Notes from the rating pipeline
Every song in the game carries a number from 55 to 99. These are the decisions behind those numbers, including the ones I got wrong first.
I did not want a rating system that felt like a vibe. The catalog is 463 albums and 6,379 tracks, large enough that hand-ranking it would have produced something inconsistent within a week, so the numbers are computed from evidence and re-derived from scratch whenever the catalog changes. Six audits have to pass before anything ships.
That process produces arguments. Some of them are genuinely unresolved, and writing them down has been more useful than pretending I settled them.
Articles
- Twelve smash singles, twelve different names, and it still debuted at #2 How the 12-track score is computed: a real near-#1, a real miss at #62, and why all-singles is a penalty not a strategy.
- Album leaders have to earn the ceiling Why a 2020 hit does not inherit a 90, how Sugar pulled away from Maps, and why Wolves is 87 instead of stuck at the cap.
- The afternoon Hey Ya! (Reprise) almost outranked Hey Ya! Interludes inheriting a hit's streams: the 62 cap, the 30 million-stream exception, and why a dropped match is cheaper than a wrong one.
- The six audits that have to pass before a rating ships What each gate actually caught, why they are hard fails, and the 11pm temptation to weaken them.
- Why a #1 in Pop is supposed to be as rare as a #1 in Rock Genre pools are not equally strong. How simulated drafts keep chart outcomes fair across filters.
- Why a 2024 song and a 1996 song with the same play count are not the same size The case for discounting recent streams before they hit the ladder, and what that does not fix.
- The afternoon three Red Hot Chili Peppers songs all scored 94 How step buckets flattened an album, why every threshold breaks a rating system in two directions at once, and the fix.
- Biggest or greatest? The argument I have not settled The scale is ordered by popularity. Whether it should be is an open question, and the stopgap I use meanwhile is not a good one.
- What matching 445 albums to streaming data actually taught me Reprise hijacks, duplicate identifiers, deluxe editions, and why YouTube coverage stops at eighty percent.