- Draft — a general LLM
- A general chat model reads your plain-English request and drafts a structured order: instrument, side, size, order type, and any stop-loss distance. It only interprets and drafts — it has no ability to execute, submit, or connect to anything.
- Compute — deterministic code, in your browser
- Fixed, inspectable arithmetic against a small illustrative contract-spec table (tick value, point value, margin, and a reference price per instrument) computes exposure, margin, tick value, max loss, and what share of a fixed illustrative daily-loss limit that max loss would consume. View source to see the exact numbers used.
- Judge — Jev
- Jev never sees dollar figures — only your original text, the draft's restatement, and bucketed signals (stop-loss present or not, how severe the daily-loss impact bucket is, the drafting model's own confidence). It answers three bounded yes/no-as-probability questions: does the draft match your intent, is the request ambiguous, and should a human review it.
- Decide — deterministic code, again
- Code owns the final gate, not Jev. A draft missing instrument, side, or quantity is blocked before Jev is ever asked. A missing stop-loss or a severe daily-loss bucket always forces review, regardless of what Jev says. Beyond that, review is also triggered if Jev's intent-match score is low or its ambiguity score is high. Jev's own "needs review" score is shown but only forces the gate on its own when unusually high — in testing it leans cautious by default for almost any real order, so code treats it as one input among several rather than a single switch.
- Paper-trade draft only
- There is no broker connection anywhere in this Lab. Every outcome — ready, review required, or blocked — is a draft, full stop. Nothing here is ever submitted or executed.