What a two-week proof should actually prove

Most two-week proofs demonstrate the part everybody already believed would work. That is a demo, not a proof, and it moves no risk off the table.

Find the killing assumption

Every project has one assumption that, if false, ends it. Latency under real data volume. Extraction accuracy on the messy documents rather than the clean ones. Integration with the system nobody has documentation for.

Build against production-shaped data

Synthetic data proves synthetic results. If we cannot get real data in week one, that itself is the finding.

A proof can end an engagement

The most valuable outcome we have delivered in two weeks is a clear recommendation not to build the thing. That is worth considerably more than the fee.