vdd

Apply a verification loop that requires the agent to expose hidden assumptions, attach evidence to every claim, and pass four gates covering time, scope, mechanism, and counter-evidence before stating completion. The agent must also switch to a fresh reviewer identity rather than re-walking its original reasoning.

Write a task contract before starting that defines core goals, design decisions, and explicit no-change boundaries. During review, allow only three verdicts: matches the contract, deviates with evidence, or the contract lacks coverage, which blocks open-ended fault-finding.

No, this method runs entirely through the agent's own reasoning discipline with no external tools, scripts, or platform integration. Its effectiveness comes from the structured loop itself, making it usable in any conversation context.

Match verification depth to the cost of being wrong. Trivial changes like copy edits need only a quick glance, while core logic, authentication, payments, or data writes justify the full multi-step verification with a switched reviewer identity.

Agents are trained to be helpful, so a prompt framed as find issues pressures them to manufacture findings even when the design is sound. Anchoring every reported problem to a written standard prevents fabricated issues and stabilizes review scope.