Let the scripts judge and the agents fix

The same shape shows up in different tools I’ve built. A deterministic layer does the judging: it’s cheap, it’s reproducible, and it’s safe to run in CI because it changes nothing. The agent is kept to the one thing it’s good at: proposing a fix a human can actually read.

In Tightrope the judge is a set of contract checks and the agent drafts the remediation. The same split shows up in destruction bundles (read-only evidence in, signed plan out, human delete) and in edit vs merge authority. Repeating-unit platforms need the same certification layer so the hundredth unit can’t invent a new return shape.

The reason is trust. A model’s output varies; a script’s does not. If the gate that certifies a change as safe to merge is itself nondeterministic, it isn’t really a gate. So the deterministic part holds the line, and the agent does the creative, reviewable work behind it.