Check your work
Read · 1 of 2
Generator and checker
Models are much better at checking an answer than at getting it right first time. So pair them: one step generates an answer, another verifies it against clear rules, and a failed check (with its reason) goes back for another try.
The best checkers are code, not opinions: does it parse, do the tests pass, are the numbers the same ones you started with? Use a model as a judge only for what code can't check.
Read this lesson as text
Generate, verify, retry. A checker makes a weak answer strong.
Generator and checker
Models are much better at checking an answer than at getting it right first time. So pair them: one step generates an answer, another verifies it against clear rules, and a failed check (with its reason) goes back for another try.
The best checkers are code, not opinions: does it parse, do the tests pass, are the numbers the same ones you started with? Use a model as a judge only for what code can't check.
Exercise: Verify, then retry
Line 1: a list of numbers the model was asked to sort. Then up to 3 lines: the model's attempts, in order.
Check each attempt: rejected: wrong numbers if it isn't the same numbers, else rejected: not sorted if it isn't ascending, else accepted. Print attempt i: … and stop at the first accepted one.
If none is accepted, finish with gave up.