The afternoon three Red Hot Chili Peppers songs all scored 94
Rating systems fail in a specific and recognisable way. Ours failed by producing an answer that was defensible for every song individually and obviously wrong the moment you saw them together.
The symptom was easy to spot. Californication, Scar Tissue and Otherside all came out at exactly 94. Three of the biggest songs from one album, sitting on the same number, with nothing to separate them.
Nobody would rank those three as a dead heat. More to the point, the game itself becomes worse when that happens: if you spin that album and every worthwhile track shows the same value, there is no decision left to make. The pick stops being interesting.
Why the bug was there
The ratings map streaming counts onto a scale through a ladder of calibrated anchors. Early on, that ladder was implemented as buckets. Songs between roughly 350 and 400 million plays land in one band, 500 to 600 million in the next, and so on, each band carrying a floor and a ceiling.
Buckets are an appealing way to build this. They are easy to reason about, easy to write down, and easy to check by hand. You can look at a table of bands and say yes, that is roughly what a song at that level should be worth.
What they cannot do is distinguish between two songs inside the same band. Those three tracks have meaningfully different play counts. All three sat inside one bucket, hit the same ceiling, and came out identical. The system had thrown away the information that separated them before it ever got to the rating.
The general problem with thresholds
The deeper issue is not that the buckets were too wide. Narrower buckets would have split those three and then produced the same collision somewhere else, because the failure is structural rather than a matter of tuning.
Any threshold in a rating system creates two artefacts, and they point in opposite directions. Inside a band, real differences vanish. At the edge of a band, trivial differences explode: two songs a rounding error apart in plays can land on either side of a boundary and finish points away from each other. The bucket that flattened the Chili Peppers was simultaneously capable of manufacturing a gap between two songs that were effectively tied.
Neither artefact is visible if you audit one song at a time. Every individual rating looked reasonable. It was only in the context of the album that the flattening became obvious, which is a good argument for auditing rating systems in groups rather than one row at a time.
The fix
The bands were replaced with continuous interpolation. The calibrated anchors stayed exactly where they were, because they encoded real judgements about what a given level of popularity is worth. What changed is what happens between them: instead of snapping to the band's ceiling, a song's floor and ceiling are now interpolated smoothly across the logarithm of its play count.
Working in log space rather than raw plays matters. The difference between 50 and 100 million streams is a large difference in kind. The difference between 950 million and a billion is a rounding error. A linear treatment would take those equally seriously, so the curve compresses as the numbers grow, which is closer to how the gap actually feels.
After the change, those three songs separate in the order their evidence supports, and no boundary exists for a song to sit unluckily on the wrong side of.
The second half of the fix
Smooth interpolation solved the collision but exposed a related problem. Run an album's tracks through an absolute ladder independently and the resulting spread is often far too wide, because streaming counts within a single album routinely differ by an order of magnitude. Album tracks were finishing nearly thirty points apart, which reads as though half the record is unlistenable.
So album ratings are calibrated after the fact toward a spread of roughly eight to sixteen points, using a compressive square-root map that keeps mid-pack songs near the leaders when the underlying stream ratios are only moderate. Order is preserved strictly. Where the pipeline has to choose, it prefers lowering a less popular track over breaking an absolute ceiling, because the ceilings are what keep ratings comparable across the whole catalog.
Equal ratings are still permitted for genuinely close songs. What is no longer permitted is two songs sharing a number purely because a bucket ate the difference between them.
What I took from it
The audit that would have caught this earlier is not "is this rating correct" but "do these ratings, seen together, tell the truth about this album". That check now runs on every build, along with one that re-derives every stored rating from the raw signals and fails if any of them disagree. A number cannot be quietly hand-edited into place and left to rot.
It also shaped a rule I apply elsewhere. When the age adjustment for recent releases was added, it was built as a smooth decay rather than a set of era brackets, specifically so it could never reorder two songs from the same year. I made that call right after watching three songs collide, and it is written up separately.
Play #1 Debut