Natural-language meaning becomes an inspectable typed creature genome; deterministic code constructs and renders the anatomy.
Jev is TypeSafe’s System One model. Instead of generating prose or images, it answers named, typed questions with probabilistic judgments. Creature Forge uses those judgments to resolve ambiguous meaning into a creature design. Ordinary code owns the literal facts, constraints, arithmetic, anatomy, provenance, geometry, and SVG rendering.
A recorded result replayed in the current Studio: Jev resolved typed decisions; the local renderer constructed the SVG. The wing inspector shows where its properties came from. No new Jev call was needed. Capture details.
Creature Forge started as a two-day experiment around a simple question: what should a model decide when meaning is uncertain, and what should remain deterministic code?
prompt → explicit literal facts → bounded Jev judgments
→ typed genome + provenance → deterministic procedural creature
For example, Jev can judge that a description implies a long neck; code maps that score into neck geometry.
Known facts stay in code. Ambiguous meaning becomes narrow decisions, and later questions can depend on earlier answers. Explicit instructions take precedence over inference. The result is a stored design you can inspect, edit, and replay: importing, replaying, or rendering an existing genome requires no Jev generation call.
Jev is used for judgment, not as an agent or code generator. Reusable anatomical parts and rules construct the creatures.
These are exact prompts from recorded runs. The excerpts separate user instructions, Jev judgments, and deterministic defaults; they are examples, not accuracy scores. Scores describe authored visual scales, not physical measurements.
Make a stealthy orange forest predator with black vertical markings, powerful shoulders, a long balancing tail and pale eyes.
- Explicit: orange body, black stripes, long tail.
- Jev inferred: feline anatomy, whip-shaped tail, pale-gold eyes.
- Code derived: four legs and a fur covering from the selected family’s rules.
The description suggests a tiger without naming one. Its orange and stripes came from the user; selecting feline anatomy was Jev’s judgment.
I want a predator adapted to chasing fast prey over open ground for long distances
The recorded genome includes:
anatomy_family = canine # Jev inferred
leg_length = high (3.00/4) # Jev inferred, ordered visual scale
leg_count = four # code derived
surface = fur # code derived
primary_color = gray # code-derived neutral default
Create a glass creature that lives in a swamp, with long legs, green clear wings, glowing orange eyes, and a stinger tail.
This is the creature in the hero image. Glass, long legs, green wings, and orange eyes are explicit. Jev selected an insect construction, six legs, two insect-style wings, and the available scorpion-tail shape. The interpretation stays within a finite design vocabulary; code constructs each part.
a wolf body with an owl head, feathered wings, and a long fox tail
The body is explicitly canine and the head explicitly avian; the finer owl source remains in the property trace. Feathered wings and tail length stay locked. Jev resolves remaining traits, while code assigns each anatomical region its owner. A local head instruction does not replace the torso’s construction.
The application sends relevant semantic state and authored questions through the official TypeSafe SDK:
| Primitive | Role in Creature Forge |
|---|---|
| Choice | Selects one bounded, authored alternative, such as a body family or wing shape. |
| Score | Places a semantic trait on an authored ordered scale; code maps the result into geometry. |
| Noul | Handles genuine yes/no judgments, such as whether a part is present. |
This is staged semantic compilation: previous typed answers determine which later question batches are relevant. Explicit locks, absent parts, and known family defaults can eliminate questions. Rich descriptions can unlock morphology and appearance decisions, while invented concepts use bounded creative choices.
The code pins jev-1.13.0 and typesafe-sdk==0.7.0. See TypeSafe’s structured workflow examples for the broader approach.
A selected part can show:
- Explicit: the prompt clause that supplied a property.
- Jev-inferred or creatively selected: the typed decision behind it.
- Derived: a deterministic family or anatomy rule, including inherited properties.
- Locally overridden: an edit made in Sculpt, the Studio’s local geometry and appearance editor.
Decisions and Raw calls expose distributions, question inputs, dependencies, model IDs, and receipts. Local edits preserve the original trace.
Python 3.10+ and a current Chromium-based browser are required. From the repository root:
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python -m backend.appOpen http://127.0.0.1:8050. Paste a TypeSafe key into the gear-shaped Settings dialog, or set TYPESAFE_API_KEY in your terminal before starting. The dialog keeps the key in server memory for that session. Stop with Ctrl+C.
See setup instructions for Windows and terminal-key commands. Keep the server on loopback; prompts are sent to TypeSafe, and local receipts can contain prompt text. See SECURITY.md.
The recorded 20 September 2026 sample contains 12 prompts and 46 Jev requests: mean 3.83, maximum 5 per creature. Its stored checks reported no explicit-constraint, dependency, inactive-part, or anatomy-slot violations. That does not establish semantic accuracy or visual recognition. The later unnamed-animal and glass-creature examples above are separate recorded runs, outside that sample.
Offline tests cover constraints, property scope, ownership, migrations, geometry and replay; browser checks exercise the Studio. See the testing guide and README evidence.
Typed output prevents out-of-schema answers; it does not guarantee a correct in-schema judgment. Jev can choose the wrong valid family or trait. Stored probabilities are model judgments, not proof of biological or visual correctness.
The renderer is a stylized 2D construction system with finite parts, materials, and colors. Some semantic details lack dedicated geometry, and faces, poses, and biological proportions remain approximate. The project has no established accuracy rate, controlled repeatability result, or end-to-end latency benchmark.
Development guardrails: Creature Forge’s current local development-pass soft cap is 3,000 attempts, with a persistent cumulative project safety ceiling of 5,000. These are local safeguards, not TypeSafe billing or account limits. Changing API keys does not reset them; the ledger is intentionally preserved.
Architecture · Setup · Testing · Security · MIT License