Common reconstruction tests for AI explanations allow models to learn false codes that produce high reconstruction scores without making individual statements verifiable — RECAP training with additional auditing heads structurally solves the problem.
