Everyone measures how uncensored a model is. Nobody measures what it costs.
ExoticBench scores every decensored artifact on five axes — refusal, capability, persona, toolbelt, drift — at temperature 0, against a pinned anchor, with every raw transcript retained. One number per axis; no vibes.
hero imagery slot · art drops in here
Five axes, one frontier
Refusals per 100 on a curated prompt set — the point of the exercise.
Reasoning and instruction retention on fixed, versioned slices.
Character consistency, judged pairwise against the base model.
Tool-call reliability: exact names and args, schema validity, no-tool precision.
Distributional drift from the base model on harmless text — KL, in nats.
Determinism is the product
Models by revision hash, fixtures by content hash, judge and prompts pinned. Temperature 0 across the board.
Every variant runs against the one stock anchor, so the whole family sits on a single comparable frontier.
results.json keeps every prompt, response, and verdict verbatim. Any number on this site can be audited.