39,000 Torrents: The Bug My Green Benchmark Never Caught

Short summary

A detailed account of building a torrent filename parser for a self-hosted PHP media server, using a 39,000-file corpus as a benchmark. The author iterated from 10.84% to 99.93% cleanliness by truncating at the first technical marker rather than enumerating all known tags. Despite a green benchmark, a silent bug turned every TV show into a film match, illustrating the gap between what metrics measure and what actually matters.

• •Truncating at the first technical marker (V2) delivered 99% of the total improvement from 10.84% to 99.83% cleanliness
• •A green benchmark masked a silent bug that matched every TV show as a film
• •AI assistant accelerated iteration but the core insight — cut at first tag — was human-derived

Generated with AI, which can make mistakes.