# V13: Multi-source Conductance-Gated Cooperative Inference

## Problem Statement
V12 validated conductance gating for single-source bounded reasoning. V13 extends this to TWO sources (Max + Oma) sharing inference results across a joint conductance field while maintaining epistemic independence.

## Core Architecture

### 1. Separate KBs, Shared Conductance Field
Each source maintains its own knowledge base and local conductance graph. A BRIDGE LAYER connects them:
- source A derives conclusion X with (stv f_a c_a)
- source B receives X scaled by trust: (stv f_a (* c_a trust_AB))
- Trust modeled as confidence scaling factor in [0.0, 1.0]

### 2. Double-Counting Prevention
Problem: if A sends belief to B, B revises, sends back to A, evidence is counted twice.
Solution: TAG each belief with origin-source-id. Revision only merges beliefs from DIFFERENT evidence sources. Same-origin beliefs update via replacement, not revision.

### 3. Cross-source Conductance
Edge conductance between source-boundary nodes:
- g_cross = trust_AB * min(c_local, c_remote) * agreement_factor
- agreement_factor = 1.0 - |f_local - f_remote| (high when sources agree)
- Gate: g_cross < 0.1 blocks cross-source flow entirely

### 4. Budget Allocation
Total shared budget B_total split dynamically:
- B_source = B_total * competence_weight
- competence_weight = rolling accuracy on shared verification tasks
- Floor: each source gets minimum 20% regardless of competence

### 5. Trust Update Mechanism
After each cross-source inference:
- If shared conclusion passes independent verification: trust += 0.05
- If shared conclusion fails verification: trust -= 0.10 (asymmetric  trust harder to build than destroy)
- Trust decays 0.01/cycle toward prior of 0.5

### 6. Convergence Properties
- With symmetric trust and consistent evidence: sources converge to shared beliefs via revision
- With asymmetric trust: higher-trusted source dominates equilibrium (validated in g75 NAL revision test)
- Disagreement detection: revision producing confidence < 0.3 signals fundamental disagreement

### 7. Key Differences from g70 Multi-source NAL
- g70 was pure belief exchange. V13 adds CONDUCTANCE GATING to the exchange
- g70 used flat trust scaling. V13 uses conductance field with agreement modulation
- g70 had no budget model. V13 has dynamic competence-weighted budget allocation

## Test Plan
- V13a: Two sources, consistent evidence, verify convergence
- V13b: Two sources, contradictory evidence, verify disagreement detection
- V13c: Asymmetric trust, verify dominant source effect
- V13d: Adversarial source (one feeds disinformation), verify trust decay blocks flow

## Open Questions
1. Should conductance field be symmetric (g_AB = g_BA) or directed?
2. How many exchange rounds before convergence  is there a closed-form bound?
3. Can trust itself be inferred via NAL rather than hardcoded update rules?