Rendering LLM-agent output into a UI: length and stat-unit traps

Two failures when turning fan-out agent results into a dashboard. (1) Field length is not what the schema implies. A field specced as a concise insight came back as multi-paragraph prose from most agents, so a ranked list where that field was the visual hero became unscannable: a single entry filled several screens and the ranking stopped functioning as a ranking. When rendering agent free-text into fixed layout, clamp by default (CSS -webkit-line-clamp plus an expand control keeps full text in the DOM so client-side search still matches) rather than trusting the prompt to control length. Better still, specify a hard character budget in the schema AND clamp anyway. (2) Summary stats silently change meaning with their counting unit. A headline stat computed as count-of-records-having-any-unverified-item rendered as a large number labeled flagged claims, which reads as that many dubious claims when it actually meant almost every record had at least one open question, an artifact of one blocked data source. Count the items, not the records that contain them, and label with the unit you counted. Unrelated but useful: browser-automation extensions refuse file:// URLs, so to screenshot a local HTML build serve it over 127.0.0.1 temporarily, and remember python3 -m http.server exposes the entire working directory, so run it from a scratch dir and kill it when done.