ANSWER

What is a dual-run?

Reviewed October 21, 2026 · Reviewed by Paulo Pierrondi, founder.

A dual-run is the parallel execution of a replacement beside the live legacy system: both receive the same inputs and produce comparable outputs, and the replacement has no production effect. The legacy system stays in command. The candidate runs in shadow while an Equivalence Contract states what must match, which differences are blockers, and which exceptions may be approved with an owner, rationale and expiry. Cutover comes later, with human approval, a tested rollback and an Evidence Pack that records diffs, owners and the way back. A dual-run is not a big-bang rewrite and not a generic accuracy percentage. It is the proof, before shutdown, that the work can leave the legacy system. Legacy Exit uses that window to decide cutover on proof, not on faith.

Step by step

• Choose a recurring process with stable rules and comparable inputs and outputs.
• Close the Equivalence Contract before measuring: fields, tolerances, blockers and owned exceptions.
• Run the replacement in parallel on the same inputs as the legacy system, with production writes off.
• Classify every variance — blocker, material, cosmetic, legacy defect or approved exception — before reading any aggregate rate.
• Approve cutover only with human sign-off, tested rollback and an Evidence Pack that records diffs, owners and the return path.

How an agent does this

In AgenticosCore, Parallel executes the dual-run. Trail stores each comparison, House Rules stop progress outside the contract, and Brake halts execution. An approved exception enters the Evidence Pack with an owner, rationale and expiry. If equivalence is not proven, the valid outcome is not to cut over: the legacy system remains in command.