Subnet 83: CliqueAI
CliqueAI is Bittensor Subnet 83. Public CliqueAI materials describe it as a maximum-clique solving subnet where miners return graph solutions and validators score those solutions for validity, quality, and diversity.
What CliqueAI Rewards
The CliqueAI README frames the subnet around a four-stage mechanism: problem selection, miner selection, scoring, and weight setting. The work target is the maximum clique problem, where a submitted answer is a set of graph vertices that should form a clique under the challenge constraints.
The mechanism documentation describes submitted solutions as vertex sets and scores them on optimality and diversity. Optimality focuses on whether the solution is valid and how its clique size compares with other returned solutions for the same problem. Diversity rewards valid solutions that are not identical to many other miners’ answers.
This scoring shape matters because maximum clique solving can converge on duplicated answers. A subnet that rewards only size can overvalue copycat solutions once a strong clique is known. CliqueAI’s diversity component gives miners a reason to search for strong alternative solutions rather than simply returning the most common answer.
Miner and Validator Roles
CliqueAI validators coordinate the challenge loop. They choose problem types, retrieve problems, filter eligible miners, sample miners with a difficulty-adjusted probability, evaluate submitted solutions, and update weights from historical performance. This makes validators the layer that turns graph-solving results into Bittensor weights.
Miners compete by producing strong clique solutions under the challenge constraints. Because the score combines solution quality with diversity, the subnet rewards miners that find large valid cliques while still leaving room for different solver approaches.
Problem Distribution Context
CliqueAI’s mechanism docs classify problems by graph scale and time limit. Harder problems are not sent to every eligible miner with the same broad reach as easier ones; the mechanism adjusts miner sampling by problem difficulty. That keeps challenge distribution tied to expected difficulty rather than treating every graph as equivalent.
The same docs also describe weight setting through an exponential moving average of historical solution scores. That means one good answer is not the whole miner reputation. Past performance continues to matter, while newer challenge results can still move a miner’s score over time.
At the subnet level, CliqueAI is therefore a repeated graph-optimization contest. Validators select problems, miners submit candidate cliques, and scoring balances correctness, relative clique size, diversity, and performance history before weights are set.
Difficulty and Diversity Boundary
The CliqueAI README describes the subnet as a four-stage mechanism: problem selection, miner selection, scoring, and weight setting. That structure is useful because it separates challenge distribution from solution quality. A miner first has to receive a problem; only then can its returned clique be scored.
The mechanism documentation adds that problem difficulty affects miner selection. Easier problems can reach more miners, while harder problems have a lower selection probability for eligible miners. The challenge pool is therefore not a flat queue where every graph has the same distribution pattern.
Difficulty also changes how scoring should be read. The mechanism docs say performance should carry more weight than uniqueness, and that difficult problems emphasize performance more because they offer more room for improvement. Diversity still matters, but it is not a replacement for solving a hard graph well.
This distinction helps explain why CliqueAI is not only a “largest clique wins” subnet. A submitted solution has to be valid, competitive in size, and not merely a duplicated answer when diversity is being rewarded. Strong miners are therefore competing on both graph-solving strength and the ability to produce useful alternatives under the current scoring rules.
The weight-setting step adds a time dimension. The mechanism docs describe miner ratings as historical solution scores updated over time, so one strong submission does not fully define a miner’s standing. Repeated valid, competitive solutions are what make the performance signal more durable.
For readers, Subnet 83 is best understood as a difficulty-aware graph-optimization market. Validators decide which problems to distribute, miners return vertex sets, and scoring combines validity, relative clique size, diversity, and historical performance before weights are updated.
References: CliqueAI README, CliqueAI mechanism documentation
On-Chain Identity
Live SN83 subnet data is available on TaoStats. The source-backed graph-solving and scoring details in this article come from public CliqueAI materials rather than from live identity fields.
Relationship to Yuma Consensus
Subnet 83 uses Yuma Consensus to convert the graph-solving performance weight vectors that validators submit into the emission shares distributed to miners and validators within the subnet each tempo. The Yuma Consensus documentation describes how validator weight submissions are aggregated into consensus weights for each miner registered on the subnet.
In CliqueAI’s context, validators evaluate submitted clique solutions by scoring validity, relative clique size compared to competing solutions, diversity against other miners’ answers, and historical performance across prior rounds, combining those dimensions into weight vectors for the subnet. The Emission documentation describes how those consensus weights determine each participant’s share of the subnet’s accumulated emission each tempo.
Development Stage Context
The Introduction to Bittensor describes subnet development as moving from localnet to testnet and then mainnet. For Subnet 83, that sequence applies to the standard Bittensor lifecycle: localnet for isolated development, testnet for shared non-production testing, and mainnet for live operation with real emissions.
On mainnet, Subnet 83 is registered as the live production subnet at netuid 83. The Bittensor Networks reference separates mainnet, testnet, and localnet. Participation examples or emission outcomes from one environment should not be read as representing production subnet performance in another environment.
Reader Boundary
Subnet 83 CliqueAI should not be read as generic Bittensor subnet documentation, a pure largest-clique leaderboard, or a guarantee that one valid submission settles a miner’s standing. The CliqueAI mechanism documentation scores vertex-set solutions on validity, relative clique size, diversity against other miners, and historical performance.
Diversity Penalizes Duplicate Answers on the Same Problem
The mechanism docs reward valid solutions that are not identical to many other miners’ answers when diversity is applied. Copying a known maximum clique can score poorly even if the clique is large, because the subnet is designed to discourage convergent copycat submissions across miners facing the same challenge.
Weight History Uses an Exponential Moving Average
The same mechanism docs describe miner ratings as an exponential moving average of historical solution scores rather than a single-round winner-take-all ranking. One strong clique can help, but sustained competitive submissions build the performance signal validators translate into weights over time.
Validator weights still flow through Yuma Consensus to determine emissions each tempo (Yuma Consensus, Emission).