16 · Evidence, not impressions
The test infrastructure that tells you whether a model change made the system better or worse, and gives a supervisor an answer that is measured rather than asserted.
Most teams ship model changes on impressions. Someone tries a new prompt, it feels better, it goes out. That works until the system is doing something consequential, at which point the absence of a harness means nobody can say whether last week's change improved anything, and nobody can answer a regulator asking how the behaviour was validated.
A harness makes the question answerable. A held-out set drawn from real cases rather than invented ones, graders that are themselves checked against human judgement, and a regression suite that runs on every change. The output is a number that moves, so a change can be rejected on evidence instead of argued about, and a supervisor gets a record of how the system behaved on cases you did not choose.
Common questions
Why not just check the outputs by hand?
Hand-checking works up to a few dozen cases and then silently stops happening. The failure is not that manual review is inaccurate, it is that it does not survive contact with a release schedule, so the checks quietly become spot checks and then become nothing. A harness runs whether or not anyone remembers to run it.
Can a model grade another model's output reliably?
Only if the grader is itself validated. An ungraded grader is a confident number with nothing behind it, which is worse than no number because people act on it. We calibrate graders against human scoring on a sample, measure the agreement, and re-check it when either model changes. Where agreement is poor, that task does not get an automated grader.
What does a supervisor actually want to see?
That the behaviour was tested against cases the firm did not choose, that the failure modes were looked for rather than stumbled upon, and that a record exists from before the incident rather than assembled after it. The technical content matters less than the evidence trail being contemporaneous.
Next capability
Systems engineering
Bring us the hard part.
Forty-five minutes with the people who would actually run the build.