Every Forge Sprint ships with an eval harness and reliability gates as a named deliverable. Not an option, not an add-on. This article explains what those two things are and why they are in the contract.
The eval harness
An eval harness checks the system's outputs against a test set. Before any change reaches production, the harness runs and reports what improved and what regressed. It turns 'the model seems better' into a number you can read.
Reliability gates
Reliability gates are the pass conditions the system must clear before release. If a gate fails, the release stops. The gates make reliability a property of the process rather than a hope about the model.
Reliability is a process property, not a model property.
Already built something?
For AI systems your company already built, the Eval and Reliability Audit applies the same machinery retroactively: a fixed fee, 2 weeks. You get an eval harness, reliability gates, and a fix plan for the system you have.