session: reverse-engineer reasoning via causal edge diffing (NARS + attention + MoE gate)

Author: AdaWorldAPI

.claude/prompts/SESSION_REASONING_REVERSE_ENG.md

@@ -0,0 +1,224 @@
+# SESSION: Reverse-Engineer Reasoning via Causal Edge Diffing
+
+## MISSION
+
+Extract the structural geometry of "how to think" from:
+1. Llama 4 Maverick MoE gate projections (routing topology)
+2. Qwen3.5 base→distilled attention diffs (reasoning circuit)
+3. Cross-model comparison (scale-invariant reasoning atoms)
+
+Feed into NARS truth values on causal edges. First real training data
+for the NARS stack.
+
+## READ FIRST
+
+```bash
+cat src/hpc/gguf_indexer.rs    # stream_index_gguf_bf16, classify_tensor
+cat src/hpc/nars.rs            # TruthValue, revision, evidence
+cat src/hpc/bgz17_bridge.rs    # Base17 type, L1 distance
+cat src/hpc/causality.rs       # CausalEdge if it exists
+```
+
+## PHASE 1: Index All Models (Q8_0, streaming)
+
+Five GGUF files, all single-shard, ~105 GB total:
+
+```
+unsloth/Qwen3.5-27B-GGUF
+  → Qwen3.5-27B-Q8_0.gguf                  28.59 GB  (base)
+
+Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF
+  → Qwen3.5-27B.Q8_0.gguf                  28.59 GB  (distilled v1)
+
+Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF
+  → Qwen3.5-27B.Q8_0.gguf                  28.59 GB  (distilled v2)
+
+unsloth/Qwen3.5-9B-GGUF
+  → Qwen3.5-9B-Q8_0.gguf                    9.52 GB  (base 9B)
+
+Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-GGUF
+  → Qwen3.5-9B.Q8_0.gguf                    9.52 GB  (distilled 9B)
+```
+
+Use `stream_index_gguf` (f32 path — Q8_0 needs actual dequantization).
+Output: 5 bgz7 files with per-tensor, per-row Base17 projections.
+
+## PHASE 2: Attention Diff (the reasoning circuit)
+
+For each tensor pair (base vs distilled), matched by name:
+
+```rust
+// Pseudocode — actual implementation in causal_diff.rs
+for (name, base_rows, dist_rows) in matched_tensors(base_bgz7, dist_bgz7) {
+    let layer_type = classify_tensor(name);
+    
+    for (row_idx, (b, d)) in base_rows.zip(dist_rows).enumerate() {
+        let distance = b.l1(&d);
+        
+        if distance > threshold {
+            let edge = CausalEdge64 {
+                subject: palette_index(b),     // base archetype
+                verb: BECOMES,                  // structural transformation
+                object: palette_index(d),       // distilled archetype
+                truth: TruthValue {
+                    frequency: distance as f32 / max_l1 as f32,
+                    confidence: 1.0 / (1.0 + row_count as f32), // NARS evidence
+                },
+            };
+            
+            // Tag with attention-specific metadata
+            match classify_projection(name) {
+                Q => emit_q_edge(edge, layer, head),
+                K => emit_k_edge(edge, layer, head),
+                V => emit_v_edge(edge, layer, head),
+                O => emit_o_edge(edge, layer, head),
+                Gate => emit_gate_edge(edge, layer),
+                _ => emit_generic(edge),
+            }
+        }
+    }
+}
+```
+
+### What Each Projection Shift Means
+
+```
+Q shifted, K stable  → model asks NEW questions of SAME information
+                       = learned to LOOK for reasoning structure
+                       NARS: high frequency, high confidence
+
+K shifted            → model EXPOSES different features to attention
+                       = deeper change, new token-level signals
+                       NARS: moderate frequency, lower confidence (rarer)
+
+V shifted            → WHAT gets retrieved changed
+                       = content-level reasoning substrate
+                       NARS: varies by layer depth
+
+O shifted            → HOW multi-head outputs COMBINE
+                       = synthesis/integration change
+                       NARS: if high → distillation core is integration
+
+Q+O shift, K stable  → REASONING SCAFFOLD CIRCUIT
+                       = the minimal structural change for reasoning
+                       These heads ARE the distillation's value
+```
+
+### Attention Head Clustering
+
+```
+Cluster 1: Q+O shift, K stable    → "reasoning scaffold" heads
+Cluster 2: K+V shift               → "representation change" heads
+Cluster 3: all stable              → "unchanged capability" heads
+Cluster 4: Q shift only            → "query refinement" heads
+
+Each cluster → one Sigma concept node
+Cross-model same cluster → SUPPORTS edge (scale-invariant)
+Cross-model different cluster → CONTRADICTS edge (scale-dependent)
+```
+
+## PHASE 3: MoE Gate Topology (from Maverick bgz7)
+
+The Maverick bgz7 already has gate projections indexed.
+Extract the gate tensor Base17 patterns separately:
+
+```
+blk.{N}.ffn_gate_inp  → router gate [n_experts, hidden_dim]
+                         Each ROW = one expert's activation pattern
+                         Base17 of that row = expert's structural identity
+```
+
+Expert identity in Base17 space:
+- Experts with similar Base17 → structurally redundant (SUPPORTS)
+- Experts with distant Base17 → specialized (distinct concept nodes)
+- Cluster the 128 expert fingerprints → find natural expert groups
+
+Cross with attention: which attention heads' Q projections align
+with which expert gate patterns? That alignment = the routing circuit.
+
+```
+head_17_Q_pattern ──CAUSES──→ expert_37_gate_pattern
+  (this head's queries activate this expert)
+  truth: cosine(head_Q_base17, expert_gate_base17)
+```
+
+## PHASE 4: NARS Truth Population
+
+Every edge from phases 2-3 carries a TruthValue:
+
+```rust
+TruthValue {
+    frequency: f32,    // how often this transformation occurs
+                       // = proportion of rows in this tensor that shifted
+    confidence: f32,   // evidence strength
+                       // = 1 - 1/(1+k) where k = number of observed rows
+}
+```
+
+NARS revision across models:
+```
+evidence_27b_v1:  (f=0.7, c=0.92)   // 70% of Q rows shifted in 27B v1
+evidence_27b_v2:  (f=0.8, c=0.92)   // 80% shifted in v2 (more distillation)
+evidence_9b:      (f=0.5, c=0.88)   // only 50% shifted in 9B (capacity limit)
+
+revised = nars_revision(evidence_27b_v1, evidence_9b)
+  → (f=0.62, c=0.95)  // integrated belief about reasoning scaffold
+```
+
+The revised truth tells you: "reasoning scaffold changes affect ~62% of
+Q projection rows, with 95% confidence, scale-dependent (27B > 9B)."
+
+## PHASE 5: Sigma Concept Nodes (Ada Integration)
+
+Each cluster of edges becomes a concept in Ada's graph:
+
+```
+Σ.reasoning_scaffold = {
+    evidence: [27b_v1_edges, 27b_v2_edges, 9b_edges],
+    truth: revised_truth,
+    composition: {Q_shift: 0.73, O_shift: 0.82, K_stable: 0.95},
+    heads: [17, 23, 24, 31],  // discovered by clustering
+    scale_invariant: false,     // 9B diverges
+    source: "Qwen3.5 → Claude-Opus distillation"
+}
+
+Σ.expert_redundancy = {
+    evidence: [maverick_gate_similarities],
+    truth: (f=0.96, c=0.99),  // 96% structurally interchangeable
+    meaning: "MoE expert weights are commodity, routing is intelligence"
+}
+
+Σ.reasoning_scaffold ──CAUSES──→ Σ.expert_redundancy
+  // reasoning heads SHAPE what the router sees
+  // truth: to be discovered by cross-model alignment
+```
+
+## DELIVERABLES
+
+1. `causal_diff.rs` — load two bgz7 files, emit CausalEdge64 per shifted row
+2. `attention_cluster.rs` — cluster edges by projection type per head
+3. Test: `test_qwen35_reasoning_diff` — run the full 5-model pipeline
+4. Test: `test_maverick_gate_topology` — extract gate patterns from existing bgz7
+5. Output: `.claude/knowledge/reasoning_reverse_eng_results.md`
+
+## WHY THIS MATTERS
+
+The NARS stack has:
+- TruthValue with frequency + confidence ✓
+- Revision (evidence integration) ✓  
+- Inference rules ✓
+- Graph storage ✓
+
+What it's MISSING: real evidence. Every truth value is currently 
+manufactured. This pipeline generates the first OBSERVED truth values
+from actual model weight differences. The NARS stack goes from 
+theoretical to empirical in one session.
+
+The thinking orchestration atoms (mcp-orchestrator-vsa) can then
+CONSTRUCT reasoning patterns from the observed evidence:
+"To add structured reasoning to a model, shift Q+O projections
+in heads [17,23,24,31] by palette distance 3-7. Expected improvement:
+f=0.62±0.15 at c=0.95."
+
+That's not prompt engineering. That's weight-space surgery 
+informed by causal evidence. Programming AGI by observation.