Right now, if your agent quietly gets worse, a customer is the one who tells you. This is the harness that tells you first.

I pull real tasks out of your own transcripts, turn them into a graded set of about fifty, and write the scoring so a run gives you a number that actually moves. It goes in CI, so a change that breaks behaviour fails the build instead of the customer.

You get the task set, the runner, the report, and one rule I work by: whatever writes the answer never grades it.

One week to build. After that it runs on every deploy, and you own it.