Assertions before judges

Most of what you care about needs no model to check

The instinct on hearing "evaluate an LLM" is to reach for another LLM. Resist it for as long as possible, because a deterministic assertion is free, instant, and cannot itself be wrong:

Check	How
is it valid JSON?	json.loads
does it have the required fields?	schema check
is the label one of three?	set membership
does it cite a real passage?	range check
does it avoid a forbidden phrase?	substring
is it short enough?	len

Every one runs in microseconds and has no API bill. A judge model is for the residue — tone, faithfulness, whether an answer actually addresses the question — and even then only after the cheap checks are green.

Your turn: write assertions(answer, required, forbidden, max_words) returning the list of failed check names, in this order: "missing:<field>" for each required word not present, "forbidden:<word>" for each forbidden one that is, and "too_long" when the answer exceeds max_words. An empty list means everything passed.

def assertions(answer, required, forbidden, max_words):
    """Return the names of the checks that failed, in order."""
    return []