Classification
How resource types are detected
You do not pick a category when submitting a magnet. After we read the file list, we score video, software, music, and games, then open the matching page.
Pipeline
- Fetch file manifest from magnet metadata
- Map extensions →
category_id(file-type layer) - Score verticals with weighted signals (config:
vertical_classifier.json) - Apply disambiguation rules (e.g. distro ISO → software)
- If top score < threshold or margin too small →
unknown(/u) + needs_review - Write
vertical+canonical_path; qualifying releases enter the matching release sitemap
Verticals & URL prefixes
- movie / video →
/v/{infohash}— video-heavy trees, subtitles, resolution tokens - software →
/s/{infohash}— distro ISO, deb/rpm/msi, source+README - music →
/m/{infohash}— audio-dominant, cue/log+flac - game →
/g/{infohash}— Unity/Unreal/Steam markers, ROM exts - unknown →
/u/{infohash}— ambiguous or weak signals
Key signals (summary)
- Video byte share > 50%, subtitle + video, 1080p/WEB-DL tokens → video
- ISO + distro dictionary, installable packages, source tree → software
- Audio share > 70%, cue/log + flac → music
*_Data, .pak clusters, steam_appid, game exe layout → game
Scores are heuristics from names and extensions only — not a legal or content authenticity judgment. All verticals use the same indexability gate when file lists and ruleset metadata are present.