kai-tastelisted

# /kai-taste — a scored diagnosis of why an AI product feels off, and what to change ## Objective Taste treated as a control system: the thing that converts stochastic model output into reliable user outcomes at minimal correction cost. Two modes share one destination — a defensible position in the three-pillar space with evidence behind every claim. **Audit mode** produces a scored diagnosis of an existing UI, product, or generated output: a number per pillar, a present/absent call on all eight failure modes with the observation that supports it, whichever north-star metrics are computable, and a prioritized fix list. **Design mode** produces a taste contract for something being built: which zones are deterministic, which are stochastic, what must never happen, and the instrumentation that will show whether the design landed. **Iron law:** taste stays subordinate to function. The moment the system's correction vector dominates the user's intent vector, the product crosses from high-fidelity to high-friction. ## Done when Work type `audit-report` — floor **E3/C4/O1** (`harness/eco-floors.yaml`). Design-mode output ships inside the feature it informs, and that feature's own work type governs its floor. - **E3** — a named human approved the exact scorecard, and every score, failure-mode call, and metric resolves to a recorded observation: a screenshot at a stated viewport, a snapshot, a source file and line, or a session recording. - **C4** — provenance holds. `banned_w