FORTY TRIALS, FIVE SETS
Each trial is a specification the model’s answer has to satisfy exactly — a number of bullets, a hard word cap, valid JSON, a banned letter, one correct extraction from messy data. You never write the answer. You write the instruction that makes the model produce it, and a parser decides whether it did.
Set 1 is free and needs no card. The other four sets are part of Harfi Plus.
SHAPE
FreeThe machine will answer your topic happily and ignore your shape entirely. These eight trials are about making the shape stick — exact counts, hard caps, and nothing either side of the answer.
SCHEMA
PlusThe step from “looks right” to “parses”. Every rule here runs a real parser, so output a downstream system would reject fails here too — a stray code fence is a failure, not a formatting quibble.
FORBIDDEN
PlusThe first set where the machine fails for a reason you can feel. Told not to say “gravity” while explaining a black hole, it reaches for the word anyway, because the word is what the topic is made of.
TRANSFORM
PlusThe most useful set here, and the only skill that doesn’t depreciate. Each trial hands you real data, and exactly one output is correct — so the question stops being “did it sound right” and becomes “did it process my data correctly”.
CONTROL
PlusEvery set before this asks whether you can make it comply once. This one asks whether you can make it comply every time — on a model you never tested against, in fifteen words, at a temperature you don’t control.
START AT THE BEGINNING
Trial 1.1 asks for three bullet points, under fifty words, with nothing before or after. It sounds trivial and it is not — which is the whole idea.