-time cliff guards, a curated model registry, and a telemetry flight recorder. Swap models without rewriting prompts.",
|
|
5
5
|
"main": "./dist/index.js",
|
|
6
6
|
"module": "./dist/index.mjs",
|
|
@@ -114,4 +114,4 @@
|
|
|
114
114
|
"typescript": "^5.7.0",
|
|
115
115
|
"vitest": "^3.1.0"
|
|
116
116
|
}
|
|
117
|
-
}
|
|
117
|
+
}
|