@aleph-ai/tinyaleph 1.3.0 → 1.4.0

package/core/rformer-crt.js

@@ -0,0 +1,892 @@
+/**
+ * CRT-Enhanced ResoFormer Layers
+ * 
+ * Integrates Chinese Remainder Theorem reconstruction with Birkhoff polytope
+ * constraints and homology-based regularization into the ResoFormer architecture.
+ * 
+ * Key additions:
+ * - CRTResonantAttention: Modular attention with per-modulus Birkhoff projection
+ * - HomologyRegularizedBlock: Detects semantic inconsistencies as topological holes
+ * - CRTResoFormer: Full model with homology loss integration
+ * 
+ * Mathematical Foundation:
+ * - Encode: r_k = softmax(W_k h + b_k) ∈ Δ(ℤ/p_k) for coprime moduli p_k
+ * - Reconstruct: L̂ = Σ_k E[r_k] · (P/p_k) · (P/p_k)^{-1} mod P via CRT
+ * - Project: A_k ∈ Birkhoff(n) via Sinkhorn-Knopp
+ * - Regularize: ℒ_homology = Σ_{cycles} f(cycle) detects Ker(ℛ) holes
+ */
+
+'use strict';
+
+const {
+  Quaternion,
+  SparsePrimeState,
+  resonanceScore,
+  resonantAttention,
+  hamiltonCompose,
+  computeCoherence
+} = require('./rformer');
+
+const {
+  ResonantMultiHeadAttention,
+  PrimeFFN,
+  PrimeLayerNorm,
+  PositionalPrimeEncoding,
+  ResoFormerBlock
+} = require('./rformer-layers');
+
+const {
+  CRTReconstructor,
+  BirkhoffProjector,
+  HomologyLoss,
+  CRTModularLayer,
+  CRTFusedAttention,
+  CoprimeSelector,
+  ResidueEncoder
+} = require('./crt-homology');
+
+const { Complex, PrimeState } = require('./hilbert');
+const { firstNPrimes, isPrime } = require('./prime');
+
+/**
+ * CRTResonantAttention - Multi-head attention with CRT-fused modular structure
+ * 
+ * Each head computes attention in a different modular field, then
+ * fuses via CRT reconstruction. Attention matrices are projected
+ * onto the Birkhoff polytope for doubly-stochastic structure.
+ */
+class CRTResonantAttention {
+  /**
+   * @param {object} config
+   * @param {number} config.numHeads - Number of attention heads (maps to moduli)
+   * @param {number} [config.numPrimes=4096] - Size of prime vocabulary
+   * @param {number} [config.activeK=32] - Sparsity per state
+   * @param {number} [config.sinkhornIterations=10] - Iterations for Birkhoff projection
+   * @param {number} [config.temperature=1.0] - Softmax temperature
+   */
+  constructor(config) {
+    this.numHeads = config.numHeads || 8;
+    this.numPrimes = config.numPrimes || 4096;
+    this.activeK = config.activeK || 32;
+    this.temperature = config.temperature || 1.0;
+    
+    // Select coprime moduli for CRT reconstruction
+    this.coprimeSelector = new CoprimeSelector(this.numHeads);
+    this.moduli = this.coprimeSelector.selectMinimal();
+    
+    // CRT components
+    this.crtReconstructor = new CRTReconstructor(this.moduli);
+    this.birkhoffProjector = new BirkhoffProjector({
+      maxIterations: config.sinkhornIterations || 10,
+      tolerance: config.birkhoffTolerance || 1e-3
+    });
+    
+    // Per-head weights [alpha, beta, gamma] for resonance scoring
+    this.headWeights = config.headWeights || this._defaultHeadWeights();
+    
+    // Output scaling
+    this.outputScale = config.outputScale || 1.0 / Math.sqrt(this.numHeads);
+  }
+  
+  /**
+   * Generate default head weights with moduli-specific emphasis
+   * @private
+   */
+  _defaultHeadWeights() {
+    const weights = [];
+    for (let h = 0; h < this.numHeads; h++) {
+      const modulus = this.moduli[h];
+      // Weight configuration varies with modulus size
+      // Larger moduli get more phase weight, smaller get more Jaccard
+      const t = Math.log(modulus) / Math.log(this.moduli[this.numHeads - 1]);
+      
+      const alpha = 0.5 - 0.2 * t;  // Jaccard weight
+      const beta = 0.3;              // Quaternion (constant)
+      const gamma = 0.2 + 0.2 * t;   // Phase weight
+      
+      weights.push([alpha, beta, gamma]);
+    }
+    return weights;
+  }
+  
+  /**
+   * Apply CRT-fused multi-head attention
+   * 
+   * @param {SparsePrimeState} query - Query state
+   * @param {SparsePrimeState[]} keys - Key states
+   * @param {SparsePrimeState[]} values - Value states
+   * @returns {object} { result, headOutputs, attentionWeights, crtResidues, homologyInfo }
+   */
+  forward(query, keys, values) {
+    const n = keys.length;
+    if (n === 0) {
+      return { 
+        result: query, 
+        headOutputs: [], 
+        attentionWeights: [],
+        crtResidues: [],
+        homologyInfo: { hasHoles: false, bettiNumbers: [1, 0] }
+      };
+    }
+    
+    const headOutputs = [];
+    const allWeights = [];
+    const crtResidues = [];
+    
+    // Apply each head (modular attention)
+    for (let h = 0; h < this.numHeads; h++) {
+      const modulus = this.moduli[h];
+      const [alpha, beta, gamma] = this.headWeights[h];
+      
+      // Compute head-specific attention with modular structure
+      const headResult = this._modularHeadAttention(
+        query, keys, values, modulus, alpha, beta, gamma
+      );
+      
+      headOutputs.push(headResult.result);
+      allWeights.push(headResult.birkhoffWeights);
+      crtResidues.push(headResult.residue);
+    }
+    
+    // Fuse head outputs via CRT reconstruction
+    const fusedResult = this._crtFuseHeads(headOutputs);
+    
+    // Detect homology holes (cycles in kernel)
+    const homologyInfo = this._detectHomologyHoles(crtResidues, allWeights);
+    
+    return {
+      result: fusedResult,
+      headOutputs,
+      attentionWeights: allWeights,
+      crtResidues,
+      homologyInfo
+    };
+  }
+  
+  /**
+   * Single head attention with modular structure and Birkhoff projection
+   * @private
+   */
+  _modularHeadAttention(query, keys, values, modulus, alpha, beta, gamma) {
+    const n = keys.length;
+    
+    // Compute resonance scores
+    const scores = keys.map((k, idx) => {
+      const primesQ = new Set(query.getActivePrimes());
+      const primesK = new Set(k.getActivePrimes());
+      
+      // Jaccard
+      const intersection = new Set([...primesQ].filter(p => primesK.has(p)));
+      const union = new Set([...primesQ, ...primesK]);
+      const jaccard = intersection.size / (union.size || 1);
+      
+      if (intersection.size === 0) {
+        return alpha * jaccard;
+      }
+      
+      // Quaternion alignment
+      let quatSum = 0;
+      for (const p of intersection) {
+        const qi = query.get(p).quaternion;
+        const qk = k.get(p).quaternion;
+        quatSum += Math.abs(qi.dot(qk));
+      }
+      const quatAlign = quatSum / intersection.size;
+      
+      // Phase coherence (modular)
+      let phaseSum = 0;
+      for (const p of intersection) {
+        const phaseQ = query.get(p).amplitude.phase();
+        const phaseK = k.get(p).amplitude.phase();
+        // Modular phase difference
+        const phaseDiff = ((phaseQ - phaseK) % (2 * Math.PI / modulus)) * modulus;
+        phaseSum += Math.cos(phaseDiff);
+      }
+      const phaseCoherence = (phaseSum / intersection.size + 1) / 2;
+      
+      return alpha * jaccard + beta * quatAlign + gamma * phaseCoherence;
+    });
+    
+    // Create attention matrix (single row for query)
+    // Expand to n×n by computing pairwise scores among keys
+    const attentionMatrix = [];
+    for (let i = 0; i < n; i++) {
+      const row = keys.map((k, j) => resonanceScore(keys[i], k, alpha, beta, gamma));
+      // Apply modular structure
+      for (let j = 0; j < n; j++) {
+        row[j] = row[j] % 1.0;  // Keep in [0, 1) range
+      }
+      attentionMatrix.push(row);
+    }
+    
+    // Project onto Birkhoff polytope (doubly-stochastic)
+    const birkhoffResult = this.birkhoffProjector.project(attentionMatrix);
+    const birkhoffMatrix = birkhoffResult && birkhoffResult.matrix ? birkhoffResult.matrix : null;
+    
+    // Extract query-specific weights (first row after projection)
+    // For proper attention, use the scores projected through Birkhoff
+    const softmaxWeights = this._softmax(scores.map(s => s / this.temperature));
+    
+    // Blend Birkhoff structure with softmax attention
+    const blendFactor = 0.3;  // How much Birkhoff structure to incorporate
+    const blendedWeights = softmaxWeights.map((w, i) => {
+      // Safely access Birkhoff weights with full null checking
+      if (birkhoffMatrix && Array.isArray(birkhoffMatrix) && birkhoffMatrix.length > 0) {
+        const birkhoffRow = birkhoffMatrix[0];
+        if (Array.isArray(birkhoffRow) && i < birkhoffRow.length && typeof birkhoffRow[i] === 'number') {
+          return (1 - blendFactor) * w + blendFactor * birkhoffRow[i];
+        }
+      }
+      return w;  // Fall back to pure softmax if Birkhoff unavailable
+    });
+    
+    // Normalize
+    const weightSum = blendedWeights.reduce((a, b) => a + b, 0);
+    const normalizedWeights = blendedWeights.map(w => w / (weightSum || 1));
+    
+    // Compute residue for CRT reconstruction
+    // Residue = weighted sum of value indices mod modulus
+    let residue = 0;
+    for (let i = 0; i < n; i++) {
+      residue = (residue + normalizedWeights[i] * i) % modulus;
+    }
+    
+    // Weighted combination of values
+    const result = new SparsePrimeState(this.numPrimes, this.activeK);
+    
+    for (let i = 0; i < n; i++) {
+      const w = normalizedWeights[i];
+      for (const [p, act] of values[i].activations) {
+        const current = result.get(p);
+        const newAmp = current.amplitude.add(act.amplitude.scale(w));
+        const newQuat = current.quaternion.add(act.quaternion.scale(w));
+        result.set(p, newAmp, newQuat.normalize());
+      }
+    }
+    
+    return { 
+      result: result.normalize(), 
+      birkhoffWeights: normalizedWeights, 
+      residue,
+      convergenceInfo: birkhoffResult.convergenceInfo
+    };
+  }
+  
+  /**
+   * Softmax function
+   * @private
+   */
+  _softmax(scores) {
+    const maxScore = Math.max(...scores);
+    const expScores = scores.map(s => Math.exp(s - maxScore));
+    const sumExp = expScores.reduce((a, b) => a + b, 0);
+    return expScores.map(e => e / (sumExp || 1));
+  }
+  
+  /**
+   * Fuse head outputs via CRT reconstruction
+   * @private
+   */
+  _crtFuseHeads(headOutputs) {
+    const result = new SparsePrimeState(this.numPrimes, this.activeK);
+    
+    // Collect all active primes across heads
+    const allPrimes = new Set();
+    for (const head of headOutputs) {
+      for (const p of head.getActivePrimes()) {
+        allPrimes.add(p);
+      }
+    }
+    
+    // For each prime, fuse activations via CRT-style combination
+    for (const p of allPrimes) {
+      // Collect residues (amplitude norms) from each head
+      const residues = headOutputs.map(h => h.get(p).amplitude.norm());
+      
+      // Compute reconstructed norm using weighted average (simpler than full CRT)
+      // This avoids issues with probability distributions of varying sizes
+      let reconstructedNorm = 0;
+      let totalWeight = 0;
+      for (let h = 0; h < headOutputs.length; h++) {
+        const w = 1 / (h + 1);  // Decaying weights by head index
+        reconstructedNorm += residues[h] * w;
+        totalWeight += w;
+      }
+      reconstructedNorm = totalWeight > 0 ? reconstructedNorm / totalWeight : 0;
+      
+      // Fuse phases and quaternions via weighted average
+      let phaseSum = 0;
+      let quatSum = Quaternion.zero();
+      let weightSum = 0;
+      
+      for (let h = 0; h < headOutputs.length; h++) {
+        const act = headOutputs[h].get(p);
+        const w = act.amplitude.norm();
+        phaseSum += w * act.amplitude.phase();
+        quatSum = quatSum.add(act.quaternion.scale(w));
+        weightSum += w;
+      }
+      
+      const avgPhase = weightSum > 0 ? phaseSum / weightSum : 0;
+      const avgQuat = weightSum > 0 ? quatSum.scale(1/weightSum).normalize() : Quaternion.one();
+      
+      // Scale by output factor
+      const finalAmp = Complex.fromPolar(
+        reconstructedNorm * this.outputScale / this.numHeads,
+        avgPhase
+      );
+      
+      result.set(p, finalAmp, avgQuat);
+    }
+    
+    return result.normalize();
+  }
+  
+  /**
+   * Detect homology holes in the kernel
+   * @private
+   */
+  _detectHomologyHoles(residues, weights) {
+    // Build adjacency from attention weights
+    const n = weights[0]?.length || 0;
+    if (n < 2) {
+      return { hasHoles: false, bettiNumbers: [1, 0], cycles: [] };
+    }
+    
+    // Create averaged adjacency matrix
+    const adjacency = [];
+    for (let i = 0; i < n; i++) {
+      adjacency.push(new Array(n).fill(0));
+    }
+    
+    for (const headWeights of weights) {
+      for (let i = 0; i < Math.min(headWeights.length, n); i++) {
+        for (let j = 0; j < Math.min(headWeights.length, n); j++) {
+          adjacency[i][j] += headWeights[i] * headWeights[j] / weights.length;
+        }
+      }
+    }
+    
+    // Threshold adjacency to binary
+    const threshold = 0.1;
+    for (let i = 0; i < n; i++) {
+      for (let j = 0; j < n; j++) {
+        adjacency[i][j] = adjacency[i][j] > threshold ? 1 : 0;
+      }
+    }
+    
+    // Compute error terms (CRT inconsistencies)
+    const errors = [];
+    for (let i = 0; i < residues.length; i++) {
+      // Error = deviation from expected value
+      errors.push(Math.abs(residues[i] - Math.floor(residues[i])));
+    }
+    
+    // Build kernel (high-error nodes)
+    const errorThreshold = 0.1;
+    const kernel = errors.map((e, i) => ({ index: i, error: e, inKernel: e > errorThreshold }));
+    const kernelNodes = kernel.filter(k => k.inKernel).map(k => k.index);
+    
+    // Detect cycles in kernel subgraph
+    const cycles = this._findCyclesInSubgraph(adjacency, kernelNodes);
+    
+    // Compute Betti numbers
+    const beta0 = this._countConnectedComponents(adjacency, kernelNodes);
+    const beta1 = cycles.length;
+    
+    return {
+      hasHoles: beta1 > 0,
+      bettiNumbers: [beta0, beta1],
+      cycles,
+      kernelNodes,
+      errors
+    };
+  }
+  
+  /**
+   * Find cycles in a subgraph
+   * @private
+   */
+  _findCyclesInSubgraph(adjacency, nodes) {
+    const cycles = [];
+    const n = nodes.length;
+    
+    if (n < 3) return cycles;
+    
+    // Simple cycle detection: look for triangles
+    for (let i = 0; i < n; i++) {
+      for (let j = i + 1; j < n; j++) {
+        for (let k = j + 1; k < n; k++) {
+          const a = nodes[i], b = nodes[j], c = nodes[k];
+          if (a < adjacency.length && b < adjacency.length && c < adjacency.length) {
+            if (adjacency[a][b] && adjacency[b][c] && adjacency[c][a]) {
+              cycles.push([a, b, c]);
+            }
+          }
+        }
+      }
+    }
+    
+    return cycles;
+  }
+  
+  /**
+   * Count connected components in subgraph
+   * @private
+   */
+  _countConnectedComponents(adjacency, nodes) {
+    if (nodes.length === 0) return 0;
+    
+    const visited = new Set();
+    let components = 0;
+    
+    const dfs = (node) => {
+      visited.add(node);
+      for (const neighbor of nodes) {
+        if (!visited.has(neighbor) && 
+            node < adjacency.length && neighbor < adjacency.length &&
+            adjacency[node][neighbor]) {
+          dfs(neighbor);
+        }
+      }
+    };
+    
+    for (const node of nodes) {
+      if (!visited.has(node)) {
+        dfs(node);
+        components++;
+      }
+    }
+    
+    return components;
+  }
+  
+  /**
+   * Set head weights (for training)
+   */
+  setHeadWeights(headIdx, weights) {
+    if (headIdx >= 0 && headIdx < this.numHeads) {
+      this.headWeights[headIdx] = weights;
+    }
+  }
+  
+  /**
+   * Get all parameters (for serialization)
+   */
+  getParameters() {
+    return {
+      numHeads: this.numHeads,
+      moduli: this.moduli,
+      headWeights: this.headWeights,
+      temperature: this.temperature,
+      outputScale: this.outputScale
+    };
+  }
+}
+
+/**
+ * HomologyRegularizedBlock - ResoFormer block with homology-based regularization
+ * 
+ * Extends ResoFormerBlock with:
+ * - CRT-fused attention
+ * - Homology loss computation
+ * - Kernel detection for semantic inconsistencies
+ */
+class HomologyRegularizedBlock {
+  /**
+   * @param {object} config
+   * @param {number} [config.numHeads=8] - Number of attention heads
+   * @param {number} [config.hiddenDim=256] - FFN hidden dimension
+   * @param {number} [config.numPrimes=4096] - Prime vocabulary size
+   * @param {number} [config.activeK=32] - Sparsity parameter
+   * @param {number} [config.dropout=0.1] - Dropout probability
+   * @param {boolean} [config.preNorm=true] - Pre-norm or post-norm
+   * @param {number} [config.homologyWeight=0.1] - Weight for homology loss
+   */
+  constructor(config = {}) {
+    this.preNorm = config.preNorm ?? true;
+    this.numPrimes = config.numPrimes || 4096;
+    this.activeK = config.activeK || 32;
+    this.homologyWeight = config.homologyWeight || 0.1;
+    
+    // CRT-enhanced attention
+    this.attention = new CRTResonantAttention({
+      numHeads: config.numHeads || 8,
+      numPrimes: this.numPrimes,
+      activeK: this.activeK,
+      temperature: config.attentionTemperature || 1.0,
+      sinkhornIterations: config.sinkhornIterations || 10
+    });
+    
+    // FFN and norms from standard block
+    this.ffn = new PrimeFFN({
+      hiddenDim: config.hiddenDim || 256,
+      numPrimes: this.numPrimes,
+      activation: config.activation || 'gelu',
+      dropout: config.dropout || 0.1
+    });
+    
+    this.norm1 = new PrimeLayerNorm();
+    this.norm2 = new PrimeLayerNorm();
+    
+    // Homology loss computer
+    this.homologyLoss = new HomologyLoss({
+      errorThreshold: config.errorThreshold || 0.1,
+      alpha: config.homologyAlpha || 0.5,
+      beta: config.homologyBeta || 1.0,
+      gamma: config.homologyGamma || 0.5
+    });
+    
+    this.dropoutRate = config.dropout || 0.1;
+    this.training = false;
+  }
+  
+  /**
+   * Forward pass with homology regularization
+   * 
+   * @param {SparsePrimeState} x - Input state
+   * @param {SparsePrimeState[]} context - Context states for attention
+   * @returns {object} { output, attentionWeights, homologyInfo, loss }
+   */
+  forward(x, context = null) {
+    const keys = context || [x];
+    const values = context || [x];
+    
+    let attnInput, ffnInput;
+    let homologyLossValue = 0;
+    let homologyInfo = null;
+    
+    if (this.preNorm) {
+      // Pre-norm: Norm -> Attn -> Add -> Norm -> FFN -> Add
+      attnInput = this.norm1.forward(x);
+      const attnOut = this.attention.forward(
+        attnInput,
+        keys.map(k => this.norm1.forward(k)),
+        values.map(v => this.norm1.forward(v))
+      );
+      
+      homologyInfo = attnOut.homologyInfo;
+      
+      // Compute homology loss if there are holes
+      if (homologyInfo.hasHoles && homologyInfo.cycles && homologyInfo.cycles.length > 0) {
+        // Compute loss directly from detected cycles
+        // f(cycle) = |cycle|^α * β^γ * Σ errors
+        for (const cycle of homologyInfo.cycles) {
+          const cycleLength = cycle.length;
+          // Sum of errors for nodes in cycle
+          let errorSum = 0;
+          for (const nodeIdx of cycle) {
+            if (homologyInfo.errors && nodeIdx < homologyInfo.errors.length) {
+              errorSum += this.homologyLoss.sigmoid(homologyInfo.errors[nodeIdx] - 0.1);
+            }
+          }
+          homologyLossValue += errorSum *
+            Math.pow(cycleLength, this.homologyLoss.alpha) *
+            Math.pow(this.homologyLoss.beta, this.homologyLoss.gamma);
+        }
+      }
+      
+      // Residual connection
+      const afterAttn = this._add(x, this._dropout(attnOut.result));
+      
+      // FFN
+      ffnInput = this.norm2.forward(afterAttn);
+      const ffnOut = this.ffn.forward(ffnInput);
+      
+      // Residual connection
+      const output = this._add(afterAttn, this._dropout(ffnOut));
+      
+      return {
+        output,
+        attentionWeights: attnOut.attentionWeights,
+        homologyInfo,
+        loss: this.homologyWeight * homologyLossValue,
+        crtResidues: attnOut.crtResidues
+      };
+      
+    } else {
+      // Post-norm
+      const attnOut = this.attention.forward(x, keys, values);
+      homologyInfo = attnOut.homologyInfo;
+      
+      const afterAttn = this.norm1.forward(this._add(x, this._dropout(attnOut.result)));
+      const ffnOut = this.ffn.forward(afterAttn);
+      const output = this.norm2.forward(this._add(afterAttn, this._dropout(ffnOut)));
+      
+      return {
+        output,
+        attentionWeights: attnOut.attentionWeights,
+        homologyInfo,
+        loss: 0,
+        crtResidues: attnOut.crtResidues
+      };
+    }
+  }
+  
+  /**
+   * Add two sparse states (residual connection)
+   * @private
+   */
+  _add(a, b) {
+    const result = new SparsePrimeState(this.numPrimes, this.activeK);
+    const allPrimes = new Set([...a.getActivePrimes(), ...b.getActivePrimes()]);
+    
+    for (const p of allPrimes) {
+      const actA = a.get(p);
+      const actB = b.get(p);
+      
+      const newAmp = actA.amplitude.add(actB.amplitude);
+      const newQuat = actA.quaternion.add(actB.quaternion);
+      
+      result.set(p, newAmp, newQuat.normalize());
+    }
+    
+    return result.normalize();
+  }
+  
+  /**
+   * Apply dropout
+   * @private
+   */
+  _dropout(state) {
+    if (!this.training || this.dropoutRate <= 0) return state;
+    
+    const result = new SparsePrimeState(this.numPrimes, this.activeK);
+    const scale = 1 / (1 - this.dropoutRate);
+    
+    for (const [p, act] of state.activations) {
+      if (Math.random() >= this.dropoutRate) {
+        result.set(p, act.amplitude.scale(scale), act.quaternion);
+      }
+    }
+    
+    return result;
+  }
+  
+  /**
+   * Set training mode
+   */
+  train(mode = true) {
+    this.training = mode;
+    this.ffn.train(mode);
+    return this;
+  }
+  
+  /**
+   * Set evaluation mode
+   */
+  eval() {
+    return this.train(false);
+  }
+}
+
+/**
+ * CRTResoFormer - Complete CRT-enhanced ResoFormer model
+ * 
+ * Stacks HomologyRegularizedBlocks with:
+ * - CRT-fused attention at each layer
+ * - Accumulated homology loss for training
+ * - Semantic hole detection across layers
+ */
+class CRTResoFormer {
+  /**
+   * @param {object} config
+   * @param {number} [config.numLayers=6] - Number of transformer blocks
+   * @param {number} [config.numHeads=8] - Attention heads per block
+   * @param {number} [config.hiddenDim=256] - FFN hidden dimension
+   * @param {number} [config.numPrimes=4096] - Prime vocabulary size
+   * @param {number} [config.activeK=32] - Sparsity parameter
+   * @param {number} [config.dropout=0.1] - Dropout probability
+   * @param {boolean} [config.usePositionalEncoding=true] - Add position encoding
+   * @param {number} [config.homologyWeight=0.1] - Weight for homology loss
+   */
+  constructor(config = {}) {
+    this.numLayers = config.numLayers || 6;
+    this.numPrimes = config.numPrimes || 4096;
+    this.activeK = config.activeK || 32;
+    
+    // Position encoding
+    this.usePositionalEncoding = config.usePositionalEncoding ?? true;
+    if (this.usePositionalEncoding) {
+      this.posEncoder = new PositionalPrimeEncoding({
+        numPrimes: this.numPrimes,
+        activeK: this.activeK
+      });
+    }
+    
+    // Stack of CRT-enhanced blocks
+    this.blocks = [];
+    for (let i = 0; i < this.numLayers; i++) {
+      this.blocks.push(new HomologyRegularizedBlock({
+        numHeads: config.numHeads || 8,
+        hiddenDim: config.hiddenDim || 256,
+        numPrimes: this.numPrimes,
+        activeK: this.activeK,
+        dropout: config.dropout || 0.1,
+        preNorm: config.preNorm ?? true,
+        homologyWeight: config.homologyWeight || 0.1,
+        sinkhornIterations: config.sinkhornIterations || 10
+      }));
+    }
+    
+    // Final normalization
+    this.finalNorm = new PrimeLayerNorm();
+  }
+  /**
+   * Forward pass through all layers
+   *
+   * @param {SparsePrimeState|SparsePrimeState[]} input - Input state(s)
+   * @returns {object} { output, layerOutputs, attentionMaps, homologyReport, totalLoss }
+   */
+  forward(input) {
+    const isSequence = Array.isArray(input);
+    let states = isSequence ? input : [input];
+    
+    // Add position encoding
+    if (this.usePositionalEncoding) {
+      states = this.posEncoder.encodeSequence(states);
+    }
+    
+    const layerOutputs = [];
+    const attentionMaps = [];
+    const homologyReports = [];
+    let totalLoss = 0;
+    
+    // Process through each block
+    for (let layer = 0; layer < this.numLayers; layer++) {
+      const block = this.blocks[layer];
+      const newStates = [];
+      const layerAttention = [];
+      const layerHomology = [];
+      
+      for (let i = 0; i < states.length; i++) {
+        const { output, attentionWeights, homologyInfo, loss } = block.forward(states[i], states);
+        newStates.push(output);
+        layerAttention.push(attentionWeights);
+        layerHomology.push(homologyInfo);
+        totalLoss += loss;
+      }
+      
+      states = newStates;
+      layerOutputs.push([...states]);
+      attentionMaps.push(layerAttention);
+      homologyReports.push(layerHomology);
+    }
+    
+    // Final normalization
+    states = states.map(s => this.finalNorm.forward(s));
+    
+    // Aggregate homology report
+    const aggregateHomology = this._aggregateHomologyReports(homologyReports);
+    
+    return {
+      output: isSequence ? states : states[0],
+      layerOutputs,
+      attentionMaps,
+      homologyReport: aggregateHomology,
+      totalLoss
+    };
+  }
+  
+  /**
+   * Aggregate homology reports across layers
+   * @private
+   */
+  _aggregateHomologyReports(reports) {
+    let totalHoles = 0;
+    let maxBeta1 = 0;
+    const allCycles = [];
+    
+    for (const layerReports of reports) {
+      for (const report of layerReports) {
+        if (report.hasHoles) {
+          totalHoles++;
+          maxBeta1 = Math.max(maxBeta1, report.bettiNumbers[1]);
+          allCycles.push(...(report.cycles || []));
+        }
+      }
+    }
+    
+    return {
+      hasHoles: totalHoles > 0,
+      totalHolesDetected: totalHoles,
+      maxBettiNumber: maxBeta1,
+      uniqueCycles: allCycles.length,
+      layerReports: reports
+    };
+  }
+  
+  /**
+   * Set training mode
+   */
+  train(mode = true) {
+    for (const block of this.blocks) {
+      block.train(mode);
+    }
+    return this;
+  }
+  
+  /**
+   * Set evaluation mode
+   */
+  eval() {
+    return this.train(false);
+  }
+  
+  /**
+   * Get total parameter count (approximate)
+   */
+  getParameterCount() {
+    const perBlock = this.numLayers * (
+      8 * 3 +  // Attention head weights
+      4 +      // FFN weights
+      2 +      // LayerNorm
+      1        // Homology params
+    );
+    return perBlock + (this.usePositionalEncoding ? this.activeK * 4 : 0);
+  }
+  
+  /**
+   * Get CRT configuration
+   */
+  getCRTConfig() {
+    if (this.blocks.length === 0) return null;
+    return this.blocks[0].attention.getParameters();
+  }
+}
+
+/**
+ * createCRTResoFormer - Factory function with sensible defaults
+ *
+ * @param {object} config - Configuration options
+ * @returns {CRTResoFormer} Configured model
+ */
+function createCRTResoFormer(config = {}) {
+  return new CRTResoFormer({
+    numLayers: config.numLayers || 6,
+    numHeads: config.numHeads || 8,
+    hiddenDim: config.hiddenDim || 256,
+    numPrimes: config.numPrimes || 4096,
+    activeK: config.activeK || 32,
+    dropout: config.dropout || 0.1,
+    usePositionalEncoding: config.usePositionalEncoding ?? true,
+    homologyWeight: config.homologyWeight || 0.1,
+    sinkhornIterations: config.sinkhornIterations || 10,
+    preNorm: config.preNorm ?? true
+  });
+}
+
+module.exports = {
+  // CRT-enhanced attention
+  CRTResonantAttention,
+  
+  // Homology-regularized block
+  HomologyRegularizedBlock,
+  
+  // Complete model
+  CRTResoFormer,
+  
+  // Factory function
+  createCRTResoFormer
+};
+