The loop and four ways to look again

The loop

1. Draft: the model writes a first version.
2. Look again: something finds problems in it.
3. Revise: the model edits the draft using what it was told.
4. Stop: nothing is wrong, or a round limit is hit.

The pattern is easy to write and easy to get wrong. The questions that decide whether it helps are: who looks (step 2), what they say (a name, a vague feeling, an exact message), and when to stop (step 4).

The running example

A customer notice about a delayed order, written by a model, must pass six checks that run on the text:

The toy writer slips on each check at a rate of 8% to 15% (writing the date as 21/3/2026, quoting the whole total as the refund, adding a long closing paragraph). About 47% of first drafts pass all six. Here is one job revised using the checks' messages:

draft 1: Your order ORD-9916 is delayed and we are sorry. It will now arrive on
2026-03-21. We have refunded $115 for the wait. Our team has been working very
hard on this and we want you to know that your patience means a great deal to
everybody here.
   check failed: The notice is longer than 40 words; cut it down.
   check failed: The refund is $23 (a fifth of the total), not another amount.
draft 2: Your order ORD-9916 is delayed and we are sorry. It will now arrive on
2026-03-21. We have refunded $23 for the wait.
   all checks pass

Note that the second message contains the answer ("a fifth of the total"). A bare "refund wrong" would leave the writer guessing; compare module 2's error messages.

Four ways to look again

Same 1000 jobs, at most three rounds:

(Self-critique assumes a critic that finds 60% of real problems and invents a problem in 10% of the parts that were fine. A "spoiled" draft was valid at first and invalid at the end.)

Three things stand out:

1. The outside check wins on every column: best quality, fewest calls (checking is code, not a model call), and it never touches a good draft.
2. Even a bare pass/fail from a program is powerful. Retrying on "not valid" reaches 92% because a check selects: it keeps trying until a draft passes. This is the same idea as sampling several answers and picking the one a verifier accepts (module 5 returns to it).
3. Self-critique costs the most and gains the least, and for 15% of drafts that were already right, it makes them wrong.

Table of the same curves:

Self-critique flattens out around 70% because the problems the critic does not notice are never fixed, and the false alarms keep damaging things. More rounds do not rescue a critic that cannot see.