Notes from building it
Notes on inference-time compute, execution evidence, and the limits we publish alongside what ships.
What runs between your request and the code you read
Inferstep treats the change your agent proposes as one candidate, generates alternatives, executes each against your own checks, and records what each demonstrated.
Read the post
Why a pass flag is a number, not an answer
A green suite tells you nothing it checked went wrong. It does not tell you how much it checked. Graded evidence records the difference.
Read the post