How scoring works
Every public node is scored on its own record, mostly over the last 30 days. Its place on the leaderboard comes from that score alone, so running more nodes never lifts a place.
The ten parts
Uptime
The share of hours in the last 30 days that we reached the node. 99.9% or more scores 100, and each missing nine costs 25: 99% scores 75, 90% scores 50.
Stability
Unplanned restarts in the last 30 days. Two are free and each one after that costs 10. A restart that came with a new version is an upgrade and costs nothing.
Software version
How current its kaspad release is. The newest official release scores 100. After a 14-day upgrade window, one release behind scores 75 and anything older 50.
Reach
How many public nodes keep its address, against typical nodes of the same age. Half again typical or more scores 100; a typical node scores 67.
Clock
How far its clock is from ours, from the time in its handshakes. Within 2 seconds scores 100, within 10 seconds 50.
Responsiveness
How long it takes to answer on top of the network's own delay. Within 25 ms scores 100, falling to 0 at 150 ms.
Sync
How far behind the network its chain was when it answered. Within 5 seconds scores 100, falling to 0 at a minute.
Track record
How many of the last 180 days it answered one of our crawls. A new node starts at 50, rising to 100 at 95% of the days we crawled.
Hosting diversity
How crowded its hosting provider is among public nodes. 10% of them or fewer scores 100, falling to 50 at 30%.
Block relay
How soon it passes on new blocks, from the sensor nearest it. Within 50 ms scores 100, falling to 0 at 500 ms.
Each part is a percentage, and the score is their weighted average out of 100. A part we could not measure for a node is left out, and the others share its weight.
How uptime is measured
Our crawler visits every public node about every three hours and checks known nodes again in between. Sensors in Nuremberg, New York and Singapore connect to each node every hour. An hour counts as up if any of them reached the node, and as down if every one that tried failed, so one lost connection on one route does not cost a node anything.
When one of our own probes reaches far fewer nodes than usual in an hour, the problem is ours: its failures that hour are left out. A probe that has never reached a node, such as a sensor without an IPv6 route, does not count against it either. Once a node is down we try it less often, so an hour nobody tried keeps the last known state for up to six hours. Uptime counts from when we first saw the node.
How restarts are counted
kaspad picks a new peer id every time it starts. When a node answers two crawls in a row with different peer ids, it restarted in between. If its user agent changed too, such as a new version, the restart was an upgrade. A restart while the node was missing from a crawl is part of an outage, which uptime already counts.
How reach is measured
kaspad picks its peers from its address book, so a node whose address more of the network keeps is chosen more often: it carries more connections, relays more blocks and helps more new nodes sync. In every crawl we ask each public node for its address book and count how many list each node, counting nodes in the same subnet once so one operator's many nodes cannot inflate it.
An address takes months to spread: a typical new node is known to about 20 public nodes after a week, 40 after four weeks and 60 after eight, against about 95 for a node that has been around for months. So each node is compared with typical nodes of the same age, and a new node is never marked down for being new. Its first week is not judged at all. Full marks take half again the typical reach, about the best-known twentieth of nodes, so healthy nodes are not all tied on 100.
The clock
Every handshake carries the node's own time, which we compare with ours. kaspad checks each new block's time against its own clock, so a node whose clock drifts judges new blocks wrongly. The part takes the median over 30 days, so one slow handshake does not count.
Responsiveness, sync, track record, hosting and relay
These five were measured and shown on node pages from October, and count towards the score from 1 November 2026. Their thresholds were set from a month of data so that a healthy node keeps full marks and only the slow tail loses points; none moves a node by more than 5.
Responsiveness is how long a node takes to answer on top of the network's own delay: we ping it over the Kaspa protocol, then time a plain connection to it straight after from the same place, and take the difference. An overloaded node answers late and passes on new blocks late. The score reaches 0 at 150 ms from 16 November 2026; it was 500 ms, a figure set before we took our crawler's own delay out of the timing. Since then no ranked node's figure has been above 80 ms.
Sync is how far behind the network its chain was when it answered. Every node tells us the newest block it has, and every synced node's block count rises at the same rate, so comparing it with the nodes furthest ahead at the same moment shows how far behind it is, without trusting anyone's clock but ours. Synced nodes sit one to three seconds behind the lead, mostly from our own timing, so only a node clearly behind loses points.
Track record is how many of the last 180 days a node answered at least one of our crawls. The other parts look at 30 days, so a node that has been there since spring and one started last month can score the same; this tells them apart. A day counts if the node answered any crawl that day, so one missed crawl never costs it. A new node starts at half marks, so it is not buried while its record builds.
Hosting diversity is how crowded a node's hosting provider is among public nodes. A provider with 10% of them or fewer scores 100, falling to 50 at 30%. Many nodes on one provider's servers look like a big network but share that company's outages and terms, and adding more there lowers every node on it. One company's several networks count as one provider.
Block relay is how soon a node passes on new blocks. Kaspa makes ten blocks a second; each hour our sensors in three regions listen to every public node for five minutes and time each announcement against the first node to make it. The distance to the sensor is taken off, and each node is judged from the sensor nearest it, so where it is matters as little as we can make it.
Getting a place
A node joins the leaderboard once it has been seen online on 10 separate days in the last 30. Until then it is listed without a place. Places come only from complete crawls, so the leaderboard starts once there are 10 days of them.
Only answers from nodes following the Kaspa network count. A node still on software from before the Toccata hard fork, or whose newest block was more than 10 minutes old when the crawl began, answered but is counted as not seen in that crawl, and as down until it answers from the current network again.
A node reachable at more than one address, such as IPv4 and IPv6, answers each crawl at each of them. It keeps one row on the leaderboard: a claimed address first, otherwise the one that answered most crawls. Its other addresses point to that row.
When a new release comes out, the one before it keeps full marks for 14 days while operators upgrade (30 days until 16 October). Versions are judged against the official releases, not what other nodes report, so an unreleased build cannot mark everyone else down. A version we cannot read from its user agent scores 50.
Ties
Nodes can still share a score. Every node gets its own place: between nodes on the same score, the place goes to the one that has gone longest without a down hour, then to the one more of the network knows, then to the one we have known longest. The leaderboard marks a shared score with =, and a node's page says how many others share it.
The latest crawl
A node's page also shows a reading from the latest crawl alone, worked out the same way over its last 24 hours. It shows how the node is doing right now; only the 30-day score sets its place.
Claiming a node
Running a node and claiming it puts your name on its row. Claiming never changes a score: every node, claimed or not, is measured the same way.