Review Moderation & Content Quality

A hospitality marketplace lives or dies on trust. Guest reviews are the single strongest driver of bookings - and the fastest way to lose a customer's confidence. A flood of spam, off-topic promotion, competitor sniping, fake praise and genuinely unsafe content erodes that trust faster than any amount of good reviews can build it, and it arrives in every language a guest happens to write in.

Manual moderation does not scale to tens of thousands of multilingual reviews a week, and pure auto-publishing is unsafe. The job is to publish the honest signal quickly, hold the genuinely uncertain cases for a human, and reject the clearly harmful - at a cost that stays controlled and an audit trail that holds up.

Auto-publishing everything is fast but unsafe; holding everything is safe but slow and expensive. This system balances the two with a confidence-aware four-band router sitting on top of a cheap deterministic spine, so the great majority of clean reviews publish instantly while only the edge cases reach a moderator.

It is built to be honest about what it can and can't decide. A low-confidence automated decision is downgraded to a human review rather than trusted blindly, and a reviewer-facing reason is attached to anything held or rejected so the moderator starts with context, not a blank verdict.

This is a faithful, downscaled reimplementation on the shared synthetic catalog - never the production system. It proves the full moderation pipeline end to end: deterministic policy and topic stages, a classifier with confidence, a selective reason for held and rejected items, and a four-band router, all surfaced honestly in an inspector and operated through a role-gated console. It deliberately simplifies the breadth of policy, languages and scale, and represents retraining with an owner overlay rather than live learning - all noted in Architecture → Out of scope. The client is not named.