OPEN METHOD · VERSION 1.0
Show your working.
A derived house index beside the original platform score. Every input is visible; every platform keeps its own list.
The formula.
H = (v × R + m × C) ÷ (v + m)
- R: the verified platform score, normalized to 10.
- v: the rating count on that exact snapshot.
- m: 5,000, the fixed prior weight chosen for version 1.0.
- C: the equal-series mean of the selected platform’s eligible reference cohort.
A hypothetical example.
With a 9.0 reference mean, a 10.0 score from 1,000 ratings becomes 9.167. A 9.5 score from 50,000 ratings becomes 9.455. The second title ranks higher under this formula. These are invented numbers to demonstrate arithmetic, not scores assigned to actual series.
What stays stable.
The baseline includes one latest verified snapshot per series on the chosen platform, with at least 1,000 ratings. Viki and iQIYI are never pooled. Genre, year, origin, length, search and higher count filters narrow the result without recalculating the baseline. The score, count, source link and retrieval date remain attached to every row. Snapshot refreshes can change the index.
How ties and gaps work.
Full-precision house index orders first, then normalized raw score, rating count, title and stable series slug. Ranks are positions within your selected list, not permanent awards. Missing or unverified scores remain outside the list. Unknown years and episode totals never satisfy filters that need them. Badges describe stored rating counts, release years or episode totals; rating counts are not unique viewer counts.
What it cannot measure.
The prior weight is an editorial design choice. Without individual vote distributions, this index is not a confidence interval or a probability that a show is good. It does not correct selection bias, coordinated voting, demographic differences or changes in a provider’s audience. It is not a cross-platform consensus, live popularity chart, critic review or guarantee of enjoyment.
Discussions add context, not points.
Our reading room uses selected public Reddit/forum conversations to write original spoiler-light notes. The sample is small and chosen deliberately; it cannot establish what all fans believe. We preserve different reactions and link the threads, which may contain spoilers. Post votes, comment counts, actor artwork, house opinions, personal device scores and sponsorship never alter the numerical index.
Method version 1.0 · October 13, 2026 · Jade Rabbit · 玉兔