Extrapolation Risk: Few Points to a Fleet

Moving from a handful of test results to a fleet design carries extrapolation risk — named openly and managed by staged builds.

The scaling problem

A fusion program cannot build hundreds of prototypes. It measures a few machines and extrapolates to a fleet. That extrapolation is a real risk: behavior that holds at test scale may shift at fleet scale or across many units, and a design tuned to a couple of data points can be over-fit.

MetroVolt inherits this risk twice — once from the breeder’s own few-unit ladder, and once from the burner’s. Program documentation flags a specific two-point-to-fleet extrapolation concern, which is why the FOAK–NOAK–BOAK ladder exists: to add data points before committing to a fleet design.

How the risk is managed

The mitigation is staged builds with instrumentation, conservative margins where data is thin, and a willingness to revise the design between stages. Each build is treated as a measurement, not just a delivery, so the extrapolation shortens with every unit.

• Risk: over-fitting a fleet design to few test points.
• Mitigation: staged FOAK→NOAK→BOAK with measurement at each step.
• Discipline: margins where data is thin; revise between stages.