What validator weights tell you about a subnet

Every subnet runs on one loop. Miners produce something, validators score it, and the chain turns those scores into emissions. The scores are called weights, and they are the closest thing a subnet has to a ledger of opinion. Reading them well explains most of what you see on a subnet's metagraph.

From scores to weights

Each validator evaluates the work of the miners on its subnet using the subnet's own mechanism: a benchmark, a task, a comparison against ground truth. It normalises those scores into one weight per miner and submits them on-chain. Laid side by side, the weights of every validator form a matrix, with a row per validator and a column per miner.

In a healthy subnet most rows look alike, as in the figure: the same miners are bright in nearly every row, because validators are measuring the same thing. The one amber row is a validator that disagrees with everyone else. That disagreement is exactly what the next step is built to detect.

The chain does not take any single row at face value. Yuma Consensus combines the rows, weighting each validator by its stake, and produces two outputs: incentive for miners and dividends for validators.

The important property is how it treats disagreement. A validator whose weights are close to what the stake-weighted majority supports keeps its influence and earns a high VTrust. A validator whose weights stray far from that majority has the excess discounted, so it counts for less and earns less.

This is also why the system resists manipulation. A validator that inflates a few favourite miners cannot move the outcome on its own, and it pays for trying through a lower VTrust and lower dividends.

The subnet metagraph on taostats lays out every neuron on a subnet. Four columns carry most of the story.

Two validators with similar stake can earn different dividends when their VTrust differs, so stake alone is a poor predictor. The Updated column is worth a glance too: our documentation suggests keeping it under about 500 blocks, and treats anything over 1,500 as a warning sign that a validator has stopped doing its job.

A few readings that are useful, with the caveat that each is a prompt for a closer look rather than a verdict.

• Concentrated incentive. If a handful of miners take most of the incentive, either the task rewards a narrow kind of excellence or the scoring is easy to game. Check which.
• A wide spread of VTrust among the largest validators. Healthy consensus usually shows the big validators clustered together. A split suggests they are measuring different things.
• Stale weights. A validator that has not updated for a long stretch is not contributing to consensus, and its VTrust will usually sag.
• Quiet disagreement. One validator that is always the outlier may be badly configured, or it may be the only one doing the work properly. The weights cannot tell you which.

There is a shortcut around all of this. A validator can skip the evaluation and copy the consensus weights instead, which earns a high VTrust for no effort. Weight copying is controversial because it adds nothing to the subnet: the checks and balances of the network depend on validators doing their own scoring.

The main defence is commit-reveal, which is on by default for every subnet. Validators submit their weights encrypted, and the chain only decrypts and uses them after a delay set by the subnet's commit-reveal period. A copier therefore only ever sees weights that are already stale, which costs it VTrust. The subnet owner can change the period or switch the feature off, so it is worth checking the setting for the subnet you care about.

On taostats, open any subnet and go to its metagraph to sort by incentive, dividends or VTrust. The same data is available from the API through the latest metagraph endpoint, filtered by subnet. The documentation explains each of these in more detail.