Rethinking — resolve unknowns, then double-check

Two matbot patterns keep answers honest: rumsfeld resolves things the model doesn't know before it answers, and inner voice critiques an answer after it's drafted.

Rumsfeld — resolve unknowns just-in-time

When a turn references a term, entity, system, or concept the agent doesn't know, it loads context for that unknown instead of guessing.

Inner voice — a second opinion before committing

An opt-in bicameral critique: draft an answer, then have a second pass challenge and sharpen it. It's a cognition skill, off by default, invoked when you want the extra rigor.

Together: don't answer around a gap (resolve it first), and don't ship the first draft blind when it matters (critique it).