The plan of prove_idea, made from the spec with no model call. Quick
programs (a brute force, the intended solution, a random generator, one
per wrong approach), then the checks our code runs on them:
stress: brute and main agree on many small random tests;
sample_check: the samples of the spec match both;
max_test: main runs a max test in the time limit;
kill_wrong: each wrong approach really fails, after the stress test,
so main can be trusted as the answer.
An ML approach gets no task: memory is not measured here.
The plan of
prove_idea, made from the spec with no model call. Quick programs (a brute force, the intended solution, a random generator, one per wrong approach), then the checks our code runs on them: