Training models to report themselves faithfully

Interpretability from the inside out: rather than reading weights, train the system to describe its own state accurately and reward it for doing so. The open question is whether faithful self-report is a capability or a performance.