Verify a draft answer by having an independent subagent re-solve the same problem from scratch and a second subagent probe via a different method, then mechanically compare. The subagents never see the draft. Use before delivering any answer where being wrong would be costly AND the claim cannot be fully tested empirically — algorithm correctness, mechanism design, non-measurable numeric estimates, safety/optimality assertions, reasoning on non-obvious invariants. Triage gate routes empirically testable claims to tests (and partially testable claims to test + audit on the uncovered surface). Hard cap 2 spawns. Reports honest disagreement; never silently picks.
