Fair checking
Why checkers rotate their addresses, how source filtering is detected, and which checker failures count against a team.
The whole scoring model rests on one assumption: the checker sees what every other team sees. A team that could tell checker traffic apart from attack traffic could serve the checker and refuse everyone else, keeping perfect SLA and defence while giving attackers nothing. Three things keep that from paying off.
The checker has no fixed address
Checks reach boxes from addresses drawn from the same range players connect from, and a different one for each check. A team can’t allowlist “the checker” without allowlisting other teams as well.
Rotation raises the cost of spotting the checker; it doesn’t make it impossible. Timing and request shape can still give a checker away under analysis, which is why detection backs it up.
Filtering is detected, not guessed
Teams are expected to firewall: rate limits, payload rules and blocking specific exploits are the defensive half of the game. What isn’t allowed is refusing every other team while answering only the checker.
Each tick, the host records connection attempts to every declared port of every box, and whether the box answered. Each port of each box is judged on its own. A port is found to be filtering only when, in that tick:
- at least three different player addresses each tried it at least twice, and none of their attempts was answered, and
- the checker, from at least two different addresses, was answered on it.
Addresses are counted, not teams: every player’s VPN connection has its own address, so three players from the same team can make up the three. One answered connection from any player address clears that port for the tick, whatever else happened; it says nothing about the box’s other ports. Too few attempts, incomplete records or traffic the host couldn’t fully observe are never a violation. UDP ports aren’t judged, because nothing generic can tell a silent UDP service from a refusing one.
This is deliberately narrow. A team that answers some players and not others isn’t caught, by design: that can’t be told apart from legitimate rate limiting. A confirmed finding marks every flag store of the service that declares that port Down on that box for the tick; nothing reopens a tick that has already been scored.
When a check fails without the team’s fault
A check recorded as Cancelled carries no evidence either way: it earns no SLA or defence, and costs none. Checks are cancelled when the platform withdraws them, when the runner can’t find the checker image it needs, and, for imported checkers, when the imported checker crashes, exits with an unexpected code or returns a result its platform can’t express.
⚠ A checker you wrote against the contract directly gets no such protection. If its container fails to start, stops answering, crashes mid-call or runs past its deadline, the check is recorded as Down against the team. Test your checker with a diagnostic before the network opens; see Diagnose failing checks.
When the checker runner reports that it can’t run checks, for example because an image is missing on the host, or the host or runner stops reporting in altogether, the game pauses rather than scoring every team down for something none of them caused. See When something fails.
Every finding, confirmed or inconclusive, is kept with the rule version that judged it, its counts and its reason, and the final data export carries them, so a filtering penalty can be explained after the event.