guardrail-designerlisted

# Guardrail Designer A guardrail is a named failure plus the condition that fires the defense. "Add validation" is a wish; "block any balance in the draft that isn't in this ticket's record" is a guardrail. ## Verification gates (defined first; output is blocked until all pass) - **G1 — Named failure + trigger, every guardrail:** each guardrail states the specific failure it prevents (an incident or a labeled class risk) and its trigger condition — the observable check that fires it. Failure-less ("add validation") or trigger-less ("prevent hallucinations") guardrails fail. - **G2 — Placed and costed:** every guardrail sits at a named point in the flow (pre-draft / post-draft / pre-send / post-send) with its action on trigger (block / flag / route / degrade) and its cost (latency, false-positive friction, build effort) stated. - **G3 — Layered honestly:** mechanical checks run before human review; the human layer is treated as a failure surface with its own guardrail (fatigue, volume), never as the universal catch-all. Known incidents drive the top of the list; invented incidents and imported compliance requirements fail. ## Steps 1. **Walk the flow as an attacker of its own outputs.** Per step: what enters, what the model does with it, what leaves, and how each can be wrong (wrong data in, wrong synthesis, wrong claim out, wrong human action). 2. **Rank the failure list:** actual incidents first (they're proven reachable), then class risks the flow's shape makes likely