How AI-first moderation keeps the directory clean
Every listing, review, and ad is scanned on submission. Here's exactly how decisions get made — and when a human steps in.
Directories live or die on trust. A listing full of scams, unmarked NSFW, or redirect-only servers poisons the whole experience. So moderation isn't a bolt-on at Discapedia — it runs on every write, from day one.
The flow
When you submit a listing, edit, review, or ad, it's classified immediately against our policies: scams and phishing, malware and token grabbers, unmarked adult content, hate, impersonation, redirect-only servers, keyword stuffing, and more.
Each decision carries a confidence score, and that score decides what happens next:
- ≥ 0.90 — auto-approve or auto-reject. No waiting.
- 0.70–0.89 — a softer action: publish with a warning, force an NSFW tag, or ask for changes.
- < 0.70 — escalate to a human moderator with the full evidence package.
Some categories — child safety, credible threats — always go to a human, no matter how confident the model is.
Explainable by default
Every decision is logged with the text it saw, the policy it cited, its confidence, the recommended action, and the final action. When an owner appeals, a moderator sees all of it. Nothing is a black box, and nothing is unappealable.
Pluggable models
The classifier is provider-agnostic. It ships with a fast heuristic pass that always runs, and it can route to a local model (Ollama) or a hosted one (Claude Haiku) for deeper review. The policy and the routing stay the same; only the brain swaps out.
The result: clean listings go live in seconds, bad ones never make it, and the genuinely hard calls get the human attention they deserve.