@warmdrift/kgauto-compiler 2.0.0-alpha.46 → 2.0.0-alpha.48

package/README.md

@@ -89,6 +89,37 @@ await record({
 });
 ```
 
+### `probeShadow()` — measure a swap candidate from an AI-SDK route (alpha.48)
+
+`call()` runs a full-IR shadow probe internally (`shadowProbe` option). AI-SDK consumers who drive `streamText()` themselves never call `call()`, so they reach the same probe via the standalone `probeShadow()` export — fire it from `onFinish`, passing the served leg you already have, wrapped in `waitUntil()`/`after()` so it never blocks the response.
+
+```ts
+import { probeShadow, configureBrain } from '@warmdrift/kgauto-compiler';
+import { after } from 'next/server'; // or the platform's waitUntil()
+
+configureBrain({ endpoint: 'https://your-app.com/api/kgauto/v2', apiKey: '...' });
+
+const t0 = Date.now();
+const result = streamText({
+  model, system, messages,
+  onFinish: ({ text, usage }) => {
+    after(() => probeShadow(ir, {
+      served: {
+        model: servedModelId,
+        responseText: text,
+        tokensIn: usage.promptTokens,
+        tokensOut: usage.completionTokens,
+        latencyMs: Date.now() - t0,
+      },
+      candidates: ['deepseek-v4-pro'], // what you'd consider swapping to
+      sampleRate: 0.05,                // pin 1 for a dogfood window
+    }));
+  },
+});
+```
+
+kgauto runs each candidate on the **same IR** and persists a `probe_outcomes` row (`replay_source='inline-full-ir'`, `prompt_fidelity=1.0`, both latency legs). Phase 1 is `judge:'off'` — rows accumulate for an offline verdict rollup; inline Opus judging is Phase 2. Fire-and-forget: never throws into your `onFinish`.
+
 ## Structured advisories API (alpha.29+)
 
 kgauto's compile-time advisories (`caching-off-on-claude`, `tool-bloat`,

package/dist/{chunk-ZHUD3I52.mjs → chunk-PKOFXEB3.mjs}

@@ -1,4 +1,15 @@
 // src/profiles.ts
+var LATENCY_TIER_MS = {
+  fast: 4e3,
+  medium: 11e3,
+  slow: 24e3
+};
+function latencyTierOf(profile) {
+  if (profile.latencyTier) return profile.latencyTier;
+  if (profile.weaknesses.includes("latency")) return "slow";
+  if (profile.strengths.includes("speed")) return "fast";
+  return "medium";
+}
 var ANTHROPIC_LOWERING_BASE = {
   system: { mode: "inline" },
   cache: {
@@ -545,6 +556,14 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "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×
+    // gemini-2.5-flash on PB's synchronous /api/analyze path. The 'flash' name
+    // is DeepSeek's, not a speed promise. This is the row that, scoring 0.85
+    // baseQuality (it carries 'reasoning') with no latency counterweight,
+    // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
+    latencyTier: "slow",
     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.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
@@ -612,6 +631,11 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "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
+    // otherwise.
+    latencyTier: "slow",
     notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.
@@ -1264,6 +1288,8 @@ function profilesByProvider(provider) {
 }
 
 export {
+  LATENCY_TIER_MS,
+  latencyTierOf,
   ALIASES,
   _setProfileBrainHook,
   getProfile,

package/dist/{chunk-6MSJQCAP.mjs → chunk-SAWTKMD4.mjs}

@@ -1,6 +1,6 @@
 import {
   tryGetProfile
-} from "./chunk-ZHUD3I52.mjs";
+} from "./chunk-PKOFXEB3.mjs";
 
 // src/brain-query.ts
 var FRESH_SNAPSHOT = {

package/dist/glassbox-routes/index.js

@@ -610,6 +610,14 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "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×
+    // gemini-2.5-flash on PB's synchronous /api/analyze path. The 'flash' name
+    // is DeepSeek's, not a speed promise. This is the row that, scoring 0.85
+    // baseQuality (it carries 'reasoning') with no latency counterweight,
+    // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
+    latencyTier: "slow",
     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.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
@@ -677,6 +685,11 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "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
+    // otherwise.
+    latencyTier: "slow",
     notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.

package/dist/glassbox-routes/index.mjs

@@ -1,10 +1,10 @@
 import {
   ARCHETYPE_FLOOR_DEFAULT,
   getDefaultFallbackChain
-} from "../chunk-6MSJQCAP.mjs";
+} from "../chunk-SAWTKMD4.mjs";
 import {
   tryGetProfile
-} from "../chunk-ZHUD3I52.mjs";
+} from "../chunk-PKOFXEB3.mjs";
 import {
   subscribe,
   subscribeApp

package/dist/index.d.mts

@@ -1,7 +1,7 @@
 import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, B as BestPracticeAdvisory, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-BBKgsUX7.mjs';
 export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, r as MutationApplied, s as NormalizedTokens, t as OutcomeKind, u as PerAxisMetricsByModel, v as PromptSection, w as SectionKind, x as ShadowProbeConfig, T as ToolCall, y as ToolDefinition } from './ir-BBKgsUX7.mjs';
 import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
-export { ALIASES, CacheStrategy, CliffRule, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, profilesByProvider, tryGetProfile } from './profiles.mjs';
+export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, OutputMode, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.mjs';
 
@@ -102,6 +102,70 @@ declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promi
  * exhausted without success.
  */
 declare function call(ir: PromptIR, opts?: CallOptions): Promise<CallResult>;
+/** Served-side result the AI-SDK consumer already holds from streamText. */
+interface ProbeShadowServed {
+    /** Model id that actually served the user (becomes `current_model` on the row). */
+    model: string;
+    /** The served response text (stored as a preview; never the prompt). */
+    responseText: string;
+    /** Served prompt tokens (AI SDK `usage.promptTokens`). */
+    tokensIn: number;
+    /** Served completion tokens (AI SDK `usage.completionTokens`). */
+    tokensOut: number;
+    /** The user's actual wait in ms (alpha.46 served-latency axis). */
+    latencyMs: number;
+}
+/** Options for the standalone {@link probeShadow} export. */
+interface ProbeShadowOptions {
+    /** The served leg the consumer supplies; kgauto runs the candidate leg. */
+    served: ProbeShadowServed;
+    /** Model id(s)/family alias(es) to shadow-test against the served model. */
+    candidates: string | string[];
+    /** Probability [0,1] a given call fires a probe. Default 0.05; pin 1 for dogfood. */
+    sampleRate?: number;
+    /** Phase 1 = 'off' (offline rollup). 'opus' inline judge is Phase 2. Default 'off'. */
+    judge?: 'off' | 'opus';
+    /** Override API keys (defaults: process.env). Gates candidate reachability. */
+    apiKeys?: ApiKeys;
+    /** Override fetch (for tests / custom transport). */
+    fetchImpl?: typeof fetch;
+    /** Provider-specific request fields shallow-merged into the candidate request. */
+    providerOverrides?: ProviderOverrides;
+    /** Forwarded to compile() for the candidate run (match the served-side compile). */
+    policy?: CompilePolicy;
+    toolRelevanceThreshold?: number;
+    compressHistoryAfter?: number;
+}
+/**
+ * s54 — standalone full-IR shadow probe for AI-SDK consumers (tt-intel hunt, IC
+ * chat) that own their own streamText round-trip and never call call(). Fire it
+ * from the streamText `onFinish` callback, wrapped in the platform's
+ * `waitUntil()` / `after()`, so it runs after the response is flushed — no
+ * blocking, no `sync:true`, no teardown race (cleaner than call()'s in-process
+ * fire-and-forget for Edge runtimes; L-086).
+ *
+ * Runs the SAME candidate loop as the call()-inline probe ({@link
+ * runProbeCandidates}): same gates (sample, self-skip, G4 judge-self,
+ * reachability, G6), same fidelity=1.0 full-IR rows, same alpha.46 latency
+ * capture. The consumer supplies the served leg it already has; kgauto runs the
+ * candidate(s) and writes the `probe_outcomes` row.
+ *
+ * Requires `configureBrain()` to have been called (same as call()).
+ * Fire-and-forget: never throws into the caller.
+ *
+ * @example
+ * const result = streamText({ model, system, messages,
+ *   onFinish: ({ text, usage }) => {
+ *     after(() => probeShadow(ir, {
+ *       served: { model: servedModelId, responseText: text,
+ *                 tokensIn: usage.promptTokens, tokensOut: usage.completionTokens,
+ *                 latencyMs: Date.now() - t0 },
+ *       candidates: ['deepseek-v4-pro'], sampleRate: 0.05,
+ *     }));
+ *   },
+ * });
+ */
+declare function probeShadow(ir: PromptIR, opts: ProbeShadowOptions): Promise<void>;
 
 /**
  * alpha.33 — AI-SDK streamText helpers.
@@ -2195,4 +2259,4 @@ declare function markPromoteReadyHandled(opts: MarkPromoteReadyHandledOptions):
  */
 declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
 
-export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, 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, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, SystemModelMessage, TRANSLATOR_FLOOR, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, buildLLMJudge, buildShadowProbeRow, call, clearBrain, compile, configureBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, profileToRow, readBrainReadEnv, record, recordOutcome, recordShadowProbe, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };
+export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, 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, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, SystemModelMessage, TRANSLATOR_FLOOR, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, buildLLMJudge, buildShadowProbeRow, call, clearBrain, compile, configureBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, probeShadow, profileToRow, readBrainReadEnv, record, recordOutcome, recordShadowProbe, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };

package/dist/index.d.ts

@@ -1,7 +1,7 @@
 import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, B as BestPracticeAdvisory, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-D3n-pBYI.js';
 export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, r as MutationApplied, s as NormalizedTokens, t as OutcomeKind, u as PerAxisMetricsByModel, v as PromptSection, w as SectionKind, x as ShadowProbeConfig, T as ToolCall, y as ToolDefinition } from './ir-D3n-pBYI.js';
 import { ModelProfile, ArchetypeConvention } from './profiles.js';
-export { ALIASES, CacheStrategy, CliffRule, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, profilesByProvider, tryGetProfile } from './profiles.js';
+export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.js';
 import { IntentArchetypeName } from './dialect.js';
 export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, OutputMode, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.js';
 
@@ -102,6 +102,70 @@ declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promi
  * exhausted without success.
  */
 declare function call(ir: PromptIR, opts?: CallOptions): Promise<CallResult>;
+/** Served-side result the AI-SDK consumer already holds from streamText. */
+interface ProbeShadowServed {
+    /** Model id that actually served the user (becomes `current_model` on the row). */
+    model: string;
+    /** The served response text (stored as a preview; never the prompt). */
+    responseText: string;
+    /** Served prompt tokens (AI SDK `usage.promptTokens`). */
+    tokensIn: number;
+    /** Served completion tokens (AI SDK `usage.completionTokens`). */
+    tokensOut: number;
+    /** The user's actual wait in ms (alpha.46 served-latency axis). */
+    latencyMs: number;
+}
+/** Options for the standalone {@link probeShadow} export. */
+interface ProbeShadowOptions {
+    /** The served leg the consumer supplies; kgauto runs the candidate leg. */
+    served: ProbeShadowServed;
+    /** Model id(s)/family alias(es) to shadow-test against the served model. */
+    candidates: string | string[];
+    /** Probability [0,1] a given call fires a probe. Default 0.05; pin 1 for dogfood. */
+    sampleRate?: number;
+    /** Phase 1 = 'off' (offline rollup). 'opus' inline judge is Phase 2. Default 'off'. */
+    judge?: 'off' | 'opus';
+    /** Override API keys (defaults: process.env). Gates candidate reachability. */
+    apiKeys?: ApiKeys;
+    /** Override fetch (for tests / custom transport). */
+    fetchImpl?: typeof fetch;
+    /** Provider-specific request fields shallow-merged into the candidate request. */
+    providerOverrides?: ProviderOverrides;
+    /** Forwarded to compile() for the candidate run (match the served-side compile). */
+    policy?: CompilePolicy;
+    toolRelevanceThreshold?: number;
+    compressHistoryAfter?: number;
+}
+/**
+ * s54 — standalone full-IR shadow probe for AI-SDK consumers (tt-intel hunt, IC
+ * chat) that own their own streamText round-trip and never call call(). Fire it
+ * from the streamText `onFinish` callback, wrapped in the platform's
+ * `waitUntil()` / `after()`, so it runs after the response is flushed — no
+ * blocking, no `sync:true`, no teardown race (cleaner than call()'s in-process
+ * fire-and-forget for Edge runtimes; L-086).
+ *
+ * Runs the SAME candidate loop as the call()-inline probe ({@link
+ * runProbeCandidates}): same gates (sample, self-skip, G4 judge-self,
+ * reachability, G6), same fidelity=1.0 full-IR rows, same alpha.46 latency
+ * capture. The consumer supplies the served leg it already has; kgauto runs the
+ * candidate(s) and writes the `probe_outcomes` row.
+ *
+ * Requires `configureBrain()` to have been called (same as call()).
+ * Fire-and-forget: never throws into the caller.
+ *
+ * @example
+ * const result = streamText({ model, system, messages,
+ *   onFinish: ({ text, usage }) => {
+ *     after(() => probeShadow(ir, {
+ *       served: { model: servedModelId, responseText: text,
+ *                 tokensIn: usage.promptTokens, tokensOut: usage.completionTokens,
+ *                 latencyMs: Date.now() - t0 },
+ *       candidates: ['deepseek-v4-pro'], sampleRate: 0.05,
+ *     }));
+ *   },
+ * });
+ */
+declare function probeShadow(ir: PromptIR, opts: ProbeShadowOptions): Promise<void>;
 
 /**
  * alpha.33 — AI-SDK streamText helpers.
@@ -2195,4 +2259,4 @@ declare function markPromoteReadyHandled(opts: MarkPromoteReadyHandledOptions):
  */
 declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
 
-export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, 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, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, SystemModelMessage, TRANSLATOR_FLOOR, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, buildLLMJudge, buildShadowProbeRow, call, clearBrain, compile, configureBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, profileToRow, readBrainReadEnv, record, recordOutcome, recordShadowProbe, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };
+export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, 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, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, SystemModelMessage, TRANSLATOR_FLOOR, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, buildLLMJudge, buildShadowProbeRow, call, clearBrain, compile, configureBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, probeShadow, profileToRow, readBrainReadEnv, record, recordOutcome, recordShadowProbe, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };

package/dist/index.js

@@ -30,6 +30,7 @@ __export(index_exports, {
   DIALECT_VERSION: () => DIALECT_VERSION,
   FamilyResolutionError: () => FamilyResolutionError,
   INTENT_ARCHETYPES: () => INTENT_ARCHETYPES,
+  LATENCY_TIER_MS: () => LATENCY_TIER_MS,
   MEASURED_GROUNDING_MIN_N: () => MEASURED_GROUNDING_MIN_N,
   PRODUCER_OWNED_RULE_CODES: () => PRODUCER_OWNED_RULE_CODES,
   PROVIDER_ENV_KEYS: () => PROVIDER_ENV_KEYS,
@@ -76,6 +77,7 @@ __export(index_exports, {
   isExclusionFindingsBrainActive: () => isExclusionFindingsBrainActive,
   isModelReachable: () => isModelReachable,
   isProviderReachable: () => isProviderReachable,
+  latencyTierOf: () => latencyTierOf,
   learningKey: () => learningKey,
   loadAliasesFromBrain: () => loadAliasesFromBrain,
   loadArchetypePerfFromBrain: () => loadArchetypePerfFromBrain,
@@ -86,6 +88,7 @@ __export(index_exports, {
   markAdvisoryResolved: () => markAdvisoryResolved,
   markExclusionFindingHandled: () => markExclusionFindingHandled,
   markPromoteReadyHandled: () => markPromoteReadyHandled,
+  probeShadow: () => probeShadow,
   profileToRow: () => profileToRow,
   profilesByProvider: () => profilesByProvider,
   readBrainReadEnv: () => readBrainReadEnv,
@@ -274,6 +277,17 @@ function mapPerAxisMetrics(raw, fallbackAppId, fallbackArchetype, fallbackModel,
 }
 
 // src/profiles.ts
+var LATENCY_TIER_MS = {
+  fast: 4e3,
+  medium: 11e3,
+  slow: 24e3
+};
+function latencyTierOf(profile) {
+  if (profile.latencyTier) return profile.latencyTier;
+  if (profile.weaknesses.includes("latency")) return "slow";
+  if (profile.strengths.includes("speed")) return "fast";
+  return "medium";
+}
 var ANTHROPIC_LOWERING_BASE = {
   system: { mode: "inline" },
   cache: {
@@ -820,6 +834,14 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "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×
+    // gemini-2.5-flash on PB's synchronous /api/analyze path. The 'flash' name
+    // is DeepSeek's, not a speed promise. This is the row that, scoring 0.85
+    // baseQuality (it carries 'reasoning') with no latency counterweight,
+    // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
+    latencyTier: "slow",
     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.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
@@ -887,6 +909,11 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "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
+    // otherwise.
+    latencyTier: "slow",
     notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.
@@ -2360,6 +2387,8 @@ function passApplyCliffs(ir, profile, estimatedInputTokens) {
   }
   return { value: { ir: nextIR, loweringHints: hints }, mutations };
 }
+var LATENCY_OVERAGE_WEIGHT = 0.6;
+var LATENCY_PENALTY_CAP = 1;
 function passScoreTargets(ir, opts) {
   const constraints = ir.constraints ?? {};
   const policy = opts.policy ?? {};
@@ -2415,7 +2444,16 @@ function passScoreTargets(ir, opts) {
     const callerOrderBoost = (modelIds.length - modelIds.indexOf(modelId)) * 0.1;
     const costPenalty = estimatedCostUsd * 5;
     const preferredBoost = preferredSet.has(modelId) ? 0.5 : 0;
-    const rank = qualityScore + callerOrderBoost - costPenalty - reasons.length * 10 + preferredBoost;
+    let latencyPenalty = 0;
+    const maxLatencyMs = constraints.maxLatencyMs;
+    if (typeof maxLatencyMs === "number" && maxLatencyMs > 0) {
+      const tierMs = LATENCY_TIER_MS[latencyTierOf(profile)];
+      if (tierMs > maxLatencyMs) {
+        const overage = (tierMs - maxLatencyMs) / maxLatencyMs;
+        latencyPenalty = Math.min(LATENCY_OVERAGE_WEIGHT * overage, LATENCY_PENALTY_CAP);
+      }
+    }
+    const rank = qualityScore + callerOrderBoost - costPenalty - reasons.length * 10 + preferredBoost - latencyPenalty;
     scores.push({
       modelId,
       estimatedCostUsd,
@@ -2448,6 +2486,15 @@ function passScoreTargets(ir, opts) {
         description: `Model ${modelId} rank boosted by CompilePolicy.preferredModels`
       });
     }
+    if (latencyPenalty > 0) {
+      const tier = latencyTierOf(profile);
+      policyMutations.push({
+        id: `latency-downrank-${modelId}`,
+        source: "latency_guard",
+        passName: "score_targets",
+        description: `Model ${modelId} rank down ${latencyPenalty.toFixed(2)} \u2014 latency tier '${tier}' (~${LATENCY_TIER_MS[tier]}ms) exceeds constraints.maxLatencyMs (${maxLatencyMs}ms)`
+      });
+    }
   }
   return { value: scores, mutations: policyMutations };
 }
@@ -6111,16 +6158,15 @@ function hashForProbe(s) {
   for (let i = 0; i < s.length; i++) h = (h << 5) - h + s.charCodeAt(i) | 0;
   return `h${(h >>> 0).toString(16)}`;
 }
-async function runShadowProbe(args) {
+async function runProbeCandidates(args) {
   try {
-    const cfg = args.opts.shadowProbe;
-    if (!shouldSampleProbe(cfg.sampleRate ?? 0.05)) return;
-    const candidates = normalizeProbeCandidates(cfg.candidates);
+    if (!shouldSampleProbe(args.cfg.sampleRate ?? 0.05)) return;
+    const candidates = normalizeProbeCandidates(args.cfg.candidates);
     const promptHash = hashForProbe(args.ir.currentTurn?.content ?? "");
     for (const candidate of candidates) {
       if (candidate === args.servedModel) continue;
       if (deriveFamilyFromModelId(candidate) === "claude-opus") continue;
-      if (!isModelReachable(candidate, { apiKeys: args.opts.apiKeys })) continue;
+      if (!isModelReachable(candidate, { apiKeys: args.exec.apiKeys })) continue;
       try {
         const candCompile = compileAndRegister(
           {
@@ -6128,13 +6174,13 @@ async function runShadowProbe(args) {
             models: args.ir.models.includes(candidate) ? args.ir.models : [candidate, ...args.ir.models],
             constraints: { ...args.ir.constraints ?? {}, forceModel: candidate }
           },
-          args.opts
+          args.exec
         );
         const candStart = Date.now();
         const exec = await execute(candCompile.request, {
-          apiKeys: args.opts.apiKeys,
-          fetchImpl: args.opts.fetchImpl,
-          providerOverrides: args.opts.providerOverrides
+          apiKeys: args.exec.apiKeys,
+          fetchImpl: args.exec.fetchImpl,
+          providerOverrides: args.exec.providerOverrides
         });
         const candidateLatencyMs = Date.now() - candStart;
         if (!exec.ok) continue;
@@ -6163,6 +6209,54 @@ async function runShadowProbe(args) {
   } catch {
   }
 }
+async function runShadowProbe(args) {
+  const cfg = args.opts.shadowProbe;
+  if (!cfg) return;
+  await runProbeCandidates({
+    ir: args.ir,
+    cfg,
+    servedModel: args.servedModel,
+    servedResponse: args.servedResponse,
+    servedLatencyMs: args.servedLatencyMs,
+    // CallOptions is a structural superset of ProbeExecOpts.
+    exec: args.opts
+  });
+}
+async function probeShadow(ir, opts) {
+  try {
+    const servedResponse = {
+      text: opts.served.responseText,
+      structuredOutput: null,
+      toolCalls: [],
+      tokens: {
+        input: opts.served.tokensIn,
+        output: opts.served.tokensOut,
+        total: opts.served.tokensIn + opts.served.tokensOut
+      },
+      raw: null
+    };
+    await runProbeCandidates({
+      ir,
+      cfg: {
+        candidates: opts.candidates,
+        sampleRate: opts.sampleRate,
+        judge: opts.judge ?? "off"
+      },
+      servedModel: opts.served.model,
+      servedResponse,
+      servedLatencyMs: opts.served.latencyMs,
+      exec: {
+        policy: opts.policy,
+        toolRelevanceThreshold: opts.toolRelevanceThreshold,
+        compressHistoryAfter: opts.compressHistoryAfter,
+        apiKeys: opts.apiKeys,
+        fetchImpl: opts.fetchImpl,
+        providerOverrides: opts.providerOverrides
+      }
+    });
+  } catch {
+  }
+}
 function validateStructuredContract(exec, ir) {
   if (!ir.constraints?.structuredOutput) {
     return { ok: true, response: exec.response };
@@ -6691,6 +6785,7 @@ function compile2(ir, opts) {
   DIALECT_VERSION,
   FamilyResolutionError,
   INTENT_ARCHETYPES,
+  LATENCY_TIER_MS,
   MEASURED_GROUNDING_MIN_N,
   PRODUCER_OWNED_RULE_CODES,
   PROVIDER_ENV_KEYS,
@@ -6737,6 +6832,7 @@ function compile2(ir, opts) {
   isExclusionFindingsBrainActive,
   isModelReachable,
   isProviderReachable,
+  latencyTierOf,
   learningKey,
   loadAliasesFromBrain,
   loadArchetypePerfFromBrain,
@@ -6747,6 +6843,7 @@ function compile2(ir, opts) {
   markAdvisoryResolved,
   markExclusionFindingHandled,
   markPromoteReadyHandled,
+  probeShadow,
   profileToRow,
   profilesByProvider,
   readBrainReadEnv,

package/dist/index.mjs

@@ -33,16 +33,18 @@ import {
   loadChainsFromBrain,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-6MSJQCAP.mjs";
+} from "./chunk-SAWTKMD4.mjs";
 import {
   ALIASES,
+  LATENCY_TIER_MS,
   _setProfileBrainHook,
   allProfiles,
   allProfilesRaw,
   getProfile,
+  latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-ZHUD3I52.mjs";
+} from "./chunk-PKOFXEB3.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -710,6 +712,8 @@ function passApplyCliffs(ir, profile, estimatedInputTokens) {
   }
   return { value: { ir: nextIR, loweringHints: hints }, mutations };
 }
+var LATENCY_OVERAGE_WEIGHT = 0.6;
+var LATENCY_PENALTY_CAP = 1;
 function passScoreTargets(ir, opts) {
   const constraints = ir.constraints ?? {};
   const policy = opts.policy ?? {};
@@ -765,7 +769,16 @@ function passScoreTargets(ir, opts) {
     const callerOrderBoost = (modelIds.length - modelIds.indexOf(modelId)) * 0.1;
     const costPenalty = estimatedCostUsd * 5;
     const preferredBoost = preferredSet.has(modelId) ? 0.5 : 0;
-    const rank = qualityScore + callerOrderBoost - costPenalty - reasons.length * 10 + preferredBoost;
+    let latencyPenalty = 0;
+    const maxLatencyMs = constraints.maxLatencyMs;
+    if (typeof maxLatencyMs === "number" && maxLatencyMs > 0) {
+      const tierMs = LATENCY_TIER_MS[latencyTierOf(profile)];
+      if (tierMs > maxLatencyMs) {
+        const overage = (tierMs - maxLatencyMs) / maxLatencyMs;
+        latencyPenalty = Math.min(LATENCY_OVERAGE_WEIGHT * overage, LATENCY_PENALTY_CAP);
+      }
+    }
+    const rank = qualityScore + callerOrderBoost - costPenalty - reasons.length * 10 + preferredBoost - latencyPenalty;
     scores.push({
       modelId,
       estimatedCostUsd,
@@ -798,6 +811,15 @@ function passScoreTargets(ir, opts) {
         description: `Model ${modelId} rank boosted by CompilePolicy.preferredModels`
       });
     }
+    if (latencyPenalty > 0) {
+      const tier = latencyTierOf(profile);
+      policyMutations.push({
+        id: `latency-downrank-${modelId}`,
+        source: "latency_guard",
+        passName: "score_targets",
+        description: `Model ${modelId} rank down ${latencyPenalty.toFixed(2)} \u2014 latency tier '${tier}' (~${LATENCY_TIER_MS[tier]}ms) exceeds constraints.maxLatencyMs (${maxLatencyMs}ms)`
+      });
+    }
   }
   return { value: scores, mutations: policyMutations };
 }
@@ -3746,16 +3768,15 @@ function hashForProbe(s) {
   for (let i = 0; i < s.length; i++) h = (h << 5) - h + s.charCodeAt(i) | 0;
   return `h${(h >>> 0).toString(16)}`;
 }
-async function runShadowProbe(args) {
+async function runProbeCandidates(args) {
   try {
-    const cfg = args.opts.shadowProbe;
-    if (!shouldSampleProbe(cfg.sampleRate ?? 0.05)) return;
-    const candidates = normalizeProbeCandidates(cfg.candidates);
+    if (!shouldSampleProbe(args.cfg.sampleRate ?? 0.05)) return;
+    const candidates = normalizeProbeCandidates(args.cfg.candidates);
     const promptHash = hashForProbe(args.ir.currentTurn?.content ?? "");
     for (const candidate of candidates) {
       if (candidate === args.servedModel) continue;
       if (deriveFamilyFromModelId(candidate) === "claude-opus") continue;
-      if (!isModelReachable(candidate, { apiKeys: args.opts.apiKeys })) continue;
+      if (!isModelReachable(candidate, { apiKeys: args.exec.apiKeys })) continue;
       try {
         const candCompile = compileAndRegister(
           {
@@ -3763,13 +3784,13 @@ async function runShadowProbe(args) {
             models: args.ir.models.includes(candidate) ? args.ir.models : [candidate, ...args.ir.models],
             constraints: { ...args.ir.constraints ?? {}, forceModel: candidate }
           },
-          args.opts
+          args.exec
         );
         const candStart = Date.now();
         const exec = await execute(candCompile.request, {
-          apiKeys: args.opts.apiKeys,
-          fetchImpl: args.opts.fetchImpl,
-          providerOverrides: args.opts.providerOverrides
+          apiKeys: args.exec.apiKeys,
+          fetchImpl: args.exec.fetchImpl,
+          providerOverrides: args.exec.providerOverrides
         });
         const candidateLatencyMs = Date.now() - candStart;
         if (!exec.ok) continue;
@@ -3798,6 +3819,54 @@ async function runShadowProbe(args) {
   } catch {
   }
 }
+async function runShadowProbe(args) {
+  const cfg = args.opts.shadowProbe;
+  if (!cfg) return;
+  await runProbeCandidates({
+    ir: args.ir,
+    cfg,
+    servedModel: args.servedModel,
+    servedResponse: args.servedResponse,
+    servedLatencyMs: args.servedLatencyMs,
+    // CallOptions is a structural superset of ProbeExecOpts.
+    exec: args.opts
+  });
+}
+async function probeShadow(ir, opts) {
+  try {
+    const servedResponse = {
+      text: opts.served.responseText,
+      structuredOutput: null,
+      toolCalls: [],
+      tokens: {
+        input: opts.served.tokensIn,
+        output: opts.served.tokensOut,
+        total: opts.served.tokensIn + opts.served.tokensOut
+      },
+      raw: null
+    };
+    await runProbeCandidates({
+      ir,
+      cfg: {
+        candidates: opts.candidates,
+        sampleRate: opts.sampleRate,
+        judge: opts.judge ?? "off"
+      },
+      servedModel: opts.served.model,
+      servedResponse,
+      servedLatencyMs: opts.served.latencyMs,
+      exec: {
+        policy: opts.policy,
+        toolRelevanceThreshold: opts.toolRelevanceThreshold,
+        compressHistoryAfter: opts.compressHistoryAfter,
+        apiKeys: opts.apiKeys,
+        fetchImpl: opts.fetchImpl,
+        providerOverrides: opts.providerOverrides
+      }
+    });
+  } catch {
+  }
+}
 function validateStructuredContract(exec, ir) {
   if (!ir.constraints?.structuredOutput) {
     return { ok: true, response: exec.response };
@@ -4325,6 +4394,7 @@ export {
   DIALECT_VERSION,
   FamilyResolutionError,
   INTENT_ARCHETYPES,
+  LATENCY_TIER_MS,
   MEASURED_GROUNDING_MIN_N,
   PRODUCER_OWNED_RULE_CODES,
   PROVIDER_ENV_KEYS,
@@ -4371,6 +4441,7 @@ export {
   isExclusionFindingsBrainActive,
   isModelReachable,
   isProviderReachable,
+  latencyTierOf,
   learningKey,
   loadAliasesFromBrain,
   loadArchetypePerfFromBrain,
@@ -4381,6 +4452,7 @@ export {
   markAdvisoryResolved,
   markExclusionFindingHandled,
   markPromoteReadyHandled,
+  probeShadow,
   profileToRow,
   profilesByProvider,
   readBrainReadEnv,

package/dist/profiles.d.mts

@@ -145,6 +145,14 @@ interface LoweringSpec {
         default?: number | 'auto' | 'off';
     };
 }
+/**
+ * Coarse latency bucket for a model. alpha.47 — the third swap axis
+ * (cost / quality / SPEED). We bucket rather than store per-model ms because
+ * served latency varies with token count; honest precision is the tier, not a
+ * false-precise number. See {@link LATENCY_TIER_MS} for the representative ms
+ * each tier maps to when compared against `constraints.maxLatencyMs`.
+ */
+type LatencyTier = 'fast' | 'medium' | 'slow';
 interface ModelProfile {
     id: string;
     provider: Provider;
@@ -163,6 +171,16 @@ interface ModelProfile {
     recovery: RecoveryRule[];
     strengths: string[];
     weaknesses: string[];
+    /**
+     * alpha.47 — explicit latency bucket. OPTIONAL: when unset, `latencyTierOf`
+     * derives it from the tags above (`weaknesses` includes `'latency'` → slow,
+     * `strengths` includes `'speed'` → fast, else medium). Set this explicitly
+     * ONLY when measured evidence contradicts the tag derivation — e.g.
+     * `deepseek-v4-flash` carries no `'latency'` weakness yet measures ~20s
+     * (alpha.46 shadow-probe). Carry provenance in an inline comment when you
+     * override, same discipline as capability data (step zero / L-081).
+     */
+    latencyTier?: LatencyTier;
     notes?: string;
     verifiedAgainstDocs?: string;
     /**
@@ -236,6 +254,26 @@ interface ModelProfile {
      */
     archetypeConventions?: ArchetypeConvention[];
 }
+/**
+ * Representative p50 latency (ms) per tier. Coarse on purpose — used only to
+ * compare a model against `constraints.maxLatencyMs`, never reported as a
+ * per-model number. Grounded: alpha.46 shadow-probe measured
+ * `deepseek-v4-flash` at 20485ms and `deepseek-v4-pro` at 47722ms (2026-06-03),
+ * both bucket `slow`; gemini-2.5-flash / haiku-4-5 serve in single-digit
+ * seconds → `fast`; mid-tier (sonnet, gemini-pro) → `medium`. Phase 2 can swap
+ * these buckets for measured per-(archetype,model) p50 from
+ * `compile_outcomes.latency_ms` once enough served rows accumulate.
+ */
+declare const LATENCY_TIER_MS: Record<LatencyTier, number>;
+/**
+ * Resolve a model's latency tier. Explicit `profile.latencyTier` wins;
+ * otherwise derive from the tags already on the profile so we don't maintain a
+ * second source of truth that can silently disagree (L-073 family):
+ *   - `weaknesses` includes `'latency'` → `'slow'`
+ *   - `strengths` includes `'speed'`    → `'fast'`
+ *   - else                              → `'medium'`
+ */
+declare function latencyTierOf(profile: ModelProfile): LatencyTier;
 declare const ALIASES: Record<string, string>;
 interface ProfileBrainHook {
     getProfile?: (canonicalId: string) => ModelProfile | undefined;
@@ -251,4 +289,4 @@ declare function allProfiles(): readonly ModelProfile[];
 declare function allProfilesRaw(): readonly ModelProfile[];
 declare function profilesByProvider(provider: Provider): readonly ModelProfile[];
 
-export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffRule, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, allProfiles, allProfilesRaw, getProfile, profilesByProvider, tryGetProfile };
+export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffRule, LATENCY_TIER_MS, type LatencyTier, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, allProfiles, allProfilesRaw, getProfile, latencyTierOf, profilesByProvider, tryGetProfile };

package/dist/profiles.d.ts

@@ -145,6 +145,14 @@ interface LoweringSpec {
         default?: number | 'auto' | 'off';
     };
 }
+/**
+ * Coarse latency bucket for a model. alpha.47 — the third swap axis
+ * (cost / quality / SPEED). We bucket rather than store per-model ms because
+ * served latency varies with token count; honest precision is the tier, not a
+ * false-precise number. See {@link LATENCY_TIER_MS} for the representative ms
+ * each tier maps to when compared against `constraints.maxLatencyMs`.
+ */
+type LatencyTier = 'fast' | 'medium' | 'slow';
 interface ModelProfile {
     id: string;
     provider: Provider;
@@ -163,6 +171,16 @@ interface ModelProfile {
     recovery: RecoveryRule[];
     strengths: string[];
     weaknesses: string[];
+    /**
+     * alpha.47 — explicit latency bucket. OPTIONAL: when unset, `latencyTierOf`
+     * derives it from the tags above (`weaknesses` includes `'latency'` → slow,
+     * `strengths` includes `'speed'` → fast, else medium). Set this explicitly
+     * ONLY when measured evidence contradicts the tag derivation — e.g.
+     * `deepseek-v4-flash` carries no `'latency'` weakness yet measures ~20s
+     * (alpha.46 shadow-probe). Carry provenance in an inline comment when you
+     * override, same discipline as capability data (step zero / L-081).
+     */
+    latencyTier?: LatencyTier;
     notes?: string;
     verifiedAgainstDocs?: string;
     /**
@@ -236,6 +254,26 @@ interface ModelProfile {
      */
     archetypeConventions?: ArchetypeConvention[];
 }
+/**
+ * Representative p50 latency (ms) per tier. Coarse on purpose — used only to
+ * compare a model against `constraints.maxLatencyMs`, never reported as a
+ * per-model number. Grounded: alpha.46 shadow-probe measured
+ * `deepseek-v4-flash` at 20485ms and `deepseek-v4-pro` at 47722ms (2026-06-03),
+ * both bucket `slow`; gemini-2.5-flash / haiku-4-5 serve in single-digit
+ * seconds → `fast`; mid-tier (sonnet, gemini-pro) → `medium`. Phase 2 can swap
+ * these buckets for measured per-(archetype,model) p50 from
+ * `compile_outcomes.latency_ms` once enough served rows accumulate.
+ */
+declare const LATENCY_TIER_MS: Record<LatencyTier, number>;
+/**
+ * Resolve a model's latency tier. Explicit `profile.latencyTier` wins;
+ * otherwise derive from the tags already on the profile so we don't maintain a
+ * second source of truth that can silently disagree (L-073 family):
+ *   - `weaknesses` includes `'latency'` → `'slow'`
+ *   - `strengths` includes `'speed'`    → `'fast'`
+ *   - else                              → `'medium'`
+ */
+declare function latencyTierOf(profile: ModelProfile): LatencyTier;
 declare const ALIASES: Record<string, string>;
 interface ProfileBrainHook {
     getProfile?: (canonicalId: string) => ModelProfile | undefined;
@@ -251,4 +289,4 @@ declare function allProfiles(): readonly ModelProfile[];
 declare function allProfilesRaw(): readonly ModelProfile[];
 declare function profilesByProvider(provider: Provider): readonly ModelProfile[];
 
-export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffRule, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, allProfiles, allProfilesRaw, getProfile, profilesByProvider, tryGetProfile };
+export { ALIASES, type ArchetypeConvention, type CacheStrategy, type CliffRule, LATENCY_TIER_MS, type LatencyTier, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, allProfiles, allProfilesRaw, getProfile, latencyTierOf, profilesByProvider, tryGetProfile };

package/dist/profiles.js

@@ -21,14 +21,27 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 var profiles_exports = {};
 __export(profiles_exports, {
   ALIASES: () => ALIASES,
+  LATENCY_TIER_MS: () => LATENCY_TIER_MS,
   _setProfileBrainHook: () => _setProfileBrainHook,
   allProfiles: () => allProfiles,
   allProfilesRaw: () => allProfilesRaw,
   getProfile: () => getProfile,
+  latencyTierOf: () => latencyTierOf,
   profilesByProvider: () => profilesByProvider,
   tryGetProfile: () => tryGetProfile
 });
 module.exports = __toCommonJS(profiles_exports);
+var LATENCY_TIER_MS = {
+  fast: 4e3,
+  medium: 11e3,
+  slow: 24e3
+};
+function latencyTierOf(profile) {
+  if (profile.latencyTier) return profile.latencyTier;
+  if (profile.weaknesses.includes("latency")) return "slow";
+  if (profile.strengths.includes("speed")) return "fast";
+  return "medium";
+}
 var ANTHROPIC_LOWERING_BASE = {
   system: { mode: "inline" },
   cache: {
@@ -575,6 +588,14 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "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×
+    // gemini-2.5-flash on PB's synchronous /api/analyze path. The 'flash' name
+    // is DeepSeek's, not a speed promise. This is the row that, scoring 0.85
+    // baseQuality (it carries 'reasoning') with no latency counterweight,
+    // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
+    latencyTier: "slow",
     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.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
@@ -642,6 +663,11 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "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
+    // otherwise.
+    latencyTier: "slow",
     notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.
@@ -1295,10 +1321,12 @@ function profilesByProvider(provider) {
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   ALIASES,
+  LATENCY_TIER_MS,
   _setProfileBrainHook,
   allProfiles,
   allProfilesRaw,
   getProfile,
+  latencyTierOf,
   profilesByProvider,
   tryGetProfile
 });

package/dist/profiles.mjs

@@ -1,18 +1,22 @@
 import {
   ALIASES,
+  LATENCY_TIER_MS,
   _setProfileBrainHook,
   allProfiles,
   allProfilesRaw,
   getProfile,
+  latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-ZHUD3I52.mjs";
+} from "./chunk-PKOFXEB3.mjs";
 export {
   ALIASES,
+  LATENCY_TIER_MS,
   _setProfileBrainHook,
   allProfiles,
   allProfilesRaw,
   getProfile,
+  latencyTierOf,
   profilesByProvider,
   tryGetProfile
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.46",
+  "version": "2.0.0-alpha.48",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",