Sign Playground

AI Frontier AI Frontier
Home › Sign › Playground | ← Back to Sign
Safari / iPhone auto-switches to JS
JS shows browser execution output • WASM experimental now shows readable results by default. Turn on debug to also show the WASM64 memory dump • On Safari / iPhone browsers, if Memory64 is unavailable, the Playground explains why and automatically falls back to JS

Sign source

Sign v0a grammar
1

Resolved AST

JSON representation 
{}

Transpiled JavaScript

white_cats ESM target 
// the transpiled code appears here…

Run output

browser console log
Console ready. Run your code.
Sign × System One Models — v2

Jev Bridge — Define Jev's choice space with a Sign schema

Jev, the first System One Models model, is a decision model that generates no text. The official API (docs.typesafe.ai) exposes three primitives — Choice / Score / Noul — and evaluates all questions in parallel and in isolation within a single call. This demo compiles a Sign schema (#{ … } = choice space, score 1..5, noul, threshold) before running the decision, then compares the Jev lane against the LLM lane (sequential generation, equivalent to the official adapter) on the same decision. On low confidence it cascades to an LLM re-check on lane 2 → human review.

Sign schema — declaring the 3 primitives

editable

Run the decision — Jev lane vs LLM lane

in-browser simulation
— the request appears after schema validation —
STAGE 1 — SIMULATED DECISION MODEL
Compare — LLM lane (sequential generation)
— the LLM's token-by-token generation streams here —
BackendLatencyCost/callMethod
— waiting to run —
About this demo: with the real API, Jev (Early Access) returns calibrated probabilities while the Sign schema guarantees — as a contract — the legality of the choices and where each branch flows (a broken schema fails compilation, and the Jev call never happens). The LLM lane reflects the idea behind system-one-adapter (the official LLM fallback) — Jev and the LLM are swappable behind the same interface, so Sign works as an abstraction over the decision provider. Responses are pseudo-probabilities derived from keywords and do not reproduce Jev's claimed calibration. Latency and cost figures are simulated (rough real-world reference: an ≈25× gap in independent testing).