Interpretability Arena

open the mechanism · intervene in the brain

Every recommendation exposes its observations, prior, pairwise compatibility, iteration-by-iteration strength, uncertainty, and response to intervention.

1 · Choose the remembered position

2 · Intervene on candidate labels

Uncheck a move to perform a causal intervention. The field is recomputed without it.

3 · Watch strengths settle

pᵢᵗ⁺¹ ∝ priorᵢ^α · pᵢᵗ · (0.35 + Σⱼ compatibility(i,j)pⱼᵗ)

4 · Compatibility contributions

5 · Observation provenance

← Relaxed Chess Brain · Chess Worlds