The Verification Stack

Your eval dashboard is green. The output shipped. A week later a customer finds the error the scores never caught, and you realize the dashboard was measuring, not deciding.

Evals measure. They do not decide. Between a score and a shipped artifact sits a missing organ: a system that turns measurements into verdicts, with evidence, at a cost you can defend. Almost nobody builds it, which is why so few agent systems reach production and the rest stall in the gap between "looks right" and "is right."

The Verification Stack architects that missing organ. It builds verification as a subsystem you design once and operate forever: machine-checkable specs that compile into gates, five layers ordered so cheap checks shield expensive ones, LLM judges deployed with known error rates, and human escalation as a designed interface rather than a fallback. It is the one moat a model release cannot delete.

Inside: the five-layer stack and the Verification Grid that maps every gate your agents need; how to verify code, content, decisions, and irreversible actions; calibrating a judge against its own error matrix; anti-gaming gates that stay meaningful after agents learn to beat them; verification budgets you can defend to a CFO; and verification-driven development, the successor to test-driven development. Every claim carries a receipt. Every chapter ends with a gate you can run.

Volume 4 of The AI-Native Builder Canon.