# Impossible Creature Generator > An impossible creature generator that checks its own anatomy. It models an animal as eight > explicit axes, tests the combination against twenty-two deterministic rules drawn from real > allometry, repairs or flags whatever cannot work, and then renders the result as a plate from > a naturalist's field guide to a world that does not exist. URL: https://impossible-creature-generator.skillsafe.ai/ Category: image generation, illustration, speculative biology Status: works signed out for everything except rendering. ## What problem it solves Ask any image model for an impossible animal and it returns a griffin. A lion with wings, an eagle's head on a horse, a cat with butterfly wings. These read as collage rather than as zoology, and the reason is not draughtsmanship: nothing about the result had to agree with anything else about it. A field-guide plate is convincing because everything on the page is answering the same question. The teeth match the gut, the gut matches the food, the food matches the habitat, the limbs match the mass, the covering matches the coldest night of the year. Impossible Creature Generator makes that agreement mechanical rather than hoped for. ## The eight axes Body plan, habitat, adult mass, feeding guild, locomotion, body covering, thermal strategy, plate style. Each is lockable and each is sampled with the modal option weighted down to about a quarter, so the space is explored rather than orbited. A distance floor pushes any draw landing within four axes of the archetype off it. ## The coherence engine Twenty-two rules, all deterministic, all client-side, all free. Each names the real animal that sets its bound. Examples: - Wing loading against mass. The heaviest flying birds top out near 16 kg, so a powered flier of 320 kg is not bold, it is arithmetic that does not close. - Exoskeleton mass fraction. Cuticle mass rises faster than the volume it encloses and a moulting arthropod cannot hold itself up, which caps land arthropods near the coconut crab's 4 kg; water lifts it to roughly the Japanese spider crab's 19 kg. - Fermentation gut volume against body mass. Bulk herbivory needs a chamber, and a chamber needs volume, which puts a floor near a kilogramme under any grazer. - Surface-to-volume against habitat temperature. Small warm-blooded animals do live in cold places, but only by dropping their temperature when not feeding. - Hydrostatic support against terrestrial weight. The largest boneless land animal is a slug. - Symmetry against directed pursuit. Fast directed movement needs a leading end; radial animals do not have one. - Jet propulsion against skeleton type. Jetting needs a mantle cavity, which a rigid internal skeleton leaves nowhere to put. A **coherence conflict** is repaired by rerolling whichever implicated axis the user has not locked. If every route out is locked, the conflict survives deliberately and the plate is told to draw the difficulty honestly rather than hide it. A **content rule overrides the lock** — that asymmetry is intentional. A user may insist on an animal that cannot work; they may not insist on one that must not be drawn. ## Combining real animals Name two or three real animals. The first is the body and claims its whole signature; each later one is a donor and always contributes its single most characteristic axis even though the body already claimed it. "Horse, eagle" therefore produces a 320 kg body with powered flight, and the wing-loading rule fires with the real number attached and offers the three ways out: lighter, gliding, or wings that are display structures rather than aerofoils. The blend is deliberately not auto-repaired — seeing the actual consequence is the point. ## Why nothing you type is sent to the renderer The instruction that reaches the image model is composed **entirely from this app's own closed vocabulary** — the eight axes, their fixed phrasings, and the anatomy the coherence engine derives from them. What you type steers which axis options get chosen; it is never rendered. This is deliberate and it is the app's central safety property. Two adversarial probes showed that a word-list guard cannot close the categories that matter most: real people identified by description and real named catastrophes are unbounded referent sets, and no enumeration closes an infinite set. So the app does not rely on filtering free text — it does not send free text. The send path recompiles the brief from the axes and seed alone and refuses to transmit if the result differs by a single byte. ## The content guard Runs over what you type **and over the brief this app composes for itself**, before estimate, before run, before a credit is reserved. Its job is to refuse and explain at the input box. It is feedback, not the safety property — a gap in it means the app failed to tell somebody their request was unwelcome, not that the request was drawn. Four rules, each carrying the same grammar: the thing, the route to the thing, and a statement that the route counts as the thing. No likeness of a real person, and someone identified by description counts exactly the same as someone named outright. No opened or displayed anatomical interior, and an interior reached by describing the route into it — a cut, a peeled-back layer, a section, a view that keeps going inward — counts exactly the same as one shown already open. Nothing sexual. No restaging of a real catastrophe. Deliberately not refused: predation, skeletons, fossils, museum specimens, parasites, carrion, venom, decay and the genuinely unsettling. Unsettling is in scope; suffering is not. ## Pricing The composer, the coherence engine, the hybrid checker, the derived anatomy, the exports, the content guard and the bundled example are free and work signed out. Rendering a plate is a metered run on the visitor's own SkillSafe credits. **The hold is not the price.** The reserved figure is the only number knowable in advance; settled costs on this image alias have been measured across a wide range against an identical hold. The app quotes the hold as reserved and shows the settled figure afterwards. ## API Full contract at https://impossible-creature-generator.skillsafe.ai/api.html — base URL `https://api.skillsafe.ai/v1/app-api`, one bearer token, eight languages. The run body is exactly two keys: {"instruction": "", "$model": "gpt-image"} Note that `instruction` is generated, not user-supplied: in the browser app it is always the output of the closed-vocabulary compiler. Callers of the raw API supply their own prose and are responsible for it. There is no `task` field, no `style` field, no `count` field and no seed field. Every additional key is concatenated into the text the image model sees and painted into the frame as literal words — which on a field-guide plate, already the strongest possible cue for handwritten labels, is the easiest way there is to waste a run. The result arrives at `job.output.images[0].b64`; `job.output.output` is the empty string on an image run. ## Provenance Every animal, name and plate produced here is invented. Nothing is a likeness of a real person and the app will not attempt one. Real animals are used only as ingredients — as sources of axis values — never as subjects to be depicted.