Preprint · August 2026

Causal Observability for Active Reinforcement Learning

When should a reinforcement-learning system acquire additional information before acting?

Contribution

Causal Observability Optimization treats coarse observation, boundary observation, probing, transfer, continuation, and abstention as cost-sensitive meta-actions. It asks which observation regime preserves the intervention-relevant distinctions needed for a declared decision.

Evidence

  • A Lean-checked finite result on the sequential control value of observation refinement.
  • Typed replay infrastructure for connecting decisions to inspectable experimental traces.
  • Controlled partially observable structural causal model validation, with MiniGrid as an external protocol check.

Research infrastructure

CogTrace is the private research infrastructure behind this work: typed event streams, replayable observation views, and claim-evidence manifests for auditable agent experiments.