v12 Design Document: Conductance-Gated Abductive Inference
Hypothesis
Resource-bounded abductive inference improves in both efficiency and epistemic quality when hypothesis selection is gated by a learned 4-layer attention landscape (adaptive SPH kernel, conductance permeability, contradiction veto, utility drift) rather than by geometric proximity or uniform budget allocation alone.
Architecture
Transport Equation
delta_STI(i->j) = W(dist, h_i) * g(e_ij) * gate(e_ij) * (STI_j - STI_i + alpha*(U_j - U_i))
Layer 1 — SPH Adaptive Kernel W(dist, h_i)
- h_i = h_base + beta*(1 - conf_i), capped at h_max
- Uncertain nodes widen search radius; confident nodes stay tight
- Validated: v11 (A=1.2, B=2.2, C=1.6, D=2.6)
Layer 2 — Conductance g(e_ij)
- g = w_rreward + w_hhebb + w_ig*info_gain (additive permeability)
- Validated: v9c (Robin-Bird=0.3097, Robin-Dead=0.126)
Layer 3 — Contradiction Gate gate(e_ij)
- Multiplicative veto: 0=blocked, 1=full flow
- Validated: v9c (compatible=1.0, contradictory=0.25)
Layer 4 — Utility Drift U_j - U_i
- Local potential differences modulated by global goal-relevance
- Validated: v11 (A=0.6, D=0.1)
Abduction Pipeline (from v20_rev3)
- SPH convergence on cyclic KB (v16: 2 iterations)
- Budget-gated flat abduction: bc-step reads conductance as budget
- Recursive backward chaining with TV composition
- Auto-revision merging multi-path evidence
Validated Components
| Component |
File |
Key Result |
| v9c conductance+gate |
ecan_v9c_assert.metta |
47x selectivity |
| v10 SPH transport |
ecan_v10_sph.metta |
A->B=-0.155 |
| v11 adaptive kernel |
ecan_v11_adaptive.metta |
h=1.2 to 2.6 |
| v11b full 4-layer |
ecan_v11b_transport.metta |
47x same-distance diff |
| v16 SPH+abduction |
abd_tv11_v16.metta |
Cyclic convergence 2 iters |
| v20_rev3 full pipeline |
abd_tv11_v20_rev3.metta |
Recursive BC+revision |
Open Questions
- Uncertainty source for adaptive radius: NAL confidence or variance of incoming signals?
- How does adaptive radius affect abduction breadth vs depth tradeoff?
- Utility drift interaction with contradiction gate under cyclic evidence
- Convergence guarantees when conductance field changes during inference
- Scaling: does 4-layer gating maintain selectivity on 100+ node KBs?
Experimental Plan
- v12a: Wire v11 conductance INTO v20 bc-step budget gate (replace raw SPHCost)
- v12b: Compare abduction quality: uniform budget vs conductance-gated on same cyclic KB
- v12c: Measure epistemic gain (revision confidence) under budget constraint
- v12d: Stress test on larger KB with distractor nodes
Core Principle (Kevin Machiels, 2026-04-23)
Attention flows when target is nearby, permeable, compatible, and locally more valuable — not just geometrically close.