Review Moderation & Content Quality

A guest submits a review in their own language. Language detection, the deterministic policy engine and the topic lexicon run first, all at $0, establishing a cheap and explainable backbone before any model is called. A classifier then emits a verdict and a confidence, and for anything destined to be held or rejected a reviewer-bound reason is generated. The four-band router maps the result to an outcome, downgrading low-confidence auto-decisions to a human review.

A clean, confident review is published instantly. An uncertain one lands in the moderator queue with the AI's reason attached, and a clearly harmful one is rejected. The moderator resolves the held items, and the audit trail records who did what - closing the loop between the automated decision and the human one.