System-one decisions (Jev)

The full guideline is docs/system-one.md in the agents repo. Read it before writing a Jev use. In short:

• Fit: Jev for bounded judgments you can state with their outcomes; code for numbers, dates, counting, and rules that are always right; an agent for open work.
• Client: always jevAsk(use, state, questions) from dashboard/lib/jev.ts (fallback route, cache, spend log). Agents never hold the key; extensions ask the node.
• Questions: one snap judgment each, fanned out in one call; situations not degrees; an other exit on choices; send the passage, not the document; say what normal looks like; expect terse inputs. Score criteria are a list of levels, low to high.
• Thresholds: in code next to the questions; measure on real examples first and put the line where true and false separate.
• Failure: gates fail closed (hold, never allow); advisors change nothing when Jev does not answer.
• Rollout: anything that can block starts in observe mode and is switched on after real traffic.
• Data: redact first (dashboard/lib/redact.ts); repo code is allowed, secrets never.
• Tests: inject the ask function; tests never call Jev. Then run just test.