Seven guards checked that content is WELL FORMED. None checked what it DOES, so a
change to a damage modifier, a hit-point formula or the armour value on a service vest
could double a creature's lethality without touching one line of that creature - and
nothing in the build would notice, because the creature did not change.
Built as a regression test rather than the hand-declared bands the creature-forge plan
described. Bands are the wrong shape here: declaring "keepers: dangerous" across
forty-six creatures means inventing forty-six judgements, and after --spread it is
clear the interesting question is not "is this dangerous" - that has no single answer -
but "is this the same as it was". A recorded baseline answers exactly that, needs no
authoring, and cannot be argued with.
Every creature is measured solo against a FROZEN party of four, two armed postings and
two trades, at a fixed seed, so the only thing that can move a number is a change to
the rules or to the creature. Tolerances are 6 points of wipe rate and 0.35 agents,
which is outside the noise floor of 200 runs - a guard that cries wolf gets deleted.
Verified by tampering: told the baseline the nuckelavee was harmless and the guard
caught it at +16.5 points and +1.67 agents down, exit 1. Runs in half a second, so it
joins the pre-build checks rather than being something to remember to run.
Also: the README's guard block was a hardcoded list of four while the suite was seven.
It had silently stopped mentioning every guard added after it was written, including
check-creatures and check-scenarios. Now all eight, and the prose says eight.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
