promptdojo_

Golden evals earn autonomy — every macro proves itself per intent — step 7 of 7

ch 50 · support studio7/7
Checkpoint

One last thing before we move on. Same surface as a write step — but the lesson doesn't complete until this passes.

Checkpoint: the autonomy ladder for the whole macro library.

There is exactly one way up and three ways to stay put. Write autonomy(macro) returning one string, checked in this order:

  1. total == 0 → "review-required (no golden set)" — zero evidence, zero trust; the demo doesn't count.
  2. passed < total → "review-required (failed eval)" — a policy grant does NOT override a failing eval.
  3. no policy_grant → "review-required (no policy grant)" — a perfect eval does NOT override the approval policy. Autonomy needs both signatures.
  4. otherwise → "suggest-first" — the macro may be offered at the top of the review loop. (Even now, nothing sends itself.)

The report loop is provided; names are padded to 15 characters. Expected output:

wismo           suggest-first
reset-password  review-required (no policy grant)
refund-status   review-required (failed eval)
return-window   review-required (no golden set)

full-screen editor opens — close anytime to keep reading.

read, then continue.