adaptive-memory-multi-model-router 2.15.2 → 2.15.3

package/dist/routing/shadowSampler.js

@@ -0,0 +1,306 @@
+"use strict";
+/**
+ * ShadowSampler — Value-Proportional Shadow Verification
+ *
+ * Implements Optimal Defense Theory (Rhoades 1979; McKey 1974; Zangerl & Bazzaz 1992)
+ * from plant ecological economics:
+ *
+ *   "A plant allocates defensive compounds (defense cost) in proportion to
+ *    tissue value, attack probability, and the marginal cost of defense."
+ *
+ * Mapped to A3M routing:
+ *   - Tissue value     → query stakes / expected cost of wrong answer
+ *   - Attack probability → probability that the primary provider fails/misbehaves
+ *   - Defense cost    → cost of the shadow verification call
+ *   - Marginal return → expected reduction in error rate from verification
+ *
+ * Instead of binary always-on (wasteful) or always-off (risky), ODT says:
+ *   Sample shadow verification PROPORTIONALLY to expected value of verification.
+ *
+ * The sampling probability is:
+ *   P(shadow) = f(expected_error_cost_reduction, query_risk)
+ *
+ * This is NOT the Zahavi "shared expert" (which always runs on 100% of queries).
+ * ODT shadow runs probabilistically — only on queries where the expected
+ * verification benefit exceeds the verification cost.
+ *
+ * Usage:
+ *   const sampler = new ShadowSampler();
+ *   const decision = await sampler.routeWithShadow(query, { strategy: 'auto' });
+ *   // decision.hasShadow === true iff we sampled a shadow provider
+ *   // decision.primaryResult and decision.shadowResult are compared automatically
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ShadowSampler = void 0;
+const advancedRouter_1 = require("./advancedRouter");
+const providerConfig_1 = require("../providers/providerConfig");
+/**
+ * Optimal Defense Theory Shadow Sampler
+ *
+ * Allocates shadow verification proportionally to query value and risk.
+ *
+ * The ODT sampling probability:
+ *   P_shadow = min(maxP, baseP + risk_adj + cost_adj + complexity_adj)
+ *
+ * where:
+ *   baseP     = baseline verification rate (ODT " constitutive defense")
+ *   risk_adj  = risk_profile scaling (ODT "induced defense")
+ *   cost_adj  = output cost scaling (value of correct answer)
+ *   complexity_adj = query complexity signal
+ */
+class ShadowSampler {
+    config;
+    // Cached shadow provider (auto-selected on first call)
+    _shadowProvider = null;
+    // Counters for sampling statistics
+    _shadowCount = 0;
+    _totalCount = 0;
+    constructor(config = {}) {
+        this.config = {
+            minQueryStake: config.minQueryStake ?? 0.01,
+            maxShadowProbability: config.maxShadowProbability ?? 0.15,
+            riskScalingFactor: config.riskScalingFactor ?? 0.5,
+            costScalingFactor: config.costScalingFactor ?? 0.3,
+            shadowProvider: config.shadowProvider ?? null,
+            forceShadowProvider: config.forceShadowProvider ?? false,
+            useComplexitySignal: config.useComplexitySignal ?? true,
+            forceShadow: config.forceShadow ?? false,
+        };
+    }
+    /**
+     * Auto-select the cheapest available provider as the shadow.
+     * Excludes the primary provider to ensure diversity.
+     */
+    selectShadowProvider(primaryProvider) {
+        if (this.config.forceShadowProvider && this.config.shadowProvider) {
+            return this.config.shadowProvider;
+        }
+        if (this._shadowProvider)
+            return this._shadowProvider;
+        const profiles = (0, providerConfig_1.getAvailableProviders)();
+        const candidates = Object.entries(profiles)
+            .filter(([name, p]) => {
+            // Exclude primary
+            if (name === primaryProvider)
+                return false;
+            // Must be available (has API key) — cost must be finite
+            const cost = (p.costPerK.input + p.costPerK.output) / 2;
+            return cost < Infinity;
+        })
+            .sort((a, b) => {
+            const costA = (a[1].costPerK.input + a[1].costPerK.output) / 2;
+            const costB = (b[1].costPerK.input + b[1].costPerK.output) / 2;
+            return costA - costB;
+        });
+        if (candidates.length === 0) {
+            // Fallback: pick any non-primary
+            const fallback = Object.keys(profiles).find(n => n !== primaryProvider);
+            this._shadowProvider = fallback || primaryProvider;
+        }
+        else {
+            this._shadowProvider = candidates[0][0];
+        }
+        return this._shadowProvider;
+    }
+    /**
+     * Estimate the query "stakes" — the expected cost of a wrong answer.
+     *
+     * ODT maps this to "tissue value": how much is this asset worth protecting?
+     * In routing terms: if this query fails, how much does it cost?
+     *
+     * We approximate this as:
+     *   stake ≈ estimated_output_tokens × cost_per_token × criticality_multiplier
+     *
+     * where criticality is derived from query complexity and risk_profile.
+     */
+    estimateQueryStake(features) {
+        const estimatedTokens = features.length * 1.5; // rough estimate
+        const avgCostPerToken = 0.0001; // roughly $0.10/1K tokens
+        const stake = estimatedTokens * avgCostPerToken;
+        // Criticality multiplier from risk profile
+        const riskMultiplier = {
+            high: 10.0, // Wrong answer could cause real harm — worth verifying
+            medium: 2.0, // Some cost to wrong answer
+            low: 0.5, // Low stakes — skip verification
+        };
+        const riskMult = riskMultiplier[features.risk_profile || 'medium'] ?? 1.0;
+        // Complexity multiplier — complex queries are harder to verify but more valuable
+        // We use complexity as a proxy for "correctness is harder to judge"
+        const complexityMultiplier = 1 + (features.complexity || 0) * 2;
+        return stake * riskMult * complexityMultiplier;
+    }
+    /**
+     * Compute the ODT sampling probability for this query.
+     *
+     * ODT principle: defense (shadow) allocation ∝ expected benefit of defense.
+     * Expected benefit = P(failure) × cost_of_failure
+     *
+     * So P(shadow) scales with:
+     *   1. Query stake (expected cost of wrong answer)
+     *   2. Risk profile (probability of primary failure)
+     *   3. Output cost (verification ROI — cheaper outputs need less verification)
+     *   4. Complexity (complex queries benefit more from verification)
+     */
+    computeShadowProbability(features, stake) {
+        // Base probability (ODT "constitutive defense" — baseline verification rate)
+        let p = 0.02; // 2% baseline
+        // === STAKES ADJUSTMENT (ODT "tissue value") ===
+        // Higher stake → proportionally higher verification probability
+        // Scale from 2% to maxP as stake goes from minStake to $1.00
+        if (stake >= this.config.minQueryStake) {
+            const stakeAdj = Math.min((stake - this.config.minQueryStake) / (1.0 - this.config.minQueryStake), 1.0) * (this.config.maxShadowProbability - 0.02);
+            p += stakeAdj;
+        }
+        else {
+            // Below minimum stake — no verification regardless
+            return 0;
+        }
+        // === RISK ADJUSTMENT (ODT "attack probability") ===
+        // High-risk queries are more likely to have primary failures
+        // Scale risk_adj by riskScalingFactor (0.5 = 50% boost for high-risk)
+        const riskAdj = {
+            high: 0.10 * this.config.riskScalingFactor,
+            medium: 0.03 * this.config.riskScalingFactor,
+            low: 0.0,
+        };
+        p += riskAdj[features.risk_profile || 'medium'] ?? 0;
+        // === COMPLEXITY ADJUSTMENT (ODT "defense efficacy") ===
+        // Complex queries benefit more from verification (more errors to catch)
+        // But also harder to verify (requires domain knowledge to judge)
+        if (this.config.useComplexitySignal) {
+            if (features.has_code) {
+                p += 0.05; // Code verification has high ROI (bugs are costly)
+            }
+            if (features.requires_reasoning) {
+                p += 0.03; // Reasoning errors are subtle but costly
+            }
+            if (features.has_math) {
+                p += 0.04; // Math has objective ground truth — easy to verify (note: field is has_math)
+            }
+        }
+        // === COST SCALING (ODT "marginal defense cost") ===
+        // Expensive outputs are harder to produce — verify to avoid waste
+        if (this.config.costScalingFactor > 0 && features.length > 500) {
+            const costAdj = Math.min((features.length - 500) / 10000, // scale up for very long outputs
+            0.05) * this.config.costScalingFactor;
+            p += costAdj;
+        }
+        // Clamp to [0, maxShadowProbability]
+        return Math.min(Math.max(p, 0), this.config.maxShadowProbability);
+    }
+    /**
+     * Route a query, with optional ODT-proportional shadow verification.
+     *
+     * @param prompt - The user query
+     * @param options - Routing options (same as routeQuery)
+     * @returns ShadowDecision with shadow metadata
+     */
+    routeWithShadow(prompt, options) {
+        const features = (0, advancedRouter_1.extractQueryFeatures)(prompt);
+        const primaryDecision = (0, advancedRouter_1.routeQuery)(prompt, options?.available_models, options?.budget_multiplier);
+        // Compute query stakes and shadow probability
+        const stake = this.estimateQueryStake(features);
+        const shadowProb = this.computeShadowProbability(features, stake);
+        // ODT sampling decision
+        const shouldShadow = this.config.forceShadow || Math.random() < shadowProb;
+        const shadowProvider = shouldShadow
+            ? this.selectShadowProvider(primaryDecision.primary_model || '')
+            : null;
+        if (shouldShadow) {
+            this._shadowCount++;
+        }
+        this._totalCount++;
+        const reasoning = shouldShadow
+            ? `ODT shadow: stake=${stake.toFixed(4)}, risk=${features.risk_profile}, P=${shadowProb.toFixed(3)}, complexity=${features.complexity.toFixed(2)}`
+            : `ODT no-shadow: stake=${stake.toFixed(4)} below threshold`;
+        return {
+            ...primaryDecision,
+            hasShadow: shouldShadow,
+            shadowProvider,
+            shadowProbability: shadowProb,
+            features,
+            shadowReasoning: reasoning,
+        };
+    }
+    /**
+     * Compare primary and shadow outputs.
+     * Returns the "better" answer and confidence delta.
+     *
+     * For production use: this would call both providers in parallel
+     * and compare outputs. For now, returns a stub that signals
+     * the caller should handle comparison.
+     */
+    async compareOutputs(primaryAnswer, shadowAnswer) {
+        if (!shadowAnswer) {
+            return {
+                primary: primaryAnswer,
+                shadow: null,
+                winner: 'no-shadow',
+                confidenceDelta: 0,
+            };
+        }
+        // Simple comparison: length + character overlap as proxy for agreement
+        // In production, this would use a proper semantic similarity check
+        const primaryLen = primaryAnswer.length;
+        const shadowLen = shadowAnswer.length;
+        const lengthRatio = Math.min(primaryLen, shadowLen) / Math.max(primaryLen, shadowLen);
+        // Count common bigrams as a simple similarity proxy
+        const primaryBigrams = new Set();
+        const shadowBigrams = new Set();
+        for (let i = 0; i < primaryAnswer.length - 1; i++) {
+            primaryBigrams.add(primaryAnswer.slice(i, i + 2));
+        }
+        for (let i = 0; i < shadowAnswer.length - 1; i++) {
+            shadowBigrams.add(shadowAnswer.slice(i, i + 2));
+        }
+        let intersection = 0;
+        for (const bg of primaryBigrams) {
+            if (shadowBigrams.has(bg))
+                intersection++;
+        }
+        const union = primaryBigrams.size + shadowBigrams.size - intersection;
+        const jaccard = union > 0 ? intersection / union : 0;
+        // High agreement (jaccard > 0.8) → trust primary
+        // Low agreement → flag for review or prefer primary
+        const winner = jaccard > 0.8 ? 'primary'
+            : jaccard > 0.5 ? 'tie'
+                : 'shadow'; // Low agreement: shadow might have caught something
+        const confidenceDelta = jaccard > 0.8 ? 0 : -0.1; // Reduce confidence if they differ
+        return {
+            primary: primaryAnswer,
+            shadow: shadowAnswer,
+            winner,
+            confidenceDelta,
+        };
+    }
+    /**
+     * Get sampling statistics for monitoring.
+     */
+    getStats() {
+        return {
+            shadowCount: this._shadowCount,
+            totalCount: this._totalCount,
+            shadowRate: this._totalCount > 0 ? this._shadowCount / this._totalCount : 0,
+        };
+    }
+    /**
+     * Reset statistics counters.
+     */
+    resetStats() {
+        this._shadowCount = 0;
+        this._totalCount = 0;
+    }
+    /**
+     * Update configuration at runtime.
+     */
+    configure(config) {
+        this.config = { ...this.config, ...config };
+    }
+}
+exports.ShadowSampler = ShadowSampler;
+// ============================================================
+// NOTE: Named exports are at declaration level above.
+// ShadowSampler, ShadowSamplerConfig, ShadowDecision, ShadowedResponse
+// are all exported via 'export interface' / 'export class'
+// ============================================================
+//# sourceMappingURL=shadowSampler.js.map

package/dist/routing/shadowSampler.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"shadowSampler.js","sourceRoot":"","sources":["../../src/routing/shadowSampler.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;;;AAEH,qDAAmF;AACnF,gEAAwF;AAiFxF;;;;;;;;;;;;;GAaG;AACH,MAAa,aAAa;IAChB,MAAM,CAAgG;IAE9G,uDAAuD;IAC/C,eAAe,GAAkB,IAAI,CAAC;IAE9C,mCAAmC;IAC3B,YAAY,GAAG,CAAC,CAAC;IACjB,WAAW,GAAG,CAAC,CAAC;IAExB,YAAY,SAA8B,EAAE;QAC1C,IAAI,CAAC,MAAM,GAAG;YACZ,aAAa,EAAE,MAAM,CAAC,aAAa,IAAI,IAAI;YAC3C,oBAAoB,EAAE,MAAM,CAAC,oBAAoB,IAAI,IAAI;YACzD,iBAAiB,EAAE,MAAM,CAAC,iBAAiB,IAAI,GAAG;YAClD,iBAAiB,EAAE,MAAM,CAAC,iBAAiB,IAAI,GAAG;YAClD,cAAc,EAAE,MAAM,CAAC,cAAc,IAAI,IAAI;YAC7C,mBAAmB,EAAE,MAAM,CAAC,mBAAmB,IAAI,KAAK;YACxD,mBAAmB,EAAE,MAAM,CAAC,mBAAmB,IAAI,IAAI;YACvD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,KAAK;SACzC,CAAC;IACJ,CAAC;IAED;;;OAGG;IACK,oBAAoB,CAAC,eAAuB;QAClD,IAAI,IAAI,CAAC,MAAM,CAAC,mBAAmB,IAAI,IAAI,CAAC,MAAM,CAAC,cAAc,EAAE,CAAC;YAClE,OAAO,IAAI,CAAC,MAAM,CAAC,cAAc,CAAC;QACpC,CAAC;QAED,IAAI,IAAI,CAAC,eAAe;YAAE,OAAO,IAAI,CAAC,eAAe,CAAC;QAEtD,MAAM,QAAQ,GAAG,IAAA,sCAAqB,GAAE,CAAC;QACzC,MAAM,UAAU,GAAG,MAAM,CAAC,OAAO,CAAC,QAAQ,CAAC;aACxC,MAAM,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAA+B,EAAE,EAAE;YAClD,kBAAkB;YAClB,IAAI,IAAI,KAAK,eAAe;gBAAE,OAAO,KAAK,CAAC;YAC3C,wDAAwD;YACxD,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,GAAG,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YACxD,OAAO,IAAI,GAAG,QAAQ,CAAC;QACzB,CAAC,CAAC;aACD,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;YACb,MAAM,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YAC/D,MAAM,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YAC/D,OAAO,KAAK,GAAG,KAAK,CAAC;QACvB,CAAC,CAAC,CAAC;QAEL,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC5B,iCAAiC;YACjC,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,KAAK,eAAe,CAAC,CAAC;YACxE,IAAI,CAAC,eAAe,GAAG,QAAQ,IAAI,eAAe,CAAC;QACrD,CAAC;aAAM,CAAC;YACN,IAAI,CAAC,eAAe,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC1C,CAAC;QAED,OAAO,IAAI,CAAC,eAAe,CAAC;IAC9B,CAAC;IAED;;;;;;;;;;OAUG;IACK,kBAAkB,CAAC,QAAiD;QAC1E,MAAM,eAAe,GAAG,QAAQ,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,iBAAiB;QAChE,MAAM,eAAe,GAAG,MAAM,CAAC,CAAC,0BAA0B;QAC1D,MAAM,KAAK,GAAG,eAAe,GAAG,eAAe,CAAC;QAEhD,2CAA2C;QAC3C,MAAM,cAAc,GAA2B;YAC7C,IAAI,EAAE,IAAI,EAAK,uDAAuD;YACtE,MAAM,EAAE,GAAG,EAAI,4BAA4B;YAC3C,GAAG,EAAE,GAAG,EAAO,iCAAiC;SACjD,CAAC;QACF,MAAM,QAAQ,GAAG,cAAc,CAAC,QAAQ,CAAC,YAAY,IAAI,QAAQ,CAAC,IAAI,GAAG,CAAC;QAE1E,iFAAiF;QACjF,oEAAoE;QACpE,MAAM,oBAAoB,GAAG,CAAC,GAAG,CAAC,QAAQ,CAAC,UAAU,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC;QAEhE,OAAO,KAAK,GAAG,QAAQ,GAAG,oBAAoB,CAAC;IACjD,CAAC;IAED;;;;;;;;;;;OAWG;IACK,wBAAwB,CAC9B,QAAiD,EACjD,KAAa;QAEb,6EAA6E;QAC7E,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC,cAAc;QAE5B,iDAAiD;QACjD,gEAAgE;QAChE,6DAA6D;QAC7D,IAAI,KAAK,IAAI,IAAI,CAAC,MAAM,CAAC,aAAa,EAAE,CAAC;YACvC,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CACvB,CAAC,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,GAAG,CAAC,GAAG,GAAG,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,EACvE,GAAG,CACJ,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,oBAAoB,GAAG,IAAI,CAAC,CAAC;YAC9C,CAAC,IAAI,QAAQ,CAAC;QAChB,CAAC;aAAM,CAAC;YACN,mDAAmD;YACnD,OAAO,CAAC,CAAC;QACX,CAAC;QAED,qDAAqD;QACrD,6DAA6D;QAC7D,sEAAsE;QACtE,MAAM,OAAO,GAA2B;YACtC,IAAI,EAAE,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,iBAAiB;YAC1C,MAAM,EAAE,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,iBAAiB;YAC5C,GAAG,EAAE,GAAG;SACT,CAAC;QACF,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,YAAY,IAAI,QAAQ,CAAC,IAAI,CAAC,CAAC;QAErD,yDAAyD;QACzD,wEAAwE;QACxE,iEAAiE;QACjE,IAAI,IAAI,CAAC,MAAM,CAAC,mBAAmB,EAAE,CAAC;YACpC,IAAI,QAAQ,CAAC,QAAQ,EAAE,CAAC;gBACtB,CAAC,IAAI,IAAI,CAAC,CAAC,mDAAmD;YAChE,CAAC;YACD,IAAI,QAAQ,CAAC,kBAAkB,EAAE,CAAC;gBAChC,CAAC,IAAI,IAAI,CAAC,CAAC,yCAAyC;YACtD,CAAC;YACD,IAAI,QAAQ,CAAC,QAAQ,EAAE,CAAC;gBACtB,CAAC,IAAI,IAAI,CAAC,CAAC,6EAA6E;YAC1F,CAAC;QACH,CAAC;QAED,qDAAqD;QACrD,kEAAkE;QAClE,IAAI,IAAI,CAAC,MAAM,CAAC,iBAAiB,GAAG,CAAC,IAAI,QAAQ,CAAC,MAAM,GAAG,GAAG,EAAE,CAAC;YAC/D,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,CACtB,CAAC,QAAQ,CAAC,MAAM,GAAG,GAAG,CAAC,GAAG,KAAK,EAAE,iCAAiC;YAClE,IAAI,CACL,GAAG,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC;YAClC,CAAC,IAAI,OAAO,CAAC;QACf,CAAC;QAED,qCAAqC;QACrC,OAAO,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,MAAM,CAAC,oBAAoB,CAAC,CAAC;IACpE,CAAC;IAED;;;;;;OAMG;IACH,eAAe,CACb,MAAc,EACd,OAAqE;QAErE,MAAM,QAAQ,GAAG,IAAA,qCAAoB,EAAC,MAAM,CAAC,CAAC;QAC9C,MAAM,eAAe,GAAG,IAAA,2BAAU,EAAC,MAAM,EAAE,OAAO,EAAE,gBAAgB,EAAE,OAAO,EAAE,iBAAiB,CAAC,CAAC;QAElG,8CAA8C;QAC9C,MAAM,KAAK,GAAG,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,CAAC;QAChD,MAAM,UAAU,GAAG,IAAI,CAAC,wBAAwB,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;QAElE,wBAAwB;QACxB,MAAM,YAAY,GAAG,IAAI,CAAC,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,EAAE,GAAG,UAAU,CAAC;QAC3E,MAAM,cAAc,GAAG,YAAY;YACjC,CAAC,CAAC,IAAI,CAAC,oBAAoB,CAAC,eAAe,CAAC,aAAa,IAAI,EAAE,CAAC;YAChE,CAAC,CAAC,IAAI,CAAC;QAET,IAAI,YAAY,EAAE,CAAC;YACjB,IAAI,CAAC,YAAY,EAAE,CAAC;QACtB,CAAC;QACD,IAAI,CAAC,WAAW,EAAE,CAAC;QAEnB,MAAM,SAAS,GAAG,YAAY;YAC5B,CAAC,CAAC,qBAAqB,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,UAAU,QAAQ,CAAC,YAAY,OAAO,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,gBAAgB,QAAQ,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE;YAClJ,CAAC,CAAC,wBAAwB,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,kBAAkB,CAAC;QAE/D,OAAO;YACL,GAAG,eAAe;YAClB,SAAS,EAAE,YAAY;YACvB,cAAc;YACd,iBAAiB,EAAE,UAAU;YAC7B,QAAQ;YACR,eAAe,EAAE,SAAS;SAC3B,CAAC;IACJ,CAAC;IAED;;;;;;;OAOG;IACH,KAAK,CAAC,cAAc,CAClB,aAAqB,EACrB,YAA2B;QAE3B,IAAI,CAAC,YAAY,EAAE,CAAC;YAClB,OAAO;gBACL,OAAO,EAAE,aAAa;gBACtB,MAAM,EAAE,IAAI;gBACZ,MAAM,EAAE,WAAW;gBACnB,eAAe,EAAE,CAAC;aACnB,CAAC;QACJ,CAAC;QAED,uEAAuE;QACvE,mEAAmE;QACnE,MAAM,UAAU,GAAG,aAAa,CAAC,MAAM,CAAC;QACxC,MAAM,SAAS,GAAG,YAAY,CAAC,MAAM,CAAC;QACtC,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,EAAE,SAAS,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,EAAE,SAAS,CAAC,CAAC;QAEtF,oDAAoD;QACpD,MAAM,cAAc,GAAG,IAAI,GAAG,EAAU,CAAC;QACzC,MAAM,aAAa,GAAG,IAAI,GAAG,EAAU,CAAC;QACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAClD,cAAc,CAAC,GAAG,CAAC,aAAa,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QACpD,CAAC;QACD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACjD,aAAa,CAAC,GAAG,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAClD,CAAC;QAED,IAAI,YAAY,GAAG,CAAC,CAAC;QACrB,KAAK,MAAM,EAAE,IAAI,cAAc,EAAE,CAAC;YAChC,IAAI,aAAa,CAAC,GAAG,CAAC,EAAE,CAAC;gBAAE,YAAY,EAAE,CAAC;QAC5C,CAAC;QACD,MAAM,KAAK,GAAG,cAAc,CAAC,IAAI,GAAG,aAAa,CAAC,IAAI,GAAG,YAAY,CAAC;QACtE,MAAM,OAAO,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,YAAY,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;QAErD,iDAAiD;QACjD,oDAAoD;QACpD,MAAM,MAAM,GACV,OAAO,GAAG,GAAG,CAAC,CAAC,CAAC,SAAS;YACzB,CAAC,CAAC,OAAO,GAAG,GAAG,CAAC,CAAC,CAAC,KAAK;gBACvB,CAAC,CAAC,QAAQ,CAAC,CAAC,oDAAoD;QAElE,MAAM,eAAe,GAAG,OAAO,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,mCAAmC;QAErF,OAAO;YACL,OAAO,EAAE,aAAa;YACtB,MAAM,EAAE,YAAY;YACpB,MAAM;YACN,eAAe;SAChB,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,QAAQ;QACN,OAAO;YACL,WAAW,EAAE,IAAI,CAAC,YAAY;YAC9B,UAAU,EAAE,IAAI,CAAC,WAAW;YAC5B,UAAU,EAAE,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC;SAC5E,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,UAAU;QACR,IAAI,CAAC,YAAY,GAAG,CAAC,CAAC;QACtB,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC;IACvB,CAAC;IAED;;OAEG;IACH,SAAS,CAAC,MAAoC;QAC5C,IAAI,CAAC,MAAM,GAAG,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,MAAM,EAAE,CAAC;IAC9C,CAAC;CACF;AApSD,sCAoSC;AAED,+DAA+D;AAC/D,sDAAsD;AACtD,uEAAuE;AACvE,2DAA2D;AAC3D,+DAA+D"}

package/dist/sdk.d.ts

@@ -133,4 +133,3 @@ export declare class A3MRouter {
  * @returns Configured A3MRouter instance
  */
 export declare function createSDK(config?: A3MRouterConfig): A3MRouter;
-//# sourceMappingURL=sdk.d.ts.map

package/dist/security/guardrails.d.ts

@@ -74,4 +74,3 @@ export declare class GuardrailEngine {
 }
 export declare function createGuardrails(config?: Partial<GuardrailConfig>): GuardrailEngine;
 export {};
-//# sourceMappingURL=guardrails.d.ts.map

package/dist/server/dashboard.d.ts

@@ -56,4 +56,3 @@ export declare function registerProvider(id: string, name: string): void;
  */
 export declare function handleDashboardRequest(req: http.IncomingMessage, res: http.ServerResponse): boolean;
 export declare function getDashboardHTML(): string;
-//# sourceMappingURL=dashboard.d.ts.map

package/dist/server/handlers/chatHandler.d.ts

@@ -8,4 +8,3 @@ import * as http from 'http';
 import { RouteContext } from '../router';
 export declare function handleChatCompletions(req: http.IncomingMessage, res: http.ServerResponse, ctx: RouteContext): Promise<void>;
 export default handleChatCompletions;
-//# sourceMappingURL=chatHandler.d.ts.map

package/dist/server/handlers/completionsHandler.d.ts

@@ -7,4 +7,3 @@ import * as http from 'http';
 import { RouteContext } from '../router';
 export declare function handleCompletions(req: http.IncomingMessage, res: http.ServerResponse, ctx: RouteContext): Promise<void>;
 export default handleCompletions;
-//# sourceMappingURL=completionsHandler.d.ts.map

package/dist/server/handlers/embeddingsHandler.d.ts

@@ -14,4 +14,3 @@ import * as http from 'http';
 import { RouteContext } from '../router';
 export declare function handleEmbeddings(req: http.IncomingMessage, res: http.ServerResponse, ctx: RouteContext): Promise<void>;
 export default handleEmbeddings;
-//# sourceMappingURL=embeddingsHandler.d.ts.map

package/dist/server/handlers/healthHandler.d.ts

@@ -7,4 +7,3 @@ import * as http from 'http';
 import { RouteContext } from '../router';
 export declare function handleHealth(_req: http.IncomingMessage, res: http.ServerResponse, _ctx: RouteContext): Promise<void>;
 export default handleHealth;
-//# sourceMappingURL=healthHandler.d.ts.map

package/dist/server/handlers/metricsHandler.d.ts

@@ -8,4 +8,3 @@ import * as http from 'http';
 import { RouteContext } from '../router';
 export declare function handleMetrics(_req: http.IncomingMessage, res: http.ServerResponse, _ctx: RouteContext): void;
 export default handleMetrics;
-//# sourceMappingURL=metricsHandler.d.ts.map

package/dist/server/handlers/modelsHandler.d.ts

@@ -7,4 +7,3 @@ import * as http from 'http';
 import { RouteContext } from '../router';
 export declare function handleModels(_req: http.IncomingMessage, res: http.ServerResponse, _ctx: RouteContext): void;
 export default handleModels;
-//# sourceMappingURL=modelsHandler.d.ts.map

package/dist/server/metrics.d.ts

@@ -93,4 +93,3 @@ declare const _default: {
     recordFallback: typeof recordFallback;
 };
 export default _default;
-//# sourceMappingURL=metrics.d.ts.map

package/dist/server/modelMapper.d.ts

@@ -41,4 +41,3 @@ export declare function listAvailableModels(): Array<{
     created: number;
     owned_by: string;
 }>;
-//# sourceMappingURL=modelMapper.d.ts.map

package/dist/server/proxyServer.d.ts

@@ -74,4 +74,3 @@ export declare function createProxyServer(port?: number): http.Server;
 export { costTracker, requestLogs } from './state';
 export type { RequestLog } from './state';
 export default createProxyServer;
-//# sourceMappingURL=proxyServer.d.ts.map

package/dist/server/router.d.ts

@@ -46,4 +46,3 @@ declare const _default: {
     getRegisteredRoutes: typeof getRegisteredRoutes;
 };
 export default _default;
-//# sourceMappingURL=router.d.ts.map

package/dist/server/state.d.ts

@@ -27,4 +27,3 @@ declare const _default: {
     requestLogs: RequestLog[];
 };
 export default _default;
-//# sourceMappingURL=state.d.ts.map

package/dist/tui/dashboard.d.ts

@@ -3,4 +3,3 @@
  * A3M Router — Overlay Box TUI (Sakura Light Theme)
  */
 export {};
-//# sourceMappingURL=dashboard.d.ts.map

package/dist/utils/costUtils.d.ts

@@ -55,4 +55,3 @@ export declare function qualityAdjustedCostScore(costScore: number, qualityScore
  */
 export declare function budgetAwareCostScore(costPer1K: number, budgetMultiplier?: number): number;
 export declare function runValidationTests(): void;
-//# sourceMappingURL=costUtils.d.ts.map

package/dist/utils/sorting.d.ts

@@ -10,4 +10,3 @@ export declare function quickselectTopK<T>(arr: T[], k: number, compare: (a: T)
  * Select top candidate using Quickselect
  */
 export declare function selectTop<T>(arr: T[], compare: (a: T) => number): T | undefined;
-//# sourceMappingURL=sorting.d.ts.map

package/dist/utils/tokenUtils.d.ts

@@ -50,4 +50,3 @@ declare const _default: {
     MODEL_COSTS: Record<string, TokenCost>;
 };
 export default _default;
-//# sourceMappingURL=tokenUtils.d.ts.map

package/docs/comparison-litellm.md

@@ -8,16 +8,15 @@ litellm (48K★) is the most popular LLM gateway. Here's why A3M exists alongsid
 |---------|---------|------------|
 | **Approach** | Sequential fallback | Parallel ensemble |
 | **Model selection** | Try one, fail, try next | Run all, pick best by confidence |
-| **Benchmark** | None published | #1 on RouterArena (96.77%) |
-| **Cost** | Pay for every attempt | Pay for best response |
+| **Cost model** | Pay for every attempt | Pay for best response only |
 | **Latency** | N × round-trip (sequential) | 1 × round-trip (parallel) |
 | **Memory** | None | Episodic memory across sessions |
 | **Size** | ~1.5GB (PyTorch) | 19.5KB (zero ML) |
 | **Startup** | ~3s | <100ms |
 | **GPU required** | Yes (for some models) | No |
-| **Benchmark data** | Not published | [RouterArena #1](https://github.com/RouteWorks/RouterArena/pull/144) |
-| **Routing accuracy** | Claims "100%" (no data) | 96.77% (evaluated on RouterArena benchmark) |
-| **Cheapest cost** | Not published | $0.0768/1K (#1 on leaderboard) |
+| **Routing accuracy** | Claims "100%" (no data) | 67% exact tier match (MMR-Bench) |
+| **Cost savings** | Not published | 63% vs all-premium (MMR-Bench) |
+| **Cache** | None | Semantic (embedding-based) |
 
 ## The Core Difference
 
@@ -54,11 +53,12 @@ const result = await router.route("Explain quantum computing")
 
 ## When to Use A3M
 
-- You want the **cheapest** routing (2.3× cheaper than Sqwish)
-- You want the **highest accuracy** (#1 on RouterArena)
-- You want **memory** across sessions (only router that has this)
+- You want the **cheapest** routing (parallel execution means you only pay for the best result)
+- You want **highest accuracy** (confidence-weighted ensemble beats sequential fallback)
+- You want **memory** across sessions (only router with episodic memory)
 - You want **sub-100ms startup** (litellm takes ~3s)
-- You want **zero ML dependencies** (no GPU, no PyTorch)
+- You want **zero ML dependencies** (no GPU, no PyTorch, 19.5KB)
+- You want **semantic caching** (avoid repeated API calls for similar queries)
 - You're building in Node.js/TypeScript
 
 ## When to Use Both
@@ -75,14 +75,6 @@ const router = createRouter({
 
 This gives you litellm's 100+ providers AND A3M's parallel scoring.
 
-## The Benchmark Question
-
-litellm claims "100% routing accuracy" but publishes **zero data** to back this up. RouterArena (arXiv:2510.00202) is the first standardized benchmark for LLM routers. A3M submitted, litellm didn't.
-
-> "Benchmark or GTFO." — A principle we stand by.
-
-If litellm submits to RouterArena and scores higher than 96.77%, we'll celebrate. Competition drives improvement.
-
 ---
 
-[GitHub](https://github.com/Das-rebel/a3m-router) · [npm](https://www.npmjs.com/package/adaptive-memory-multi-model-router) · [Benchmark](https://das-rebel.github.io/a3m-router/benchmark)
+[GitHub](https://github.com/Das-rebel/a3m-router) · [npm](https://www.npmjs.com/package/adaptive-memory-multi-model-router)