No language model decides anything that affects a child's lungs.

The school's circular, written in Cedar — AWS's open-source policy language, not a prompt. A period stays outdoors only if Cedar returns allow for that class, that period and that hour's evidence. A principal can argue with a number here and change it without touching the application.

Saans runs on a global model's forecast, so the fair question is whether that forecast is right. There is no keyless ground sensor feed for Delhi to check against, and quoting an accuracy figure without one would be dishonest. So it answers the question that matters for a safety tool instead: does the decision survive the forecast being wrong? Every call is re-decided across ±40% forecast error, wider than anyone claims for this model.

Replayed over real days that happened, with measured air. The point is not that the protocol fires. It is that on a clean day it does nothing at all — a protocol that fires on a good day gets switched off within a fortnight, and then it is not there on the day that matters.

Decide, draft the notice, write the ledger, store the notice, send it. The agent drafts the opening line a parent reads first, and a guard checks that line against the record — a small model writing a fluent sentence with a wrong number in it is the likeliest way this hurts somebody. Rejected drafts fall back to the deterministic summary, and the page says so.

A parent asking why their daughter was kept inside on the sixth should get the same answer the principal got that morning. The agent answers from the ledger and the protocol only, and cites the clause.