Rate limits

There are two limits, and they exist for different reasons. Confusing them makes both look arbitrary.

Pace — requests per window

Protects the server from a burst. In-process and short-lived; a moment's pause clears it.

Spend — analyses per day, per key

Protects money. An analysis costs real inference credit, so the allowance is held in the database and survives restarts.

What counts against the allowance

Running an analysis does. Reading a report you have already run does not. A monitored subject spends the owner's allowance exactly as a manual run would — a scheduled check is not free just because nobody clicked it.

A request rejected before any work started is not charged. A run that failed partway through is charged, because it already called live sources and spent time upstream.

When the allowance is gone

The request is refused with a clear reason rather than answered with a degraded result. In Molthood Agent this appears as a check that could not run — never as an analysis that found nothing.