In balanced mode, the default, the rules answer at the edge only the texts they block outright; every other text is sent to the model tier, with the questions and nothing about who sent it, because a word list cannot read rudeness or abuse written in ordinary words. On mostly clean traffic that is nearly every check.
Some products cannot allow even that: a medical forum, a children’s app under a data agreement, a customer under a contract that names every processor. For them, switch the project to Edge-only in the dashboard.
What it guarantees
- Every check on the project is answered by the rules.
engine.routeis alwaysrules. - The model client is never constructed for the project. This is not a flag checked before a call; the provider object does not exist, and a test in the source holds that an Edge-only check handed a provider throws before any text could travel.
modeis ignored:thoroughbehaves asfast.
What it costs
The Scunthorpe Test reports the measured rules-only F1 and false-positive rate for the named corpus and run. Edge-only gives up the model’s reading of running text with no word-list hit, ambiguous insults or threats, and rudeness, which only the model scores. Check the results for your language and context, then replay representative examples from your own application before choosing a policy.
Per request
mode: "fast" gives the same guarantee for one request on any project.