@warmdrift/kgauto-compiler 2.0.0-alpha.95 → 2.0.0-alpha.97

package/dist/glassbox-routes/index.js

@@ -843,18 +843,22 @@ var PROFILES_RAW = [
     maxContextTokens: 1e6,
     maxOutputTokens: 384e3,
     maxTools: 16,
-    parallelToolCalls: false,
+    // alpha.96 — CORRECTED from `false`, which was AI-generated at greenfield
+    // (2026-04-25) and never re-checked. Measured 2026-08-28 direct to
+    // api.deepseek.com, no kgauto in the path, 6 tools offered, 3 trials:
+    // 3 parallel tool calls returned, 3/3. L-040's original evidence said
+    // "7-8 calls per step" — which was never a claim that only one tool may be
+    // offered. See the note in lower.ts:lowerDeepSeek.
+    parallelToolCalls: true,
     structuredOutput: "native",
     systemPromptMode: "inline",
     streaming: true,
-    cliffs: [
-      {
-        metric: "tool_count",
-        threshold: 1,
-        action: "drop_to_top_relevant",
-        reason: "Sequential tool calls only \u2014 L-040"
-      }
-    ],
+    // alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
+    // It restated `parallelToolCalls: false` as a threshold and was falsified by
+    // the same measurement. It is not kept-and-flagged (the usual lifecycle)
+    // because it was never a real failure mode: it was a second enforcement of a
+    // wrong belief, and keeping it active would keep amputating tools.
+    cliffs: [],
     costInputPer1m: 0.44,
     costOutputPer1m: 1.32,
     lowering: {
@@ -866,7 +870,12 @@ var PROFILES_RAW = [
       { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
-    weaknesses: ["parallel_tools", "large_tool_sets"],
+    // alpha.96 — 'parallel_tools' REMOVED: measurably false (2026-08-28, direct
+    // to api.deepseek.com, 6 tools, 3 trials, 3 parallel calls, both models).
+    // 'large_tool_sets' KEPT and explicitly unverified — the probe offered 6 of
+    // a declared maxTools 16, so it says nothing about behaviour at 16. Not
+    // falsified is not the same as confirmed.
+    weaknesses: ["large_tool_sets"],
     // alpha.47 — explicit slow override. Tag derivation would say 'medium'
     // (no 'latency' weakness, no 'speed' strength), but the alpha.46 shadow
     // probe MEASURED deepseek-v4-flash at 20485ms served (2026-06-03) — ~2.3×
@@ -878,7 +887,11 @@ var PROFILES_RAW = [
     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.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
-    // Falls off on hunt (sequential tools — L-040) and reasoning depth.
+    // alpha.96 — the hunt half of this sentence is RETRACTED: "sequential tools
+    // — L-040" was falsified 2026-08-28 (direct measurement, 3 parallel tool
+    // calls, both models). Do not use this comment to re-lower `hunt`; it is
+    // the exact prose that justified the wrong score. The reasoning-depth
+    // rationale is untouched and still carries `critique`.
     archetypePerf: {
       classify: 7,
       // brain-validated, 169 rows
@@ -890,8 +903,16 @@ var PROFILES_RAW = [
       plan: 5,
       extract: 5,
       critique: 4,
-      hunt: 4
-      // sequential tool calls only — L-040
+      // alpha.96 — was 4. Its ONLY written justification was "sequential tool
+      // calls only — L-040", which is falsified. A score derived from a false
+      // premise cannot be kept, and re-guessing a penalty would just relaunder
+      // the same belief. Set to the floor (6) = "no evidence of deficiency",
+      // NOT an endorsement. This is a JUDGMENT, not a measurement, and it is
+      // the honest direction of the two: the only hunt-shaped evidence we have
+      // is positive (a research-agent fixture, 3 of 3 demanded tools correctly
+      // selected from a 6-tool menu, 3/3 trials, both models). Replace with a
+      // real number from the golden-eval hunt corpus — that is what it is for.
+      hunt: 6
     },
     // alpha.43 — family-level conventions for the `deepseek-chat` family
     // (V4-Flash + future non-thinking-mode members). V4-Flash struggles
@@ -916,18 +937,22 @@ var PROFILES_RAW = [
     maxContextTokens: 1e6,
     maxOutputTokens: 384e3,
     maxTools: 16,
-    parallelToolCalls: false,
+    // alpha.96 — CORRECTED from `false`, which was AI-generated at greenfield
+    // (2026-04-25) and never re-checked. Measured 2026-08-28 direct to
+    // api.deepseek.com, no kgauto in the path, 6 tools offered, 3 trials:
+    // 3 parallel tool calls returned, 3/3. L-040's original evidence said
+    // "7-8 calls per step" — which was never a claim that only one tool may be
+    // offered. See the note in lower.ts:lowerDeepSeek.
+    parallelToolCalls: true,
     structuredOutput: "native",
     systemPromptMode: "inline",
     streaming: true,
-    cliffs: [
-      {
-        metric: "tool_count",
-        threshold: 1,
-        action: "drop_to_top_relevant",
-        reason: "Sequential tool calls only \u2014 L-040"
-      }
-    ],
+    // alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
+    // It restated `parallelToolCalls: false` as a threshold and was falsified by
+    // the same measurement. It is not kept-and-flagged (the usual lifecycle)
+    // because it was never a real failure mode: it was a second enforcement of a
+    // wrong belief, and keeping it active would keep amputating tools.
+    cliffs: [],
     // Verified against the live DeepSeek pricing page 2026-06-22: the 75%-off
     // launch "promo" ($0.435/$0.87) did NOT revert on 2026-05-31 — it became the
     // standing price (docs show $0.435 in / $0.87 out cache-miss, $0.003625
@@ -945,7 +970,12 @@ var PROFILES_RAW = [
       { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
-    weaknesses: ["parallel_tools", "large_tool_sets"],
+    // alpha.96 — 'parallel_tools' REMOVED: measurably false (2026-08-28, direct
+    // to api.deepseek.com, 6 tools, 3 trials, 3 parallel calls, both models).
+    // 'large_tool_sets' KEPT and explicitly unverified — the probe offered 6 of
+    // a declared maxTools 16, so it says nothing about behaviour at 16. Not
+    // falsified is not the same as confirmed.
+    weaknesses: ["large_tool_sets"],
     // alpha.47 — explicit slow override. Measured 47722ms on the alpha.46
     // shadow probe (2026-06-03) — it's an extended-thinking reasoner, slowest
     // of the served set. Tag derivation would say 'medium'; the measurement says
@@ -964,14 +994,24 @@ var PROFILES_RAW = [
       summarize: 7,
       extract: 6,
       transform: 6,
-      hunt: 4
-      // sequential tools — same as V4-Flash
+      // alpha.96 — was 4. Its ONLY written justification was "sequential tool
+      // calls only — L-040", which is falsified. A score derived from a false
+      // premise cannot be kept, and re-guessing a penalty would just relaunder
+      // the same belief. Set to the floor (6) = "no evidence of deficiency",
+      // NOT an endorsement. This is a JUDGMENT, not a measurement, and it is
+      // the honest direction of the two: the only hunt-shaped evidence we have
+      // is positive (a research-agent fixture, 3 of 3 demanded tools correctly
+      // selected from a 6-tool menu, 3/3 trials, both models). Replace with a
+      // real number from the golden-eval hunt corpus — that is what it is for.
+      hunt: 6
     },
     // alpha.43 — family-level conventions for the `deepseek-reasoner` family
     // (V4-Pro is the family rep; future thinking-mode members inherit).
     //
-    // The decisive-archetype suffix is the load-bearing piece: 2026-05-28
-    // first real V4-Pro probe on tt-intel/classify (exclusion-finding ID 20)
+    // The decisive-archetype suffix is the load-bearing piece. ⚠️ alpha.96:
+    // its evidence was purged as poison on 2026-05-29 (s51) — retained as a
+    // hypothesis, not a finding. Original claim, for the record: 2026-05-28
+    // first real V4-Pro probe on tt-intel/classify
     // showed 8/10 judge rationales citing "candidate hedges and fails to
     // commit to a single classification" or "candidate refuses to pick
     // among the labels." Without the forcing-function suffix, every
@@ -992,24 +1032,35 @@ var PROFILES_RAW = [
       {
         archetype: "classify",
         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.",
-        reason: "Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. Evidence: 2026-05-28 V4-Pro probe on tt-intel/classify (exclusion-finding ID 20, brain probe_outcomes); 8 of 10 judge rationales cited hedging or refusal-to-commit. Forcing-function suffix converts reasoner output to decisive labels."
+        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.'
       },
       {
         archetype: "extract",
         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.",
-        reason: "Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. Evidence: same family failure shape as classify (exclusion-finding ID 20, 2026-05-28). Forcing-function suffix asks for commitment instead of epistemic-humility hedging."
+        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.'
       },
       {
         archetype: "ask",
         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.",
-        reason: "Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. Evidence: same family failure shape as classify (exclusion-finding ID 20, 2026-05-28). Forcing-function suffix asks for commitment instead of epistemic-humility hedging."
-      },
-      {
-        archetype: "hunt",
-        whenToolCountAtLeast: 3,
-        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.",
-        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."
+        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.'
       }
+      // alpha.96 — the `hunt` convention was DELETED, not corrected.
+      // It warned that "hunt archetypes with 3+ parallel tools lose substantial
+      // throughput" and steered callers to gemini-flash. Measured false
+      // 2026-08-28: 6 tools offered, 3 of 3 demanded served in parallel, 3/3
+      // trials, both V4 models, direct to the provider.
+      //
+      // Its cited evidence did not say what it claimed. exclusion-finding ID 20
+      // is `intent_archetype: 'classify'` (not hunt), `verdict:
+      // 'recommend-probe'` (i.e. NOBODY PROBED), `confidence: 'low'`, its
+      // message reads "deepseek-v4-pro has zero calls", it was later
+      // `declined`, and its own `parallel_tool_hint: false` was this profile's
+      // own belief round-tripped back as input. A no-data finding on the wrong
+      // archetype was paraphrased into "probe evidence confirms".
+      //
+      // Deleted rather than re-scoped because a cliffWarning is
+      // capability-REDUCING (it steers traffic away from a model) and nothing
+      // evidenced remains. Re-add only from a measured hunt eval.
     ]
   },
   // ── OpenAI ──
@@ -1908,6 +1959,19 @@ var PROFILES_RAW = [
     id: "glm-5.2",
     verifiedAgainstDocs: "2026-07-18",
     provider: "zai",
+    probedLimits: [
+      {
+        metric: "tool_count",
+        testedAt: 6,
+        demanded: 3,
+        served: [3, 3, 3],
+        outcome: "held",
+        trials: 3,
+        lastProbed: "2026-08-29",
+        route: "openrouter:baidu",
+        claim: "Measured 2026-08-29 via OpenRouter (served-by Baidu): 6 tools offered, 3 of 3 demanded served in parallel, 3/3 trials. Route-scoped \u2014 direct z.ai is unmeasured (no key held)."
+      }
+    ],
     status: "preview",
     maxContextTokens: 1048576,
     maxOutputTokens: 131072,
@@ -1918,7 +1982,8 @@ var PROFILES_RAW = [
     // NOTHING about multiple tool_calls per response. Conservative false so
     // hunt-shaped ranking doesn't over-promise; flip on doc or probe
     // evidence (L-081: undocumented ≠ true).
-    parallelToolCalls: false,
+    parallelToolCalls: true,
+    // alpha.96 — see weaknesses note; was an unevidenced blanket false
     structuredOutput: "native",
     systemPromptMode: "inline",
     streaming: true,
@@ -1941,7 +2006,17 @@ var PROFILES_RAW = [
       { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
     ],
     strengths: ["cost", "1m_context", "code", "tool_use", "json_output", "reasoning"],
-    weaknesses: ["parallel_tools"],
+    // alpha.96 — 'parallel_tools' REMOVED. Measured 2026-08-29 via OpenRouter
+    // (served-by Baidu), 6 tools offered, 3 of 3 demanded served in PARALLEL,
+    // 3/3 trials. ROUTE-SCOPED: per the alpha.92 execute-leg contract there is
+    // no cross-route verdict transfer, and direct z.ai is UNMEASURED (we hold
+    // no zai key). But `parallelToolCalls: false` was a blanket claim with no
+    // route qualification and no evidence at all, and it is capability-
+    // REDUCING — it drives requires-adapter/reject. An unsubstantiated reducer
+    // must not reduce. Contradicted on the one route we can reach is enough to
+    // stop enforcing it; it is not enough to claim the opposite, which is why
+    // probedLimits below carries the route.
+    weaknesses: [],
     notes: "GLM-5.2 (Z.ai, 2026-06): agentic-coding flagship. 1,048,576 ctx / 131,072 max out. $1.40/$4.40 per 1M; cached input $0.26/M (automatic caching, no marker; storage limited-time free as of 2026-07-18). Thinking on by default (`thinking.type`), `reasoning_effort` max..none. Parallel tool calls UNDOCUMENTED on the hosted API \u2014 profile says false until doc or probe evidence. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
     // Starter hypotheses — verify with telemetry/probes; NO brain evidence
     // yet. Anchored on the vendor's agentic-coding positioning at a
@@ -1959,8 +2034,11 @@ var PROFILES_RAW = [
       extract: 6,
       critique: 6,
       classify: 6,
-      hunt: 5
-      // tools work; parallel unproven → discounted
+      // alpha.96 — was 5, discounted for "parallel unproven". Parallel is no
+      // longer unproven on the reachable route (3/3, 2026-08-29). The discount
+      // had no other stated basis, so it is withdrawn to the neutral floor.
+      // Judgment, not measurement — same standard as the DeepSeek correction.
+      hunt: 6
     }
   },
   // ── Moonshot (Kimi) ──

package/dist/glassbox-routes/index.mjs

@@ -1,10 +1,10 @@
 import {
   ARCHETYPE_FLOOR_DEFAULT,
   getDefaultFallbackChain
-} from "../chunk-BJD75WCL.mjs";
+} from "../chunk-77AHSX6Y.mjs";
 import {
   tryGetProfile
-} from "../chunk-D3BVYXNA.mjs";
+} from "../chunk-HHIXHL4O.mjs";
 import {
   subscribe,
   subscribeApp

package/dist/index.d.mts

@@ -1241,7 +1241,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.95";
+declare const LIBRARY_VERSION = "2.0.0-alpha.97";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.

package/dist/index.d.ts

@@ -1241,7 +1241,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.95";
+declare const LIBRARY_VERSION = "2.0.0-alpha.97";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.