affective-taxis

Models affective valence as the directional derivative of an interoceptive energy landscape (Sennesh & Ramstead 2025), providing alignment validation through structural conservation laws.

Use When

• Building alignment-aware RL agents that respect structural invariants
• Validating GF(3) conservation across reward trajectories
• Training Langevin-based policies as an alternative to PPO (which breaks conservation)
• Implementing fold-change detection reward signals in POMDP environments
• Bridging affective neuroscience models with RL training loops

Workflow

1. Define energy landscape: set attractant/repellent positions, sigmas, and amplitudes
2. Compute fold-change detection signal: r(t) = nabla_z log gamma(z; beta) . v
3. Run Langevin dynamics: dz/dt = nabla_z log gamma(z; beta) + sqrt(2) dW(t)
4. Classify trit per timestep: positive derivative → +1, orthogonal → 0, negative → -1
5. Verify conservation: sum(trits) === 0 (mod 3) across trajectory