Gaming it: weight copying, collusion, commit-reveal
This lesson is about the gap between judging honestly and just looking like it — and the specific fix the network built for the most obvious way to fake it.
Once judging finishes on a contest, everyone's numbers become public. That creates an obvious shortcut: a validator can skip evaluating anyone at all and simply submit a stake-weighted average of what everyone else already revealed. Because the network's consensus rule rewards agreement, copiers like this have historically earned better standing and better pay per unit of stake than the honest judges they copied — while contributing nothing of their own. [source]
The fix is to stop anyone from reading anyone else's answer until it's too late to just copy it. A validator's submitted numbers are now sealed with timelock encryption: nobody can open them — not even their own author — until a public, verifiable countdown elsewhere on the internet reaches a specific moment. Only then does the chain decrypt and apply them automatically, with no manual step for anyone to skip or abuse. [source] [source]
A copier isn't just delayed by this, either — the numbers it eventually sees are guaranteed to be at least one full concealment window stale. If the true ranking has actually moved since then, copying it lands away from wherever consensus has since settled, which is exactly the outcome that damages a copier's standing. The one case this doesn't fix: a contest whose true ranking barely changes between windows, where the stale numbers are still accurate and copying still pays just as well as judging. [source] [source]
This encryption creates its own trap for an owner who isn't careful: a subnet's immunity period has to stay longer than the concealment window, or a brand-new miner can be evicted before its very first scores ever get revealed and counted at all. [source]
None of this touches a different kind of gaming: coordinated agreement. The median that survives consensus is whichever point a majority of stake backs, and money doesn't announce whose hands it's in or whether those hands are acting together — a handful of validators voting the same way, on purpose, moves the outcome exactly as one much bigger one would. [source]
How it actually works
The randomness this all hangs on comes from drand's Quicknet beacon, which publishes a fresh, publicly verifiable signature every 3 seconds worldwide, with no keys or accounts involved on either end. On-chain, a dedicated pallet runs an off-chain worker that pulls each new pulse and submits it as a signed transaction, verified with optimized host functions rather than re-deriving the cryptography inside the runtime itself. [source] [source]
Submitting a timelocked commit means encrypting a SCALE-encoded weights payload into a TLECiphertext and posting the result on-chain, tagged with the DRAND round whose output should trigger its reveal and a version number that has to match the chain's current commit-reveal version. A hotkey is capped at 10 unrevealed commits at once, on top of its own separate rate limit. [source]
The consensus clip's guarantee is narrower than it sounds: kappa, defaulting to roughly half of active stake, sets the bar a rank has to clear — and the chain has no concept of which validators are related, commonly owned, or simply agreeing by arrangement. It only ever sees a pile of stake-weighted numbers. [source]