How the score works

A — activity (45%). Seven-day median concurrent players, logarithmic within the category. A peak can be manufactured for an hour; a median cannot. The module is measured from inside and could be forged, so it never stands alone: a Fixed server's module median can confirm or lower this figure, but never raise it past 1.25× the sighted count — the number the server shows an outside client. A server that lies to outside clients too is a matter for the drift log and the random hand audits, not the formula.

S — stability (25%). Uptime reduced by how much the population jitters.

R — reputation (20%). A Bayesian average of verified-player reviews, drawn only from the sentiment axes (rates, cash shop, staff, community). Stability is left out on purpose — it is already measured in S, and a fact should not be re-voted. An average review counts the same as no review: scores are shrunk toward a neutral 0.70 with a weight of 5 reviews, weighted by recency (a 180-day half-life) and how long the character has been verified. Reputation can move a score within a bounded band of ±0.15 — at most 3 of 100 points — so it can never overturn a real measured-population gap. Playtime is a gate to review, never a weight; hidden reviews are removed only for rules violations, logged, and counted in the open.

G — growth (10%). Today's median against the immutable day-seven record on the horizon, both sides computed by the same mediation rule as A. Holding or growing the launch crowd earns full marks; decline maps linearly to the retained fraction. Before a record is due — no launch declared, a launch younger than seven days, a wipe opening a new one, or a launch we could never have watched — G carries a neutral 1.00. Once a record is due, an unmeasured side scores 0.50, like a 50% decline: dodging the day-seven measurement never beats being measured for any server that held at least half its launch crowd, and it is never branded as proven emptiness either. This holds for listings we cannot measure at all — declaring a launch date is a commitment to be measured. Ninety days after a launch its record retires and G returns to neutral: retention is a launch question, not ancient history.

C — confidence. Set by the position class below. An unmeasured server can never outrank a measured one on equal figures.

Every listed server is probed twelve times an hour, around the clock, from monitoring nodes in separate networks. A single missed probe is noise; a pattern of them is an outage and is scored as one.

A protocol read is the number the server itself hands to every connecting client. It can be padded: fake clients parked in town inflate it, and modified server files can report an invented figure outright. A bare protocol read cannot tell a bot from a person. What it does catch is the arithmetic: a population that never dips at 5 a.m., a count that stays flat while the login queue is empty, a claim ten times the measured line. Those patterns are exactly what the drift detector flags.

The in-server module closes most of the remaining gap: it reports the count from inside the same process that runs the game, so the outside read and the inside report have to agree. When they stop agreeing, that disagreement is published, not smoothed over. What no toplist can measure is fun. We publish the numbers and the evidence; the conclusion stays yours.

A position in the rankings. A change to a score. The removal of a review. The removal of a drift flag.

Paid placement exists only as banners and featured cards rendered outside the ranking list, always visually distinct and always labelled. There is no code path from a payment to a score.