P273 The Explanation-Behaviour Gap

Pattern 273

The Explanation-Behaviour Gap — Pawar, Ramanayake, Kemal, Kandath, O’Neill, Bourgeon, Chatbri (BNY). arXiv:2609.05385. Catalog was 222 at ship.

Live: #pattern-273

Evaluates whether LLM explanations that look good actually track downstream behaviour. Reported Spearman correlations in the 0.349–0.580 range — positive but far from identity. Wellbeing stake: fluent rationales can mask unstable action policies.