@warmdrift/kgauto-compiler 2.0.0-alpha.94 → 2.0.0-alpha.96
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-VXSSZFES.mjs → chunk-26IS3Y6Z.mjs} +1 -1
- package/dist/{chunk-OAJZLZBN.mjs → chunk-2D4H7HGZ.mjs} +1 -1
- package/dist/{chunk-57XCZGN5.mjs → chunk-AMV3M2KH.mjs} +187 -36
- package/dist/{chunk-TN6UIX56.mjs → chunk-SRPRNTDT.mjs} +17 -7
- package/dist/glassbox-routes/index.js +187 -36
- 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 +205 -42
- package/dist/index.mjs +5 -4
- package/dist/key-health.js +1 -1
- package/dist/key-health.mjs +1 -1
- package/dist/probe.js +196 -40
- package/dist/probe.mjs +3 -3
- package/dist/profiles.d.mts +45 -2
- package/dist/profiles.d.ts +45 -2
- package/dist/profiles.js +187 -36
- package/dist/profiles.mjs +1 -1
- package/package.json +2 -2
|
@@ -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",
|
|
@@ -803,18 +843,22 @@ var PROFILES_RAW = [
|
|
|
803
843
|
maxContextTokens: 1e6,
|
|
804
844
|
maxOutputTokens: 384e3,
|
|
805
845
|
maxTools: 16,
|
|
806
|
-
|
|
846
|
+
// alpha.96 — CORRECTED from `false`, which was AI-generated at greenfield
|
|
847
|
+
// (2026-04-25) and never re-checked. Measured 2026-08-28 direct to
|
|
848
|
+
// api.deepseek.com, no kgauto in the path, 6 tools offered, 3 trials:
|
|
849
|
+
// 3 parallel tool calls returned, 3/3. L-040's original evidence said
|
|
850
|
+
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
851
|
+
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
852
|
+
parallelToolCalls: true,
|
|
807
853
|
structuredOutput: "native",
|
|
808
854
|
systemPromptMode: "inline",
|
|
809
855
|
streaming: true,
|
|
810
|
-
|
|
811
|
-
|
|
812
|
-
|
|
813
|
-
|
|
814
|
-
|
|
815
|
-
|
|
816
|
-
}
|
|
817
|
-
],
|
|
856
|
+
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
857
|
+
// It restated `parallelToolCalls: false` as a threshold and was falsified by
|
|
858
|
+
// the same measurement. It is not kept-and-flagged (the usual lifecycle)
|
|
859
|
+
// because it was never a real failure mode: it was a second enforcement of a
|
|
860
|
+
// wrong belief, and keeping it active would keep amputating tools.
|
|
861
|
+
cliffs: [],
|
|
818
862
|
costInputPer1m: 0.44,
|
|
819
863
|
costOutputPer1m: 1.32,
|
|
820
864
|
lowering: {
|
|
@@ -826,7 +870,12 @@ var PROFILES_RAW = [
|
|
|
826
870
|
{ signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
|
|
827
871
|
],
|
|
828
872
|
strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
|
|
829
|
-
|
|
873
|
+
// alpha.96 — 'parallel_tools' REMOVED: measurably false (2026-08-28, direct
|
|
874
|
+
// to api.deepseek.com, 6 tools, 3 trials, 3 parallel calls, both models).
|
|
875
|
+
// 'large_tool_sets' KEPT and explicitly unverified — the probe offered 6 of
|
|
876
|
+
// a declared maxTools 16, so it says nothing about behaviour at 16. Not
|
|
877
|
+
// falsified is not the same as confirmed.
|
|
878
|
+
weaknesses: ["large_tool_sets"],
|
|
830
879
|
// alpha.47 — explicit slow override. Tag derivation would say 'medium'
|
|
831
880
|
// (no 'latency' weakness, no 'speed' strength), but the alpha.46 shadow
|
|
832
881
|
// probe MEASURED deepseek-v4-flash at 20485ms served (2026-06-03) — ~2.3×
|
|
@@ -838,7 +887,11 @@ var PROFILES_RAW = [
|
|
|
838
887
|
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
839
888
|
// Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
|
|
840
889
|
// classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
|
|
841
|
-
//
|
|
890
|
+
// alpha.96 — the hunt half of this sentence is RETRACTED: "sequential tools
|
|
891
|
+
// — L-040" was falsified 2026-08-28 (direct measurement, 3 parallel tool
|
|
892
|
+
// calls, both models). Do not use this comment to re-lower `hunt`; it is
|
|
893
|
+
// the exact prose that justified the wrong score. The reasoning-depth
|
|
894
|
+
// rationale is untouched and still carries `critique`.
|
|
842
895
|
archetypePerf: {
|
|
843
896
|
classify: 7,
|
|
844
897
|
// brain-validated, 169 rows
|
|
@@ -850,8 +903,16 @@ var PROFILES_RAW = [
|
|
|
850
903
|
plan: 5,
|
|
851
904
|
extract: 5,
|
|
852
905
|
critique: 4,
|
|
853
|
-
|
|
854
|
-
//
|
|
906
|
+
// alpha.96 — was 4. Its ONLY written justification was "sequential tool
|
|
907
|
+
// calls only — L-040", which is falsified. A score derived from a false
|
|
908
|
+
// premise cannot be kept, and re-guessing a penalty would just relaunder
|
|
909
|
+
// the same belief. Set to the floor (6) = "no evidence of deficiency",
|
|
910
|
+
// NOT an endorsement. This is a JUDGMENT, not a measurement, and it is
|
|
911
|
+
// the honest direction of the two: the only hunt-shaped evidence we have
|
|
912
|
+
// is positive (a research-agent fixture, 3 of 3 demanded tools correctly
|
|
913
|
+
// selected from a 6-tool menu, 3/3 trials, both models). Replace with a
|
|
914
|
+
// real number from the golden-eval hunt corpus — that is what it is for.
|
|
915
|
+
hunt: 6
|
|
855
916
|
},
|
|
856
917
|
// alpha.43 — family-level conventions for the `deepseek-chat` family
|
|
857
918
|
// (V4-Flash + future non-thinking-mode members). V4-Flash struggles
|
|
@@ -876,18 +937,22 @@ var PROFILES_RAW = [
|
|
|
876
937
|
maxContextTokens: 1e6,
|
|
877
938
|
maxOutputTokens: 384e3,
|
|
878
939
|
maxTools: 16,
|
|
879
|
-
|
|
940
|
+
// alpha.96 — CORRECTED from `false`, which was AI-generated at greenfield
|
|
941
|
+
// (2026-04-25) and never re-checked. Measured 2026-08-28 direct to
|
|
942
|
+
// api.deepseek.com, no kgauto in the path, 6 tools offered, 3 trials:
|
|
943
|
+
// 3 parallel tool calls returned, 3/3. L-040's original evidence said
|
|
944
|
+
// "7-8 calls per step" — which was never a claim that only one tool may be
|
|
945
|
+
// offered. See the note in lower.ts:lowerDeepSeek.
|
|
946
|
+
parallelToolCalls: true,
|
|
880
947
|
structuredOutput: "native",
|
|
881
948
|
systemPromptMode: "inline",
|
|
882
949
|
streaming: true,
|
|
883
|
-
|
|
884
|
-
|
|
885
|
-
|
|
886
|
-
|
|
887
|
-
|
|
888
|
-
|
|
889
|
-
}
|
|
890
|
-
],
|
|
950
|
+
// alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
|
|
951
|
+
// It restated `parallelToolCalls: false` as a threshold and was falsified by
|
|
952
|
+
// the same measurement. It is not kept-and-flagged (the usual lifecycle)
|
|
953
|
+
// because it was never a real failure mode: it was a second enforcement of a
|
|
954
|
+
// wrong belief, and keeping it active would keep amputating tools.
|
|
955
|
+
cliffs: [],
|
|
891
956
|
// Verified against the live DeepSeek pricing page 2026-06-22: the 75%-off
|
|
892
957
|
// launch "promo" ($0.435/$0.87) did NOT revert on 2026-05-31 — it became the
|
|
893
958
|
// standing price (docs show $0.435 in / $0.87 out cache-miss, $0.003625
|
|
@@ -905,7 +970,12 @@ var PROFILES_RAW = [
|
|
|
905
970
|
{ signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
|
|
906
971
|
],
|
|
907
972
|
strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
|
|
908
|
-
|
|
973
|
+
// alpha.96 — 'parallel_tools' REMOVED: measurably false (2026-08-28, direct
|
|
974
|
+
// to api.deepseek.com, 6 tools, 3 trials, 3 parallel calls, both models).
|
|
975
|
+
// 'large_tool_sets' KEPT and explicitly unverified — the probe offered 6 of
|
|
976
|
+
// a declared maxTools 16, so it says nothing about behaviour at 16. Not
|
|
977
|
+
// falsified is not the same as confirmed.
|
|
978
|
+
weaknesses: ["large_tool_sets"],
|
|
909
979
|
// alpha.47 — explicit slow override. Measured 47722ms on the alpha.46
|
|
910
980
|
// shadow probe (2026-06-03) — it's an extended-thinking reasoner, slowest
|
|
911
981
|
// of the served set. Tag derivation would say 'medium'; the measurement says
|
|
@@ -924,14 +994,24 @@ var PROFILES_RAW = [
|
|
|
924
994
|
summarize: 7,
|
|
925
995
|
extract: 6,
|
|
926
996
|
transform: 6,
|
|
927
|
-
|
|
928
|
-
//
|
|
997
|
+
// alpha.96 — was 4. Its ONLY written justification was "sequential tool
|
|
998
|
+
// calls only — L-040", which is falsified. A score derived from a false
|
|
999
|
+
// premise cannot be kept, and re-guessing a penalty would just relaunder
|
|
1000
|
+
// the same belief. Set to the floor (6) = "no evidence of deficiency",
|
|
1001
|
+
// NOT an endorsement. This is a JUDGMENT, not a measurement, and it is
|
|
1002
|
+
// the honest direction of the two: the only hunt-shaped evidence we have
|
|
1003
|
+
// is positive (a research-agent fixture, 3 of 3 demanded tools correctly
|
|
1004
|
+
// selected from a 6-tool menu, 3/3 trials, both models). Replace with a
|
|
1005
|
+
// real number from the golden-eval hunt corpus — that is what it is for.
|
|
1006
|
+
hunt: 6
|
|
929
1007
|
},
|
|
930
1008
|
// alpha.43 — family-level conventions for the `deepseek-reasoner` family
|
|
931
1009
|
// (V4-Pro is the family rep; future thinking-mode members inherit).
|
|
932
1010
|
//
|
|
933
|
-
// The decisive-archetype suffix is the load-bearing piece:
|
|
934
|
-
//
|
|
1011
|
+
// The decisive-archetype suffix is the load-bearing piece. ⚠️ alpha.96:
|
|
1012
|
+
// its evidence was purged as poison on 2026-05-29 (s51) — retained as a
|
|
1013
|
+
// hypothesis, not a finding. Original claim, for the record: 2026-05-28
|
|
1014
|
+
// first real V4-Pro probe on tt-intel/classify
|
|
935
1015
|
// showed 8/10 judge rationales citing "candidate hedges and fails to
|
|
936
1016
|
// commit to a single classification" or "candidate refuses to pick
|
|
937
1017
|
// among the labels." Without the forcing-function suffix, every
|
|
@@ -952,24 +1032,35 @@ var PROFILES_RAW = [
|
|
|
952
1032
|
{
|
|
953
1033
|
archetype: "classify",
|
|
954
1034
|
promptSuffix: "\n\nOutput exactly one classification label. Do not hedge, do not add preamble, do not explain your reasoning. If uncertain, pick the most likely label and commit to it.",
|
|
955
|
-
reason:
|
|
1035
|
+
reason: 'Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. UNEVIDENCED HYPOTHESIS as of alpha.96 \u2014 this reason used to cite a 2026-05-28 V4-Pro probe on tt-intel/classify ("8 of 10 judge rationales cited hedging") and named `brain probe_outcomes` as its source. That source is EMPTY: probed 2026-08-28, zero probe_outcomes rows exist before 2026-06-01 and zero `classify` rows have ever existed. The probe was deleted on 2026-05-29 by the s51 poison purge \u2014 it replayed against truncated `prompt_preview`, which s51 established yields meaningless verdicts. The conclusion outlived the purge of its own evidence and kept shaping prompts for three months. The pointer was wrong too: exclusion-finding ID 20 is a zero-calls `recommend-probe` on classify, confidence low, later declined \u2014 the probe TRIGGER, never its result. KEPT because a promptSuffix is capability-PRESERVING (adds an instruction, removes nothing a raw call had) and is visible in mutations_applied. Re-evidence from the golden-eval corpus or drop it.'
|
|
956
1036
|
},
|
|
957
1037
|
{
|
|
958
1038
|
archetype: "extract",
|
|
959
1039
|
promptSuffix: "\n\nOutput exactly the requested structured fields. Do not hedge, do not add preamble, do not explain your reasoning. If a field is unknown, emit the field with a null/empty value and commit to that \u2014 do not refuse to extract.",
|
|
960
|
-
reason:
|
|
1040
|
+
reason: 'Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. UNEVIDENCED HYPOTHESIS as of alpha.96 \u2014 the supporting evidence NO LONGER EXISTS. The cited 2026-05-28 V4-Pro classify probe ("8/10 judge rationales") was deleted on 2026-05-29 by the s51 poison purge, because it replayed against truncated `prompt_preview` and s51 established such verdicts are meaningless. Probed: zero probe_outcomes rows exist before 2026-06-01 and zero `classify` rows have EVER existed. The pointer was also wrong \u2014 exclusion-finding ID 20 is a zero-calls `recommend-probe` on classify, confidence low, later declined; it was the trigger for the probe, never its result. KEPT anyway because a promptSuffix is capability-PRESERVING (it adds an instruction and removes nothing a raw call had), so the cost of being wrong is bounded and a consumer can see it in mutations_applied. Re-evidence it from the golden-eval corpus or drop it.'
|
|
961
1041
|
},
|
|
962
1042
|
{
|
|
963
1043
|
archetype: "ask",
|
|
964
1044
|
promptSuffix: "\n\nAnswer the question directly. Do not hedge, do not add preamble, do not explain your reasoning unless explicitly asked. If uncertain, give your best answer and commit to it.",
|
|
965
|
-
reason:
|
|
966
|
-
},
|
|
967
|
-
{
|
|
968
|
-
archetype: "hunt",
|
|
969
|
-
whenToolCountAtLeast: 3,
|
|
970
|
-
cliffWarning: "Reasoners (deepseek-reasoner family) produce sequential reasoning chains by design; hunt archetypes with 3+ parallel tools lose substantial throughput. Consider gemini-flash family (parallel-strong) instead \u2014 see L-040 / L-041.",
|
|
971
|
-
reason: "L-040 + L-041 architectural reading; 2026-05-28 V4-Pro probe evidence (exclusion-finding ID 20) confirms reasoner-family hunt fit is structurally weak."
|
|
1045
|
+
reason: 'Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. UNEVIDENCED HYPOTHESIS as of alpha.96 \u2014 the supporting evidence NO LONGER EXISTS. The cited 2026-05-28 V4-Pro classify probe ("8/10 judge rationales") was deleted on 2026-05-29 by the s51 poison purge, because it replayed against truncated `prompt_preview` and s51 established such verdicts are meaningless. Probed: zero probe_outcomes rows exist before 2026-06-01 and zero `classify` rows have EVER existed. The pointer was also wrong \u2014 exclusion-finding ID 20 is a zero-calls `recommend-probe` on classify, confidence low, later declined; it was the trigger for the probe, never its result. KEPT anyway because a promptSuffix is capability-PRESERVING (it adds an instruction and removes nothing a raw call had), so the cost of being wrong is bounded and a consumer can see it in mutations_applied. Re-evidence it from the golden-eval corpus or drop it.'
|
|
972
1046
|
}
|
|
1047
|
+
// alpha.96 — the `hunt` convention was DELETED, not corrected.
|
|
1048
|
+
// It warned that "hunt archetypes with 3+ parallel tools lose substantial
|
|
1049
|
+
// throughput" and steered callers to gemini-flash. Measured false
|
|
1050
|
+
// 2026-08-28: 6 tools offered, 3 of 3 demanded served in parallel, 3/3
|
|
1051
|
+
// trials, both V4 models, direct to the provider.
|
|
1052
|
+
//
|
|
1053
|
+
// Its cited evidence did not say what it claimed. exclusion-finding ID 20
|
|
1054
|
+
// is `intent_archetype: 'classify'` (not hunt), `verdict:
|
|
1055
|
+
// 'recommend-probe'` (i.e. NOBODY PROBED), `confidence: 'low'`, its
|
|
1056
|
+
// message reads "deepseek-v4-pro has zero calls", it was later
|
|
1057
|
+
// `declined`, and its own `parallel_tool_hint: false` was this profile's
|
|
1058
|
+
// own belief round-tripped back as input. A no-data finding on the wrong
|
|
1059
|
+
// archetype was paraphrased into "probe evidence confirms".
|
|
1060
|
+
//
|
|
1061
|
+
// Deleted rather than re-scoped because a cliffWarning is
|
|
1062
|
+
// capability-REDUCING (it steers traffic away from a model) and nothing
|
|
1063
|
+
// evidenced remains. Re-add only from a measured hunt eval.
|
|
973
1064
|
]
|
|
974
1065
|
},
|
|
975
1066
|
// ── OpenAI ──
|
|
@@ -1097,6 +1188,26 @@ var PROFILES_RAW = [
|
|
|
1097
1188
|
},
|
|
1098
1189
|
{
|
|
1099
1190
|
id: "gpt-5.4-mini",
|
|
1191
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 34.
|
|
1192
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1193
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1194
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1195
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1196
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1197
|
+
probedLimits: [
|
|
1198
|
+
{
|
|
1199
|
+
metric: "tool_count",
|
|
1200
|
+
testedAt: 20,
|
|
1201
|
+
demanded: 3,
|
|
1202
|
+
served: [3, 3, 3],
|
|
1203
|
+
outcome: "held",
|
|
1204
|
+
trials: 3,
|
|
1205
|
+
lastProbed: "2026-08-28",
|
|
1206
|
+
probeRunId: 34,
|
|
1207
|
+
route: "direct",
|
|
1208
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1209
|
+
}
|
|
1210
|
+
],
|
|
1100
1211
|
verifiedAgainstDocs: "2026-05-17",
|
|
1101
1212
|
provider: "openai",
|
|
1102
1213
|
status: "current",
|
|
@@ -1140,6 +1251,26 @@ var PROFILES_RAW = [
|
|
|
1140
1251
|
},
|
|
1141
1252
|
{
|
|
1142
1253
|
id: "gpt-5.4-nano",
|
|
1254
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 35.
|
|
1255
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1256
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1257
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1258
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1259
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1260
|
+
probedLimits: [
|
|
1261
|
+
{
|
|
1262
|
+
metric: "tool_count",
|
|
1263
|
+
testedAt: 20,
|
|
1264
|
+
demanded: 3,
|
|
1265
|
+
served: [3, 3, 3],
|
|
1266
|
+
outcome: "held",
|
|
1267
|
+
trials: 3,
|
|
1268
|
+
lastProbed: "2026-08-28",
|
|
1269
|
+
probeRunId: 35,
|
|
1270
|
+
route: "direct",
|
|
1271
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1272
|
+
}
|
|
1273
|
+
],
|
|
1143
1274
|
verifiedAgainstDocs: "2026-05-17",
|
|
1144
1275
|
provider: "openai",
|
|
1145
1276
|
status: "current",
|
|
@@ -1755,6 +1886,26 @@ var PROFILES_RAW = [
|
|
|
1755
1886
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1756
1887
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1757
1888
|
id: "claude-sonnet-5",
|
|
1889
|
+
// alpha.95 — MEASURED, 2026-08-28, guard-probe-suite --discover run 36.
|
|
1890
|
+
// This model declared no cliffs, which is why the offline check had
|
|
1891
|
+
// nothing measured to say about it. The claim states its denominator:
|
|
1892
|
+
// the fixture demands 3 specific tools from a 20-tool menu, so this is
|
|
1893
|
+
// evidence of correct selection + 3-way parallel emission at that menu
|
|
1894
|
+
// size — NOT evidence of 20 concurrent calls.
|
|
1895
|
+
probedLimits: [
|
|
1896
|
+
{
|
|
1897
|
+
metric: "tool_count",
|
|
1898
|
+
testedAt: 20,
|
|
1899
|
+
demanded: 3,
|
|
1900
|
+
served: [3, 3, 3],
|
|
1901
|
+
outcome: "held",
|
|
1902
|
+
trials: 3,
|
|
1903
|
+
lastProbed: "2026-08-28",
|
|
1904
|
+
probeRunId: 36,
|
|
1905
|
+
route: "direct",
|
|
1906
|
+
claim: "Measured 2026-08-28: served 3 of 3 demanded tools from a 20-tool menu, 3/3 trials, direct route."
|
|
1907
|
+
}
|
|
1908
|
+
],
|
|
1758
1909
|
verifiedAgainstDocs: "2026-08-14",
|
|
1759
1910
|
provider: "anthropic",
|
|
1760
1911
|
status: "current",
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import {
|
|
2
2
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
3
3
|
getDefaultFallbackChain
|
|
4
|
-
} from "../chunk-
|
|
4
|
+
} from "../chunk-2D4H7HGZ.mjs";
|
|
5
5
|
import {
|
|
6
6
|
tryGetProfile
|
|
7
|
-
} from "../chunk-
|
|
7
|
+
} from "../chunk-AMV3M2KH.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.96";
|
|
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.96";
|
|
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 };
|