web-agent-bridge 3.2.0 → 3.4.0

package/server/services/agent-learning.js

@@ -1,575 +1,601 @@
-/**
- * Agent Learning Engine — Local Reinforcement Learning
- *
- * Agents learn from user decisions, building behavioral models locally
- * without sending data to external LLMs. The engine tracks:
- *   - Decision patterns (what the user chooses and when)
- *   - Reward signals (accepted/rejected/modified outcomes)
- *   - Policy weights (which factors matter most to this user)
- *   - Prediction accuracy over time
- *
- * Learning algorithms:
- *   - Multi-armed bandit (UCB1) for exploration/exploitation
- *   - Linear policy model with sigmoid activation and gradient descent
- *   - Temporal discount for preference freshness (recent > old)
- *   - Sequential pattern mining for behavior chains
- *   - Confidence estimation: volume × accuracy × recency
- */
-
-const crypto = require('crypto');
-const { db } = require('../models/db');
-
-// ─── Schema ──────────────────────────────────────────────────────────
-
-db.exec(`
-  CREATE TABLE IF NOT EXISTS learning_decisions (
-    id TEXT PRIMARY KEY,
-    site_id TEXT NOT NULL,
-    agent_id TEXT NOT NULL,
-    domain TEXT NOT NULL,
-    action TEXT NOT NULL,
-    context TEXT DEFAULT '{}',
-    outcome TEXT DEFAULT 'pending',
-    reward REAL DEFAULT 0.0,
-    predicted_reward REAL,
-    features TEXT DEFAULT '{}',
-    created_at TEXT DEFAULT (datetime('now'))
-  );
-
-  CREATE TABLE IF NOT EXISTS learning_policies (
-    id TEXT PRIMARY KEY,
-    site_id TEXT NOT NULL,
-    agent_id TEXT NOT NULL,
-    domain TEXT NOT NULL,
-    feature TEXT NOT NULL,
-    weight REAL DEFAULT 0.0,
-    update_count INTEGER DEFAULT 0,
-    last_error REAL DEFAULT 0.0,
-    created_at TEXT DEFAULT (datetime('now')),
-    updated_at TEXT DEFAULT (datetime('now')),
-    UNIQUE(site_id, agent_id, domain, feature)
-  );
-
-  CREATE TABLE IF NOT EXISTS learning_patterns (
-    id TEXT PRIMARY KEY,
-    site_id TEXT NOT NULL,
-    agent_id TEXT NOT NULL,
-    pattern_type TEXT NOT NULL,
-    sequence TEXT NOT NULL,
-    frequency INTEGER DEFAULT 1,
-    confidence REAL DEFAULT 0.5,
-    last_seen TEXT DEFAULT (datetime('now')),
-    created_at TEXT DEFAULT (datetime('now'))
-  );
-
-  CREATE TABLE IF NOT EXISTS learning_bandit_arms (
-    id TEXT PRIMARY KEY,
-    site_id TEXT NOT NULL,
-    agent_id TEXT NOT NULL,
-    domain TEXT NOT NULL,
-    action TEXT NOT NULL,
-    pulls INTEGER DEFAULT 0,
-    total_reward REAL DEFAULT 0.0,
-    avg_reward REAL DEFAULT 0.0,
-    ucb_score REAL DEFAULT 0.0,
-    created_at TEXT DEFAULT (datetime('now')),
-    updated_at TEXT DEFAULT (datetime('now')),
-    UNIQUE(site_id, agent_id, domain, action)
-  );
-
-  CREATE TABLE IF NOT EXISTS learning_sessions (
-    id TEXT PRIMARY KEY,
-    site_id TEXT NOT NULL,
-    agent_id TEXT NOT NULL,
-    decisions_made INTEGER DEFAULT 0,
-    correct_predictions INTEGER DEFAULT 0,
-    accuracy REAL DEFAULT 0.0,
-    started_at TEXT DEFAULT (datetime('now')),
-    ended_at TEXT
-  );
-
-  CREATE INDEX IF NOT EXISTS idx_learn_dec_site ON learning_decisions(site_id, agent_id);
-  CREATE INDEX IF NOT EXISTS idx_learn_dec_domain ON learning_decisions(domain);
-  CREATE INDEX IF NOT EXISTS idx_learn_dec_outcome ON learning_decisions(outcome);
-  CREATE INDEX IF NOT EXISTS idx_learn_pol_lookup ON learning_policies(site_id, agent_id, domain);
-  CREATE INDEX IF NOT EXISTS idx_learn_pat_seq ON learning_patterns(site_id, agent_id, pattern_type);
-  CREATE INDEX IF NOT EXISTS idx_learn_bandit ON learning_bandit_arms(site_id, agent_id, domain);
-`);
-
-// ─── Config ──────────────────────────────────────────────────────────
-
-const LEARNING_RATE = 0.1;
-const DISCOUNT_FACTOR = 0.95;     // Temporal discount per decision step
-const DECAY_RATE = 0.01;          // Recency decay per hour
-const UCB_EXPLORATION = 1.414;    // √2 for UCB1
-const MIN_CONFIDENCE = 0.01;
-const MAX_SEQUENCE_LENGTH = 5;
-
-// ─── Prepared Statements ─────────────────────────────────────────────
-
-const stmts = {
-  insertDecision: db.prepare('INSERT INTO learning_decisions (id, site_id, agent_id, domain, action, context, predicted_reward, features) VALUES (?, ?, ?, ?, ?, ?, ?, ?)'),
-  updateOutcome: db.prepare('UPDATE learning_decisions SET outcome = ?, reward = ? WHERE id = ?'),
-  getDecision: db.prepare('SELECT * FROM learning_decisions WHERE id = ?'),
-  getRecentDecisions: db.prepare('SELECT * FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY created_at DESC LIMIT ?'),
-  getDecisionsByOutcome: db.prepare("SELECT * FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome = ? ORDER BY created_at DESC LIMIT ?"),
-  getAllDomainDecisions: db.prepare('SELECT * FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY created_at DESC'),
-  countDecisions: db.prepare('SELECT COUNT(*) as count FROM learning_decisions WHERE site_id = ? AND agent_id = ?'),
-  getRecentRewards: db.prepare("SELECT reward, created_at FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome != 'pending' ORDER BY created_at DESC LIMIT ?"),
-
-  upsertPolicy: db.prepare("INSERT INTO learning_policies (id, site_id, agent_id, domain, feature, weight) VALUES (?, ?, ?, ?, ?, ?) ON CONFLICT(site_id, agent_id, domain, feature) DO UPDATE SET weight = ?, update_count = update_count + 1, last_error = ?, updated_at = datetime('now')"),
-  getPolicies: db.prepare('SELECT * FROM learning_policies WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY ABS(weight) DESC'),
-  getPolicy: db.prepare('SELECT * FROM learning_policies WHERE site_id = ? AND agent_id = ? AND domain = ? AND feature = ?'),
-
-  insertPattern: db.prepare('INSERT INTO learning_patterns (id, site_id, agent_id, pattern_type, sequence, confidence) VALUES (?, ?, ?, ?, ?, ?)'),
-  findPattern: db.prepare('SELECT * FROM learning_patterns WHERE site_id = ? AND agent_id = ? AND sequence = ?'),
-  updatePattern: db.prepare("UPDATE learning_patterns SET frequency = frequency + 1, confidence = ?, last_seen = datetime('now') WHERE id = ?"),
-  getTopPatterns: db.prepare('SELECT * FROM learning_patterns WHERE site_id = ? AND agent_id = ? AND pattern_type = ? ORDER BY frequency DESC, confidence DESC LIMIT ?'),
-
-  upsertArm: db.prepare('INSERT INTO learning_bandit_arms (id, site_id, agent_id, domain, action) VALUES (?, ?, ?, ?, ?) ON CONFLICT(site_id, agent_id, domain, action) DO NOTHING'),
-  getArms: db.prepare('SELECT * FROM learning_bandit_arms WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY ucb_score DESC'),
-  getArm: db.prepare('SELECT * FROM learning_bandit_arms WHERE site_id = ? AND agent_id = ? AND domain = ? AND action = ?'),
-  updateArm: db.prepare("UPDATE learning_bandit_arms SET pulls = pulls + 1, total_reward = total_reward + ?, avg_reward = ?, ucb_score = ?, updated_at = datetime('now') WHERE site_id = ? AND agent_id = ? AND domain = ? AND action = ?"),
-
-  insertSession: db.prepare('INSERT INTO learning_sessions (id, site_id, agent_id) VALUES (?, ?, ?)'),
-  updateSession: db.prepare("UPDATE learning_sessions SET decisions_made = ?, correct_predictions = ?, accuracy = ?, ended_at = datetime('now') WHERE id = ?"),
-  getSessionHistory: db.prepare('SELECT * FROM learning_sessions WHERE site_id = ? AND agent_id = ? ORDER BY started_at DESC LIMIT ?'),
-
-  getStats: db.prepare(`SELECT
-    (SELECT COUNT(*) FROM learning_decisions WHERE site_id = ? AND agent_id = ?) as total_decisions,
-    (SELECT COUNT(*) FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome = 'accepted') as accepted,
-    (SELECT COUNT(*) FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome = 'rejected') as rejected,
-    (SELECT AVG(reward) FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome != 'pending') as avg_reward,
-    (SELECT COUNT(DISTINCT domain) FROM learning_policies WHERE site_id = ? AND agent_id = ?) as policy_domains,
-    (SELECT COUNT(*) FROM learning_patterns WHERE site_id = ? AND agent_id = ?) as total_patterns`),
-
-  deletePolicies: db.prepare('DELETE FROM learning_policies WHERE site_id = ? AND agent_id = ? AND domain = ?'),
-  deletePatterns: db.prepare('DELETE FROM learning_patterns WHERE site_id = ? AND agent_id = ?'),
-  deleteArms: db.prepare('DELETE FROM learning_bandit_arms WHERE site_id = ? AND agent_id = ? AND domain = ?'),
-  deleteDecisions: db.prepare('DELETE FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND domain = ?'),
-};
-
-// ─── Core Learning API ───────────────────────────────────────────────
-
-/**
- * Record a decision the agent is about to make, with predicted reward.
- */
-function recordDecision(siteId, agentId, domain, action, context = {}, features = {}) {
-  const id = crypto.randomUUID();
-  const extractedFeatures = { ..._extractFeatures(context), ...features };
-  const predictedReward = _predict(siteId, agentId, domain, extractedFeatures);
-
-  stmts.insertDecision.run(id, siteId, agentId, domain, action,
-    JSON.stringify(context), predictedReward, JSON.stringify(extractedFeatures));
-
-  // Ensure bandit arm exists
-  stmts.upsertArm.run(crypto.randomUUID(), siteId, agentId, domain, action);
-
-  return { decisionId: id, predictedReward, confidence: _getConfidence(siteId, agentId, domain) };
-}
-
-/**
- * Provide feedback on a decision — the outcome and actual reward.
- * This is the core learning signal.
- */
-function feedback(decisionId, outcome, reward) {
-  const decision = stmts.getDecision.get(decisionId);
-  if (!decision) throw new Error('Decision not found');
-
-  stmts.updateOutcome.run(outcome, reward, decisionId);
-
-  const features = JSON.parse(decision.features || '{}');
-  const predError = reward - (decision.predicted_reward || 0);
-
-  // Update policy weights via gradient descent with temporal discount
-  _updatePolicies(decision.site_id, decision.agent_id, decision.domain, features, predError);
-
-  // Update bandit arm with actual reward
-  _updateBanditArm(decision.site_id, decision.agent_id, decision.domain, decision.action, reward);
-
-  // Mine patterns from recent decisions
-  _minePatterns(decision.site_id, decision.agent_id, decision.domain);
-
-  return {
-    decisionId,
-    predictionError: Math.round(predError * 1000) / 1000,
-    updatedConfidence: _getConfidence(decision.site_id, decision.agent_id, decision.domain),
-    accuracy: Math.round((1 - Math.abs(predError)) * 1000) / 1000,
-  };
-}
-
-/**
- * Batch feedback — provide multiple outcomes at once.
- */
-function batchFeedback(feedbackList) {
-  const results = [];
-  const txn = db.transaction(() => {
-    for (const fb of feedbackList) {
-      try {
-        results.push(feedback(fb.decisionId, fb.outcome, fb.reward));
-      } catch (err) {
-        results.push({ decisionId: fb.decisionId, error: err.message });
-      }
-    }
-  });
-  txn();
-  return results;
-}
-
-/**
- * Get the best action for a domain using learned policies + bandit scores.
- * UCB scores are normalized to [0,1] before blending with policy prediction.
- */
-function recommend(siteId, agentId, domain, availableActions, context = {}) {
-  const features = _extractFeatures(context);
-
-  // Get all arms to find normalization bounds
-  const allArms = stmts.getArms.all(siteId, agentId, domain);
-  const armMap = {};
-  for (const arm of allArms) armMap[arm.action] = arm;
-
-  // Normalize UCB scores to [0,1]
-  let minUCB = Infinity, maxUCB = -Infinity;
-  for (const arm of allArms) {
-    if (arm.pulls > 0) {
-      if (arm.ucb_score < minUCB) minUCB = arm.ucb_score;
-      if (arm.ucb_score > maxUCB) maxUCB = arm.ucb_score;
-    }
-  }
-  const ucbRange = maxUCB - minUCB;
-
-  const scored = availableActions.map((action) => {
-    const arm = armMap[action] || _getOrCreateArm(siteId, agentId, domain, action);
-    const policyScore = _predict(siteId, agentId, domain, { ...features, [`action:${action}`]: 1 });
-
-    // Normalize bandit score to [0,1]
-    let normalizedBandit;
-    if (arm.pulls === 0) {
-      normalizedBandit = 1.0; // unexplored arms get maximum exploration bonus
-    } else if (ucbRange > 0) {
-      normalizedBandit = (arm.ucb_score - minUCB) / ucbRange;
-    } else {
-      normalizedBandit = arm.avg_reward; // single arm — use raw avg
-    }
-
-    // Blend: as confidence grows, lean more on policy, less on exploration
-    const confidence = _getConfidence(siteId, agentId, domain);
-    const policyWeight = 0.4 + confidence * 0.4; // [0.4, 0.8]
-    const banditWeight = 1 - policyWeight;        // [0.2, 0.6]
-    const blended = policyWeight * policyScore + banditWeight * normalizedBandit;
-
-    return {
-      action,
-      score: Math.round(blended * 1000) / 1000,
-      policyScore: Math.round(policyScore * 1000) / 1000,
-      banditScore: Math.round(normalizedBandit * 1000) / 1000,
-      pulls: arm.pulls,
-      avgReward: Math.round((arm.avg_reward || 0) * 1000) / 1000,
-    };
-  });
-
-  scored.sort((a, b) => b.score - a.score);
-
-  const confidence = _getConfidence(siteId, agentId, domain);
-  const topPatterns = stmts.getTopPatterns.all(siteId, agentId, 'action_sequence', 5);
-
-  return {
-    recommended: scored[0]?.action || availableActions[0],
-    rankings: scored,
-    confidence,
-    explorationLevel: confidence < 0.3 ? 'high' : confidence < 0.6 ? 'medium' : 'low',
-    patterns: topPatterns.map((p) => ({
-      sequence: p.sequence, frequency: p.frequency, confidence: p.confidence
-    })),
-  };
-}
-
-/**
- * Get learned preference summary for a domain.
- */
-function getPreferences(siteId, agentId, domain) {
-  const policies = stmts.getPolicies.all(siteId, agentId, domain);
-  const decisions = stmts.getRecentDecisions.all(siteId, agentId, domain, 50);
-  const patterns = stmts.getTopPatterns.all(siteId, agentId, 'action_sequence', 10);
-
-  const accepted = decisions.filter((d) => d.outcome === 'accepted');
-  const rejected = decisions.filter((d) => d.outcome === 'rejected');
-
-  // Build preference profile from weights
-  const profile = {};
-  for (const p of policies) {
-    if (Math.abs(p.weight) > 0.05) {
-      profile[p.feature] = {
-        weight: Math.round(p.weight * 1000) / 1000,
-        direction: p.weight > 0 ? 'preferred' : 'avoided',
-        strength: Math.abs(p.weight) > 0.5 ? 'strong' : Math.abs(p.weight) > 0.2 ? 'moderate' : 'weak',
-        updates: p.update_count,
-      };
-    }
-  }
-
-  // Compute action frequencies
-  const actionFreqs = {};
-  for (const d of decisions) {
-    actionFreqs[d.action] = (actionFreqs[d.action] || 0) + 1;
-  }
-
-  return {
-    domain,
-    profile,
-    acceptRate: decisions.length > 0 ? Math.round((accepted.length / decisions.length) * 1000) / 1000 : 0,
-    rejectRate: decisions.length > 0 ? Math.round((rejected.length / decisions.length) * 1000) / 1000 : 0,
-    totalDecisions: decisions.length,
-    avgReward: decisions.length > 0
-      ? Math.round((decisions.reduce((s, d) => s + d.reward, 0) / decisions.length) * 1000) / 1000
-      : 0,
-    topActions: Object.entries(actionFreqs)
-      .sort(([, a], [, b]) => b - a)
-      .slice(0, 5)
-      .map(([action, count]) => ({ action, count, percentage: Math.round((count / decisions.length) * 100) })),
-    topPatterns: patterns.map((p) => ({ sequence: p.sequence, frequency: p.frequency })),
-    confidence: _getConfidence(siteId, agentId, domain),
-  };
-}
-
-/**
- * Get reward history — recent rewards over time for charting.
- */
-function getRewardHistory(siteId, agentId, limit = 30) {
-  return stmts.getRecentRewards.all(siteId, agentId, limit).reverse();
-}
-
-// ─── Learning Sessions ───────────────────────────────────────────────
-
-function startSession(siteId, agentId) {
-  const id = crypto.randomUUID();
-  stmts.insertSession.run(id, siteId, agentId);
-  return { sessionId: id };
-}
-
-function endSession(sessionId, decisionsMade, correctPredictions) {
-  const accuracy = decisionsMade > 0 ? correctPredictions / decisionsMade : 0;
-  stmts.updateSession.run(decisionsMade, correctPredictions, accuracy, sessionId);
-  return { accuracy: Math.round(accuracy * 1000) / 1000 };
-}
-
-// ─── Reset ───────────────────────────────────────────────────────────
-
-/**
- * Reset all learned data for a specific domain.
- */
-function resetDomain(siteId, agentId, domain) {
-  const txn = db.transaction(() => {
-    stmts.deletePolicies.run(siteId, agentId, domain);
-    stmts.deleteArms.run(siteId, agentId, domain);
-    stmts.deleteDecisions.run(siteId, agentId, domain);
-  });
-  txn();
-  return { reset: true, domain };
-}
-
-/**
- * Reset all patterns for an agent.
- */
-function resetPatterns(siteId, agentId) {
-  stmts.deletePatterns.run(siteId, agentId);
-  return { reset: true };
-}
-
-// ─── Stats ───────────────────────────────────────────────────────────
-
-function getStats(siteId, agentId) {
-  const row = stmts.getStats.get(siteId, agentId, siteId, agentId, siteId, agentId, siteId, agentId, siteId, agentId, siteId, agentId);
-  const sessions = stmts.getSessionHistory.all(siteId, agentId, 10);
-  const recentAccuracy = sessions.length > 0 ? sessions.reduce((s, sess) => s + sess.accuracy, 0) / sessions.length : 0;
-  const rewardHistory = stmts.getRecentRewards.all(siteId, agentId, 30).reverse();
-
-  return {
-    ...row,
-    avg_reward: row.avg_reward !== null ? Math.round(row.avg_reward * 1000) / 1000 : 0,
-    recentAccuracy: Math.round(recentAccuracy * 1000) / 1000,
-    sessionsCount: sessions.length,
-    acceptRate: row.total_decisions > 0
-      ? Math.round((row.accepted / row.total_decisions) * 1000) / 1000
-      : 0,
-    rewardHistory,
-  };
-}
-
-// ─── Internal: Prediction via Linear Model ───────────────────────────
-
-function _predict(siteId, agentId, domain, features) {
-  const policies = stmts.getPolicies.all(siteId, agentId, domain);
-  if (policies.length === 0) return 0.5; // No data yet — neutral prediction
-
-  let score = 0;
-  let matchedFeatures = 0;
-  for (const p of policies) {
-    const featureVal = features[p.feature];
-    if (featureVal !== undefined) {
-      const fv = typeof featureVal === 'number' ? featureVal : (featureVal ? 1 : 0);
-
-      // Apply temporal discount: older policies (fewer recent updates) matter less
-      const recencyBoost = p.update_count > 0 ? Math.pow(DISCOUNT_FACTOR, Math.max(0, 10 - p.update_count)) : 1;
-      score += p.weight * fv * recencyBoost;
-      matchedFeatures++;
-    }
-  }
-
-  // Sigmoid squash to [0, 1]
-  return 1 / (1 + Math.exp(-score));
-}
-
-function _updatePolicies(siteId, agentId, domain, features, error) {
-  for (const [feature, value] of Object.entries(features)) {
-    const fv = typeof value === 'number' ? value : (value ? 1 : 0);
-    if (fv === 0) continue; // Skip zero-valued features
-
-    const gradient = error * fv * LEARNING_RATE;
-    const existing = stmts.getPolicy.get(siteId, agentId, domain, feature);
-
-    // Apply weight decay to prevent unbounded growth
-    const currentWeight = existing ? existing.weight * DISCOUNT_FACTOR : 0;
-    const newWeight = currentWeight + gradient;
-
-    // Clamp weights to [-5, 5] to prevent extreme values
-    const clampedWeight = Math.max(-5, Math.min(5, newWeight));
-
-    stmts.upsertPolicy.run(
-      crypto.randomUUID(), siteId, agentId, domain, feature, clampedWeight,
-      clampedWeight, Math.abs(error)
-    );
-  }
-}
-
-// ─── Internal: Multi-Armed Bandit ────────────────────────────────────
-
-function _getOrCreateArm(siteId, agentId, domain, action) {
-  stmts.upsertArm.run(crypto.randomUUID(), siteId, agentId, domain, action);
-  const arm = stmts.getArm.get(siteId, agentId, domain, action);
-  return arm || { pulls: 0, ucb_score: 0, avg_reward: 0, total_reward: 0 };
-}
-
-function _updateBanditArm(siteId, agentId, domain, action, reward) {
-  const arm = stmts.getArm.get(siteId, agentId, domain, action);
-  if (!arm) {
-    stmts.upsertArm.run(crypto.randomUUID(), siteId, agentId, domain, action);
-    return;
-  }
-
-  const newPulls = arm.pulls + 1;
-  const newTotalReward = arm.total_reward + reward;
-  const newAvgReward = newTotalReward / newPulls;
-
-  // UCB1: avg_reward + C * sqrt(ln(N) / n_i)
-  // We need total pulls across all arms in this domain
-  const arms = stmts.getArms.all(siteId, agentId, domain);
-  const totalPulls = arms.reduce((s, a) => s + a.pulls, 0) + 1; // +1 for this pull
-
-  const exploration = UCB_EXPLORATION * Math.sqrt(Math.log(totalPulls) / newPulls);
-  const ucbScore = newAvgReward + exploration;
-
-  stmts.updateArm.run(reward, newAvgReward, ucbScore, siteId, agentId, domain, action);
-}
-
-// ─── Internal: Pattern Mining ────────────────────────────────────────
-
-function _minePatterns(siteId, agentId, domain) {
-  const decisions = stmts.getRecentDecisions.all(siteId, agentId, domain, 20);
-  if (decisions.length < 3) return;
-
-  // Extract action sequences of length 2-5
-  for (let len = 2; len <= Math.min(MAX_SEQUENCE_LENGTH, decisions.length); len++) {
-    const sequence = decisions.slice(0, len).map((d) => d.action).reverse().join(' → ');
-    const existing = stmts.findPattern.get(siteId, agentId, sequence);
-
-    if (existing) {
-      // Asymptotic approach to 1.0 — confidence grows slower as it increases
-      const newConf = Math.min(0.99, existing.confidence + 0.05 * (1 - existing.confidence));
-      stmts.updatePattern.run(newConf, existing.id);
-    } else {
-      stmts.insertPattern.run(crypto.randomUUID(), siteId, agentId, 'action_sequence', sequence, 0.3);
-    }
-  }
-}
-
-// ─── Internal: Feature Extraction ────────────────────────────────────
-
-function _extractFeatures(context) {
-  const features = {};
-
-  if (context.price !== undefined) {
-    features.price = context.price;
-    // Bucketize price for discrete learning
-    if (context.price < 10) features['price_bucket:cheap'] = 1;
-    else if (context.price < 50) features['price_bucket:moderate'] = 1;
-    else if (context.price < 200) features['price_bucket:premium'] = 1;
-    else features['price_bucket:luxury'] = 1;
-  }
-  if (context.quantity !== undefined) features.quantity = context.quantity;
-  if (context.discount !== undefined) {
-    features.discount = context.discount;
-    features.has_discount = context.discount > 0 ? 1 : 0;
-  }
-  if (context.rating !== undefined) {
-    features.rating = context.rating;
-    features.high_rated = context.rating >= 4.0 ? 1 : 0;
-  }
-  if (context.category) features[`category:${context.category}`] = 1;
-  if (context.brand) features[`brand:${context.brand}`] = 1;
-  if (context.timeOfDay !== undefined) {
-    features.morning = context.timeOfDay < 12 ? 1 : 0;
-    features.afternoon = context.timeOfDay >= 12 && context.timeOfDay < 18 ? 1 : 0;
-    features.evening = context.timeOfDay >= 18 ? 1 : 0;
-  }
-  if (context.isRepeat !== undefined) features.repeat_visit = context.isRepeat ? 1 : 0;
-  if (context.urgency !== undefined) features.urgency = context.urgency;
-  if (context.inStock !== undefined) features.in_stock = context.inStock ? 1 : 0;
-
-  // Pass through any raw numeric features
-  for (const [k, v] of Object.entries(context)) {
-    if (features[k] === undefined && typeof v === 'number') {
-      features[k] = v;
-    }
-  }
-
-  return features;
-}
-
-// ─── Internal: Confidence Estimation ─────────────────────────────────
-
-function _getConfidence(siteId, agentId, domain) {
-  const decisions = stmts.getRecentDecisions.all(siteId, agentId, domain, 50);
-  if (decisions.length === 0) return 0;
-
-  const withOutcome = decisions.filter((d) => d.outcome !== 'pending');
-  if (withOutcome.length === 0) return MIN_CONFIDENCE;
-
-  // Volume component: log scale, saturates around 30 decisions
-  const volumeConf = Math.min(1, withOutcome.length / 30);
-
-  // Accuracy component: how close predictions were to actual rewards
-  let accuracySum = 0;
-  for (const d of withOutcome) {
-    if (d.predicted_reward !== null) {
-      const error = Math.abs(d.reward - d.predicted_reward);
-      accuracySum += Math.max(0, 1 - error);
-    }
-  }
-  const accuracyConf = withOutcome.length > 0 ? accuracySum / withOutcome.length : 0.5;
-
-  // Recency component: exponential decay based on age of newest data
-  const latestTs = new Date(withOutcome[0].created_at).getTime();
-  const ageHours = (Date.now() - latestTs) / 3600000;
-  const recencyConf = Math.exp(-DECAY_RATE * ageHours);
-
-  return Math.max(MIN_CONFIDENCE, Math.min(0.99,
-    volumeConf * 0.3 + accuracyConf * 0.5 + recencyConf * 0.2
-  ));
-}
-
-module.exports = {
-  recordDecision, feedback, batchFeedback, recommend, getPreferences,
-  getRewardHistory, startSession, endSession,
-  resetDomain, resetPatterns, getStats,
-};
+/**
+ * Agent Learning Engine — Local Reinforcement Learning
+ *
+ * Agents learn from user decisions, building behavioral models locally
+ * without sending data to external LLMs. The engine tracks:
+ *   - Decision patterns (what the user chooses and when)
+ *   - Reward signals (accepted/rejected/modified outcomes)
+ *   - Policy weights (which factors matter most to this user)
+ *   - Prediction accuracy over time
+ *
+ * Learning algorithms:
+ *   - Multi-armed bandit (UCB1) for exploration/exploitation
+ *   - Linear policy model with sigmoid activation and gradient descent
+ *   - Temporal discount for preference freshness (recent > old)
+ *   - Sequential pattern mining for behavior chains
+ *   - Confidence estimation: volume × accuracy × recency
+ */
+
+const crypto = require('crypto');
+const { db } = require('../models/db');
+let rewardGuard;
+try { rewardGuard = require('../security/reward-guard'); } catch { rewardGuard = null; }
+
+// ─── Schema ──────────────────────────────────────────────────────────
+
+db.exec(`
+  CREATE TABLE IF NOT EXISTS learning_decisions (
+    id TEXT PRIMARY KEY,
+    site_id TEXT NOT NULL,
+    agent_id TEXT NOT NULL,
+    domain TEXT NOT NULL,
+    action TEXT NOT NULL,
+    context TEXT DEFAULT '{}',
+    outcome TEXT DEFAULT 'pending',
+    reward REAL DEFAULT 0.0,
+    predicted_reward REAL,
+    features TEXT DEFAULT '{}',
+    created_at TEXT DEFAULT (datetime('now'))
+  );
+
+  CREATE TABLE IF NOT EXISTS learning_policies (
+    id TEXT PRIMARY KEY,
+    site_id TEXT NOT NULL,
+    agent_id TEXT NOT NULL,
+    domain TEXT NOT NULL,
+    feature TEXT NOT NULL,
+    weight REAL DEFAULT 0.0,
+    update_count INTEGER DEFAULT 0,
+    last_error REAL DEFAULT 0.0,
+    created_at TEXT DEFAULT (datetime('now')),
+    updated_at TEXT DEFAULT (datetime('now')),
+    UNIQUE(site_id, agent_id, domain, feature)
+  );
+
+  CREATE TABLE IF NOT EXISTS learning_patterns (
+    id TEXT PRIMARY KEY,
+    site_id TEXT NOT NULL,
+    agent_id TEXT NOT NULL,
+    pattern_type TEXT NOT NULL,
+    sequence TEXT NOT NULL,
+    frequency INTEGER DEFAULT 1,
+    confidence REAL DEFAULT 0.5,
+    last_seen TEXT DEFAULT (datetime('now')),
+    created_at TEXT DEFAULT (datetime('now'))
+  );
+
+  CREATE TABLE IF NOT EXISTS learning_bandit_arms (
+    id TEXT PRIMARY KEY,
+    site_id TEXT NOT NULL,
+    agent_id TEXT NOT NULL,
+    domain TEXT NOT NULL,
+    action TEXT NOT NULL,
+    pulls INTEGER DEFAULT 0,
+    total_reward REAL DEFAULT 0.0,
+    avg_reward REAL DEFAULT 0.0,
+    ucb_score REAL DEFAULT 0.0,
+    created_at TEXT DEFAULT (datetime('now')),
+    updated_at TEXT DEFAULT (datetime('now')),
+    UNIQUE(site_id, agent_id, domain, action)
+  );
+
+  CREATE TABLE IF NOT EXISTS learning_sessions (
+    id TEXT PRIMARY KEY,
+    site_id TEXT NOT NULL,
+    agent_id TEXT NOT NULL,
+    decisions_made INTEGER DEFAULT 0,
+    correct_predictions INTEGER DEFAULT 0,
+    accuracy REAL DEFAULT 0.0,
+    started_at TEXT DEFAULT (datetime('now')),
+    ended_at TEXT
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_learn_dec_site ON learning_decisions(site_id, agent_id);
+  CREATE INDEX IF NOT EXISTS idx_learn_dec_domain ON learning_decisions(domain);
+  CREATE INDEX IF NOT EXISTS idx_learn_dec_outcome ON learning_decisions(outcome);
+  CREATE INDEX IF NOT EXISTS idx_learn_pol_lookup ON learning_policies(site_id, agent_id, domain);
+  CREATE INDEX IF NOT EXISTS idx_learn_pat_seq ON learning_patterns(site_id, agent_id, pattern_type);
+  CREATE INDEX IF NOT EXISTS idx_learn_bandit ON learning_bandit_arms(site_id, agent_id, domain);
+`);
+
+// ─── Config ──────────────────────────────────────────────────────────
+
+const LEARNING_RATE = 0.1;
+const DISCOUNT_FACTOR = 0.95;     // Temporal discount per decision step
+const DECAY_RATE = 0.01;          // Recency decay per hour
+const UCB_EXPLORATION = 1.414;    // √2 for UCB1
+const MIN_CONFIDENCE = 0.01;
+const MAX_SEQUENCE_LENGTH = 5;
+
+// ─── Prepared Statements ─────────────────────────────────────────────
+
+const stmts = {
+  insertDecision: db.prepare('INSERT INTO learning_decisions (id, site_id, agent_id, domain, action, context, predicted_reward, features) VALUES (?, ?, ?, ?, ?, ?, ?, ?)'),
+  updateOutcome: db.prepare('UPDATE learning_decisions SET outcome = ?, reward = ? WHERE id = ?'),
+  getDecision: db.prepare('SELECT * FROM learning_decisions WHERE id = ?'),
+  getRecentDecisions: db.prepare('SELECT * FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY created_at DESC LIMIT ?'),
+  getDecisionsByOutcome: db.prepare("SELECT * FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome = ? ORDER BY created_at DESC LIMIT ?"),
+  getAllDomainDecisions: db.prepare('SELECT * FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY created_at DESC'),
+  countDecisions: db.prepare('SELECT COUNT(*) as count FROM learning_decisions WHERE site_id = ? AND agent_id = ?'),
+  getRecentRewards: db.prepare("SELECT reward, created_at FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome != 'pending' ORDER BY created_at DESC LIMIT ?"),
+
+  upsertPolicy: db.prepare("INSERT INTO learning_policies (id, site_id, agent_id, domain, feature, weight) VALUES (?, ?, ?, ?, ?, ?) ON CONFLICT(site_id, agent_id, domain, feature) DO UPDATE SET weight = ?, update_count = update_count + 1, last_error = ?, updated_at = datetime('now')"),
+  getPolicies: db.prepare('SELECT * FROM learning_policies WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY ABS(weight) DESC'),
+  getPolicy: db.prepare('SELECT * FROM learning_policies WHERE site_id = ? AND agent_id = ? AND domain = ? AND feature = ?'),
+
+  insertPattern: db.prepare('INSERT INTO learning_patterns (id, site_id, agent_id, pattern_type, sequence, confidence) VALUES (?, ?, ?, ?, ?, ?)'),
+  findPattern: db.prepare('SELECT * FROM learning_patterns WHERE site_id = ? AND agent_id = ? AND sequence = ?'),
+  updatePattern: db.prepare("UPDATE learning_patterns SET frequency = frequency + 1, confidence = ?, last_seen = datetime('now') WHERE id = ?"),
+  getTopPatterns: db.prepare('SELECT * FROM learning_patterns WHERE site_id = ? AND agent_id = ? AND pattern_type = ? ORDER BY frequency DESC, confidence DESC LIMIT ?'),
+
+  upsertArm: db.prepare('INSERT INTO learning_bandit_arms (id, site_id, agent_id, domain, action) VALUES (?, ?, ?, ?, ?) ON CONFLICT(site_id, agent_id, domain, action) DO NOTHING'),
+  getArms: db.prepare('SELECT * FROM learning_bandit_arms WHERE site_id = ? AND agent_id = ? AND domain = ? ORDER BY ucb_score DESC'),
+  getArm: db.prepare('SELECT * FROM learning_bandit_arms WHERE site_id = ? AND agent_id = ? AND domain = ? AND action = ?'),
+  updateArm: db.prepare("UPDATE learning_bandit_arms SET pulls = pulls + 1, total_reward = total_reward + ?, avg_reward = ?, ucb_score = ?, updated_at = datetime('now') WHERE site_id = ? AND agent_id = ? AND domain = ? AND action = ?"),
+
+  insertSession: db.prepare('INSERT INTO learning_sessions (id, site_id, agent_id) VALUES (?, ?, ?)'),
+  updateSession: db.prepare("UPDATE learning_sessions SET decisions_made = ?, correct_predictions = ?, accuracy = ?, ended_at = datetime('now') WHERE id = ?"),
+  getSessionHistory: db.prepare('SELECT * FROM learning_sessions WHERE site_id = ? AND agent_id = ? ORDER BY started_at DESC LIMIT ?'),
+
+  getStats: db.prepare(`SELECT
+    (SELECT COUNT(*) FROM learning_decisions WHERE site_id = ? AND agent_id = ?) as total_decisions,
+    (SELECT COUNT(*) FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome = 'accepted') as accepted,
+    (SELECT COUNT(*) FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome = 'rejected') as rejected,
+    (SELECT AVG(reward) FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND outcome != 'pending') as avg_reward,
+    (SELECT COUNT(DISTINCT domain) FROM learning_policies WHERE site_id = ? AND agent_id = ?) as policy_domains,
+    (SELECT COUNT(*) FROM learning_patterns WHERE site_id = ? AND agent_id = ?) as total_patterns`),
+
+  deletePolicies: db.prepare('DELETE FROM learning_policies WHERE site_id = ? AND agent_id = ? AND domain = ?'),
+  deletePatterns: db.prepare('DELETE FROM learning_patterns WHERE site_id = ? AND agent_id = ?'),
+  deleteArms: db.prepare('DELETE FROM learning_bandit_arms WHERE site_id = ? AND agent_id = ? AND domain = ?'),
+  deleteDecisions: db.prepare('DELETE FROM learning_decisions WHERE site_id = ? AND agent_id = ? AND domain = ?'),
+};
+
+// ─── Core Learning API ───────────────────────────────────────────────
+
+/**
+ * Record a decision the agent is about to make, with predicted reward.
+ */
+function recordDecision(siteId, agentId, domain, action, context = {}, features = {}) {
+  const id = crypto.randomUUID();
+  const extractedFeatures = { ..._extractFeatures(context), ...features };
+  const predictedReward = _predict(siteId, agentId, domain, extractedFeatures);
+
+  stmts.insertDecision.run(id, siteId, agentId, domain, action,
+    JSON.stringify(context), predictedReward, JSON.stringify(extractedFeatures));
+
+  // Ensure bandit arm exists
+  stmts.upsertArm.run(crypto.randomUUID(), siteId, agentId, domain, action);
+
+  return { decisionId: id, predictedReward, confidence: _getConfidence(siteId, agentId, domain) };
+}
+
+/**
+ * Provide feedback on a decision — the outcome and actual reward.
+ * This is the core learning signal.
+ */
+function feedback(decisionId, outcome, reward, opts = {}) {
+  const decision = stmts.getDecision.get(decisionId);
+  if (!decision) throw new Error('Decision not found');
+
+  // ── Reward guard: clamp / block / flag malicious or anomalous rewards ──
+  let safeReward = reward;
+  let guardDecision = 'accepted';
+  if (rewardGuard) {
+    const sanitized = rewardGuard.sanitizeReward({
+      siteId: decision.site_id,
+      agentId: decision.agent_id,
+      domain: decision.domain,
+      action: decision.action,
+      reward,
+      actorId: opts.actorId,
+      approvedBy: opts.approvedBy,
+    });
+    safeReward = sanitized.reward;
+    guardDecision = sanitized.decision;
+    if (guardDecision === 'blocked') {
+      // Don't propagate to policy / bandit.
+      stmts.updateOutcome.run(outcome, safeReward, decisionId);
+      return { decisionId, blocked: true, reason: sanitized.reason, guardDecision };
+    }
+  }
+
+  stmts.updateOutcome.run(outcome, safeReward, decisionId);
+
+  const features = JSON.parse(decision.features || '{}');
+  const predError = safeReward - (decision.predicted_reward || 0);
+
+  // Update policy weights via gradient descent with temporal discount
+  _updatePolicies(decision.site_id, decision.agent_id, decision.domain, features, predError);
+
+  // Update bandit arm with actual reward
+  _updateBanditArm(decision.site_id, decision.agent_id, decision.domain, decision.action, safeReward);
+
+  // Mine patterns from recent decisions
+  _minePatterns(decision.site_id, decision.agent_id, decision.domain);
+
+  return {
+    decisionId,
+    predictionError: Math.round(predError * 1000) / 1000,
+    updatedConfidence: _getConfidence(decision.site_id, decision.agent_id, decision.domain),
+    accuracy: Math.round((1 - Math.abs(predError)) * 1000) / 1000,
+    guardDecision,
+    appliedReward: safeReward,
+  };
+}
+
+/**
+ * Batch feedback — provide multiple outcomes at once.
+ */
+function batchFeedback(feedbackList) {
+  const results = [];
+  const txn = db.transaction(() => {
+    for (const fb of feedbackList) {
+      try {
+        results.push(feedback(fb.decisionId, fb.outcome, fb.reward));
+      } catch (err) {
+        results.push({ decisionId: fb.decisionId, error: err.message });
+      }
+    }
+  });
+  txn();
+  return results;
+}
+
+/**
+ * Get the best action for a domain using learned policies + bandit scores.
+ * UCB scores are normalized to [0,1] before blending with policy prediction.
+ */
+function recommend(siteId, agentId, domain, availableActions, context = {}) {
+  const features = _extractFeatures(context);
+
+  // Get all arms to find normalization bounds
+  const allArms = stmts.getArms.all(siteId, agentId, domain);
+  const armMap = {};
+  for (const arm of allArms) armMap[arm.action] = arm;
+
+  // Normalize UCB scores to [0,1]
+  let minUCB = Infinity, maxUCB = -Infinity;
+  for (const arm of allArms) {
+    if (arm.pulls > 0) {
+      if (arm.ucb_score < minUCB) minUCB = arm.ucb_score;
+      if (arm.ucb_score > maxUCB) maxUCB = arm.ucb_score;
+    }
+  }
+  const ucbRange = maxUCB - minUCB;
+
+  const scored = availableActions.map((action) => {
+    const arm = armMap[action] || _getOrCreateArm(siteId, agentId, domain, action);
+    const policyScore = _predict(siteId, agentId, domain, { ...features, [`action:${action}`]: 1 });
+
+    // Normalize bandit score to [0,1]
+    let normalizedBandit;
+    if (arm.pulls === 0) {
+      normalizedBandit = 1.0; // unexplored arms get maximum exploration bonus
+    } else if (ucbRange > 0) {
+      normalizedBandit = (arm.ucb_score - minUCB) / ucbRange;
+    } else {
+      normalizedBandit = arm.avg_reward; // single arm — use raw avg
+    }
+
+    // Blend: as confidence grows, lean more on policy, less on exploration
+    const confidence = _getConfidence(siteId, agentId, domain);
+    const policyWeight = 0.4 + confidence * 0.4; // [0.4, 0.8]
+    const banditWeight = 1 - policyWeight;        // [0.2, 0.6]
+    const blended = policyWeight * policyScore + banditWeight * normalizedBandit;
+
+    return {
+      action,
+      score: Math.round(blended * 1000) / 1000,
+      policyScore: Math.round(policyScore * 1000) / 1000,
+      banditScore: Math.round(normalizedBandit * 1000) / 1000,
+      pulls: arm.pulls,
+      avgReward: Math.round((arm.avg_reward || 0) * 1000) / 1000,
+    };
+  });
+
+  scored.sort((a, b) => b.score - a.score);
+
+  const confidence = _getConfidence(siteId, agentId, domain);
+  const topPatterns = stmts.getTopPatterns.all(siteId, agentId, 'action_sequence', 5);
+
+  return {
+    recommended: scored[0]?.action || availableActions[0],
+    rankings: scored,
+    confidence,
+    explorationLevel: confidence < 0.3 ? 'high' : confidence < 0.6 ? 'medium' : 'low',
+    patterns: topPatterns.map((p) => ({
+      sequence: p.sequence, frequency: p.frequency, confidence: p.confidence
+    })),
+  };
+}
+
+/**
+ * Get learned preference summary for a domain.
+ */
+function getPreferences(siteId, agentId, domain) {
+  const policies = stmts.getPolicies.all(siteId, agentId, domain);
+  const decisions = stmts.getRecentDecisions.all(siteId, agentId, domain, 50);
+  const patterns = stmts.getTopPatterns.all(siteId, agentId, 'action_sequence', 10);
+
+  const accepted = decisions.filter((d) => d.outcome === 'accepted');
+  const rejected = decisions.filter((d) => d.outcome === 'rejected');
+
+  // Build preference profile from weights
+  const profile = {};
+  for (const p of policies) {
+    if (Math.abs(p.weight) > 0.05) {
+      profile[p.feature] = {
+        weight: Math.round(p.weight * 1000) / 1000,
+        direction: p.weight > 0 ? 'preferred' : 'avoided',
+        strength: Math.abs(p.weight) > 0.5 ? 'strong' : Math.abs(p.weight) > 0.2 ? 'moderate' : 'weak',
+        updates: p.update_count,
+      };
+    }
+  }
+
+  // Compute action frequencies
+  const actionFreqs = {};
+  for (const d of decisions) {
+    actionFreqs[d.action] = (actionFreqs[d.action] || 0) + 1;
+  }
+
+  return {
+    domain,
+    profile,
+    acceptRate: decisions.length > 0 ? Math.round((accepted.length / decisions.length) * 1000) / 1000 : 0,
+    rejectRate: decisions.length > 0 ? Math.round((rejected.length / decisions.length) * 1000) / 1000 : 0,
+    totalDecisions: decisions.length,
+    avgReward: decisions.length > 0
+      ? Math.round((decisions.reduce((s, d) => s + d.reward, 0) / decisions.length) * 1000) / 1000
+      : 0,
+    topActions: Object.entries(actionFreqs)
+      .sort(([, a], [, b]) => b - a)
+      .slice(0, 5)
+      .map(([action, count]) => ({ action, count, percentage: Math.round((count / decisions.length) * 100) })),
+    topPatterns: patterns.map((p) => ({ sequence: p.sequence, frequency: p.frequency })),
+    confidence: _getConfidence(siteId, agentId, domain),
+  };
+}
+
+/**
+ * Get reward history — recent rewards over time for charting.
+ */
+function getRewardHistory(siteId, agentId, limit = 30) {
+  return stmts.getRecentRewards.all(siteId, agentId, limit).reverse();
+}
+
+// ─── Learning Sessions ───────────────────────────────────────────────
+
+function startSession(siteId, agentId) {
+  const id = crypto.randomUUID();
+  stmts.insertSession.run(id, siteId, agentId);
+  return { sessionId: id };
+}
+
+function endSession(sessionId, decisionsMade, correctPredictions) {
+  const accuracy = decisionsMade > 0 ? correctPredictions / decisionsMade : 0;
+  stmts.updateSession.run(decisionsMade, correctPredictions, accuracy, sessionId);
+  return { accuracy: Math.round(accuracy * 1000) / 1000 };
+}
+
+// ─── Reset ───────────────────────────────────────────────────────────
+
+/**
+ * Reset all learned data for a specific domain.
+ */
+function resetDomain(siteId, agentId, domain) {
+  const txn = db.transaction(() => {
+    stmts.deletePolicies.run(siteId, agentId, domain);
+    stmts.deleteArms.run(siteId, agentId, domain);
+    stmts.deleteDecisions.run(siteId, agentId, domain);
+  });
+  txn();
+  return { reset: true, domain };
+}
+
+/**
+ * Reset all patterns for an agent.
+ */
+function resetPatterns(siteId, agentId) {
+  stmts.deletePatterns.run(siteId, agentId);
+  return { reset: true };
+}
+
+// ─── Stats ───────────────────────────────────────────────────────────
+
+function getStats(siteId, agentId) {
+  const row = stmts.getStats.get(siteId, agentId, siteId, agentId, siteId, agentId, siteId, agentId, siteId, agentId, siteId, agentId);
+  const sessions = stmts.getSessionHistory.all(siteId, agentId, 10);
+  const recentAccuracy = sessions.length > 0 ? sessions.reduce((s, sess) => s + sess.accuracy, 0) / sessions.length : 0;
+  const rewardHistory = stmts.getRecentRewards.all(siteId, agentId, 30).reverse();
+
+  return {
+    ...row,
+    avg_reward: row.avg_reward !== null ? Math.round(row.avg_reward * 1000) / 1000 : 0,
+    recentAccuracy: Math.round(recentAccuracy * 1000) / 1000,
+    sessionsCount: sessions.length,
+    acceptRate: row.total_decisions > 0
+      ? Math.round((row.accepted / row.total_decisions) * 1000) / 1000
+      : 0,
+    rewardHistory,
+  };
+}
+
+// ─── Internal: Prediction via Linear Model ───────────────────────────
+
+function _predict(siteId, agentId, domain, features) {
+  const policies = stmts.getPolicies.all(siteId, agentId, domain);
+  if (policies.length === 0) return 0.5; // No data yet — neutral prediction
+
+  let score = 0;
+  let matchedFeatures = 0;
+  for (const p of policies) {
+    const featureVal = features[p.feature];
+    if (featureVal !== undefined) {
+      const fv = typeof featureVal === 'number' ? featureVal : (featureVal ? 1 : 0);
+
+      // Apply temporal discount: older policies (fewer recent updates) matter less
+      const recencyBoost = p.update_count > 0 ? Math.pow(DISCOUNT_FACTOR, Math.max(0, 10 - p.update_count)) : 1;
+      score += p.weight * fv * recencyBoost;
+      matchedFeatures++;
+    }
+  }
+
+  // Sigmoid squash to [0, 1]
+  return 1 / (1 + Math.exp(-score));
+}
+
+function _updatePolicies(siteId, agentId, domain, features, error) {
+  for (const [feature, value] of Object.entries(features)) {
+    const fv = typeof value === 'number' ? value : (value ? 1 : 0);
+    if (fv === 0) continue; // Skip zero-valued features
+
+    const gradient = error * fv * LEARNING_RATE;
+    const existing = stmts.getPolicy.get(siteId, agentId, domain, feature);
+
+    // Apply weight decay to prevent unbounded growth
+    const currentWeight = existing ? existing.weight * DISCOUNT_FACTOR : 0;
+    const newWeight = currentWeight + gradient;
+
+    // Clamp weights to [-5, 5] to prevent extreme values
+    const clampedWeight = Math.max(-5, Math.min(5, newWeight));
+
+    stmts.upsertPolicy.run(
+      crypto.randomUUID(), siteId, agentId, domain, feature, clampedWeight,
+      clampedWeight, Math.abs(error)
+    );
+  }
+}
+
+// ─── Internal: Multi-Armed Bandit ────────────────────────────────────
+
+function _getOrCreateArm(siteId, agentId, domain, action) {
+  stmts.upsertArm.run(crypto.randomUUID(), siteId, agentId, domain, action);
+  const arm = stmts.getArm.get(siteId, agentId, domain, action);
+  return arm || { pulls: 0, ucb_score: 0, avg_reward: 0, total_reward: 0 };
+}
+
+function _updateBanditArm(siteId, agentId, domain, action, reward) {
+  const arm = stmts.getArm.get(siteId, agentId, domain, action);
+  if (!arm) {
+    stmts.upsertArm.run(crypto.randomUUID(), siteId, agentId, domain, action);
+    return;
+  }
+
+  const newPulls = arm.pulls + 1;
+  const newTotalReward = arm.total_reward + reward;
+  const newAvgReward = newTotalReward / newPulls;
+
+  // UCB1: avg_reward + C * sqrt(ln(N) / n_i)
+  // We need total pulls across all arms in this domain
+  const arms = stmts.getArms.all(siteId, agentId, domain);
+  const totalPulls = arms.reduce((s, a) => s + a.pulls, 0) + 1; // +1 for this pull
+
+  const exploration = UCB_EXPLORATION * Math.sqrt(Math.log(totalPulls) / newPulls);
+  const ucbScore = newAvgReward + exploration;
+
+  stmts.updateArm.run(reward, newAvgReward, ucbScore, siteId, agentId, domain, action);
+}
+
+// ─── Internal: Pattern Mining ────────────────────────────────────────
+
+function _minePatterns(siteId, agentId, domain) {
+  const decisions = stmts.getRecentDecisions.all(siteId, agentId, domain, 20);
+  if (decisions.length < 3) return;
+
+  // Extract action sequences of length 2-5
+  for (let len = 2; len <= Math.min(MAX_SEQUENCE_LENGTH, decisions.length); len++) {
+    const sequence = decisions.slice(0, len).map((d) => d.action).reverse().join(' → ');
+    const existing = stmts.findPattern.get(siteId, agentId, sequence);
+
+    if (existing) {
+      // Asymptotic approach to 1.0 — confidence grows slower as it increases
+      const newConf = Math.min(0.99, existing.confidence + 0.05 * (1 - existing.confidence));
+      stmts.updatePattern.run(newConf, existing.id);
+    } else {
+      stmts.insertPattern.run(crypto.randomUUID(), siteId, agentId, 'action_sequence', sequence, 0.3);
+    }
+  }
+}
+
+// ─── Internal: Feature Extraction ────────────────────────────────────
+
+function _extractFeatures(context) {
+  const features = {};
+
+  if (context.price !== undefined) {
+    features.price = context.price;
+    // Bucketize price for discrete learning
+    if (context.price < 10) features['price_bucket:cheap'] = 1;
+    else if (context.price < 50) features['price_bucket:moderate'] = 1;
+    else if (context.price < 200) features['price_bucket:premium'] = 1;
+    else features['price_bucket:luxury'] = 1;
+  }
+  if (context.quantity !== undefined) features.quantity = context.quantity;
+  if (context.discount !== undefined) {
+    features.discount = context.discount;
+    features.has_discount = context.discount > 0 ? 1 : 0;
+  }
+  if (context.rating !== undefined) {
+    features.rating = context.rating;
+    features.high_rated = context.rating >= 4.0 ? 1 : 0;
+  }
+  if (context.category) features[`category:${context.category}`] = 1;
+  if (context.brand) features[`brand:${context.brand}`] = 1;
+  if (context.timeOfDay !== undefined) {
+    features.morning = context.timeOfDay < 12 ? 1 : 0;
+    features.afternoon = context.timeOfDay >= 12 && context.timeOfDay < 18 ? 1 : 0;
+    features.evening = context.timeOfDay >= 18 ? 1 : 0;
+  }
+  if (context.isRepeat !== undefined) features.repeat_visit = context.isRepeat ? 1 : 0;
+  if (context.urgency !== undefined) features.urgency = context.urgency;
+  if (context.inStock !== undefined) features.in_stock = context.inStock ? 1 : 0;
+
+  // Pass through any raw numeric features
+  for (const [k, v] of Object.entries(context)) {
+    if (features[k] === undefined && typeof v === 'number') {
+      features[k] = v;
+    }
+  }
+
+  return features;
+}
+
+// ─── Internal: Confidence Estimation ─────────────────────────────────
+
+function _getConfidence(siteId, agentId, domain) {
+  const decisions = stmts.getRecentDecisions.all(siteId, agentId, domain, 50);
+  if (decisions.length === 0) return 0;
+
+  const withOutcome = decisions.filter((d) => d.outcome !== 'pending');
+  if (withOutcome.length === 0) return MIN_CONFIDENCE;
+
+  // Volume component: log scale, saturates around 30 decisions
+  const volumeConf = Math.min(1, withOutcome.length / 30);
+
+  // Accuracy component: how close predictions were to actual rewards
+  let accuracySum = 0;
+  for (const d of withOutcome) {
+    if (d.predicted_reward !== null) {
+      const error = Math.abs(d.reward - d.predicted_reward);
+      accuracySum += Math.max(0, 1 - error);
+    }
+  }
+  const accuracyConf = withOutcome.length > 0 ? accuracySum / withOutcome.length : 0.5;
+
+  // Recency component: exponential decay based on age of newest data
+  const latestTs = new Date(withOutcome[0].created_at).getTime();
+  const ageHours = (Date.now() - latestTs) / 3600000;
+  const recencyConf = Math.exp(-DECAY_RATE * ageHours);
+
+  return Math.max(MIN_CONFIDENCE, Math.min(0.99,
+    volumeConf * 0.3 + accuracyConf * 0.5 + recencyConf * 0.2
+  ));
+}
+
+module.exports = {
+  recordDecision, feedback, batchFeedback, recommend, getPreferences,
+  getRewardHistory, startSession, endSession,
+  resetDomain, resetPatterns, getStats,
+};