CliqueAI - AI-Powered Maximum Clique Solver Network
SN 83. Rank 27 of 32 by emission
CliqueAI
Emission1.1%of the block reward · rank 27 of 32
Participants232 miners · 7 validators
Registration0.1 TAO
Verified2026-09-09links checked and scores merged
The programother · github.com@0d3488a · read 2026-09-19
The work
A clique: a set of vertices in a validator-supplied graph in which every pair is connected, returned inside a time limit of 6 to 30 seconds. Graphs are general, 290 to 900 vertices in the code. Problems come from the validators' own database, each served at most once across all validators; no outside party supplies a graph.
code · raw.githubusercontent.com · docs/mechanism.md
A `solution` is a set of vertices
How it is scored
Validity first: an invalid or non-maximal clique scores zero. Optimality is exp of minus the share of miners with a larger clique divided by the miner's size relative to the round's largest, normalised to the round's best. Diversity is one over the count of miners returning the identical vertex set, normalised. The final score is optimality times one plus difficulty, plus diversity.
code · raw.githubusercontent.com · CliqueAI/scoring/clique_scoring.py
rewards = optimality * (1 + self.difficulty) + diversity
Sources disagree docs/mechanism.md gives three problem classes at difficulty 0.2, 0.4 and 1; CliqueAI/selection/problem_selector.py at the same commit ships four classes at 0.7, 0.8, 0.9 and 1.0, the fourth at 890 to 900 vertices.
How the pay splits
Top-heavy, not proportional. Per-problem scores fold into a debiased exponential moving average with alpha 0.01, a half-life of about 69 samples. At weight time the averages are min-max normalised, passed through a sigmoid centred on the median with the interquartile range as steepness, then raised to a power chosen so the top half of miners takes 80 percent of the weight.
code · raw.githubusercontent.com · common/base/validator.py
POWER_TARGET_TOP_HALF_SHARE: float = 0.80
Who judges
Seven validators were active on the 2026-09-09 chain snapshot, against 232 miners. Commit-reveal is off. A validator checks each answer itself, that it is a valid maximal clique in the graph it sent, so judging is verification rather than re-running. Each eligible miner is sampled with the same probability, lowered for harder problems, after validator hotkeys and recently updated miners are filtered out.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled False, weights_version 1; chain-2026-09-09.json active_validators 7, active_miners 232
How often it turns
Emission settles every 360 blocks, about 72 minutes, and the validator sets weights every 361 blocks by default. Problems are dispatched continuously, one at a time, with time limits sampled from 6 to 30 seconds. A new registration is immune from deregistration for 6,000 blocks, about 20 hours. The repository's last commit is 2026-07-22.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 6000
Who buys
Unknown. Looked for a customer, a price or a way to submit a graph on cliqueai.toptensor.ai, which is a title and an empty shell to a fetcher, in the README, docs/mechanism.md, min_compute.yml, the wandb table link, the chain identity and the narrative map; found none. The only paying party is emission.
not found
What has been gamed
Copying another miner's answer: identical vertex sets split the diversity credit, one over the count of miners returning them. Memorising answers: each problem is served at most once, even across validators. Invalid or non-maximal cliques score zero. Recently updated miners and validator hotkeys are excluded from sampling. No incident is recorded in the repository or on the site.
code · raw.githubusercontent.com · CliqueAI/scoring/clique_scoring.py
unq = np.array([1 / counts[sol] for sol in canonical_responses])
6 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN83 CliqueAI is a timed puzzle league: solvers get a network of a few hundred dots and a few seconds to find the largest group where every dot touches every other, and the league pays the best solvers.
The commodity, explained from zero
The puzzle is the maximum clique problem. Draw dots, connect some pairs with lines, and find the biggest set of dots in which every pair is connected. The site's headline calls this approximating NP algorithms, problems whose answers are quick to check but slow to find. No application of the problem is named anywhere in the subnet's materials.
So the commodity is a solution to a puzzle the validators themselves wrote. Validators keep a database of graphs with 290 to 700 dots and hand them out with time limits between 6 and 30 seconds. A solution "is a set of vertices". Nobody buys these solutions. For someone to pay, the subnet would need a door for a real graph, a scheduling conflict or a social network, and the materials describe none.
Why it is on Bittensor at all
The materials do not say. The mechanism document's stated aim is internal: "To empower the whole subnet, we consider both performance and diversity." No page says what decentralization changes for this puzzle, or compares the subnet to a solver run on one laptop. The honest answer from the materials is nothing yet. The comparison breaks here: a puzzle league has a sponsor who wants the puzzles solved, and this league's only paying party is the emission, the subnet's share of newly minted TAO.
How the work gets done
Miners (the solvers) run a program that receives a graph and returns a set of dots inside the time limit; a GPU is not required, and the stated minimum is 4 CPU cores and 16 GB of memory. Validators (the judges) check that the answer is a valid clique, then score it on optimality (size relative to the best answer that round), on difficulty (a factor set by graph size), and on diversity (a bonus for an answer no other miner returned). The final score is optimality times one plus difficulty, plus diversity, smoothed into a rating with a half-life of about 69 samples, which sets the weights that direct emission. The site copy says miners are chosen by stake while the mechanism document says each eligible miner is sampled with equal probability; the sources disagree.
How you would know it works
No such artifact exists on the subnet's own site. The MINER SCOREBOARD on the front page still reads "Loading" after five seconds in a browser. The docs link a Weights & Biases table of validator runs, which logs activity but shows no benchmark or comparison.
What is missing
There is no buyer, no price, no product and no support path. The chain contact CliqueAI@toptensor.ai appears nowhere on the site or in the repository, and toptensor.ai does not resolve. The Discord link is a channel in the shared Opentensor server. The site is a blank shell to anything but a browser, with no about page. The miner README is one command with no install steps; the last release, v0.0.13, is dated 2026-04-14, and the repository was last pushed 2026-07-21. The chain shows 232 active miners against 7 validators.
Go deeper
Sources for this explainer
Metaphor: a timed puzzle league with no sponsor. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 6/7- subnet_name
- CliqueAI
- github_repo
- github.com
- subnet_contact
- CliqueAI@toptensor.ai
- subnet_url
- cliqueai.toptensor.ai
- discord
- discord.com
- description
- CliqueAI - AI-Powered Maximum Clique Solver Network
- additional
- not set
Stakers and validators●●●●● 2.3
Should I allocate here? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
CliqueAI - AI-Powered Maximum Clique Solver Network
and a site headline about approximating NP algorithms, output is clique solutions to operator graphs, no unit and no buyer, needs graph theory
Q2Who it is for●●●●● 1
W&B Dashboard: https://wandb.ai/toptensor-ai/CliqueAI/table
logs validator runs, nothing on users, demand or revenue anywhere
Q3How it resists gaming or fails●●●●● 3
mechanism doc one click from the README with optimality and diversity formulas and a debiased EMA, no failure modes, site copy says stake weighted selection while the doc says equal probability
Q4Identity and documentation●●●●● 2
identity 6 of 7, discord is a channel in the shared Opentensor server (absent), url is a JS shell, github resolves, contact CliqueAI@toptensor.ai not on the site, toptensor.ai dead, no staker page
Miners●●●●● 2.8
Can I compete, and what wins? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
A solution is a set of vertices
for a general graph of 290 to 700 vertices under a 6 to 30 second limit, request and response shape only in the protocol code, no worked example
Q2Who it is for●●●●● 2
required: False
for GPU with 4 cores and 16 GB RAM minimum, a YAML file not linked from the README, registration cost not stated (burn 0.1 TAO today), no competitiveness statement
Q3How it resists gaming or fails●●●●● 4
final score is optimality times (1 plus difficulty) plus diversity in prose and math, EMA alpha 0.01 with a 69 sample half life, no example numbers, no immunity or deregistration policy
Q4Identity and documentation●●●●● 2
one start_miner.sh command and no install steps, last release v0.0.13 dated 2026-04-14, repo pushed 2026-07-21, min_compute.yml says version 0.0.17
Buyers and enterprises●●●●● 0.5
Can I use this today? · rank 30 of 32 for this audience
Q1What it is●●●●● 1
Four-stage autonomous mechanism for distributed problem solving
is the closest thing to a product statement, no API, no app, no way to submit a graph
Q2Who it is for●●●●● 0
no sourcesecondhandunchecked
no pricing and no customer type in any own or third party material
Q3How it resists gaming or fails●●●●● 0
no sourcesecondhandunchecked
no status page, no SLA, no support address, the only Discord link is a channel in the shared Opentensor server
Q4Identity and documentation●●●●● 1
no API docs, chain contact CliqueAI@toptensor.ai appears nowhere on the site or in the repo, toptensor.ai does not resolve
Newcomers●●●●● 2.0
What is this and why does it matter? · rank 29 of 32 for this audience
Q1What it is●●●●● 2
To evaluate different strengths in solving the Maximum Clique Problem, we design challenges across varying scales, time constraints, and graph structures.
JS only, jargon unexplained
Q2Who it is for●●●●● 1
To empower the whole subnet, we consider both performance and diversity.
No statement of what decentralization changes, no comparison
Q3How it resists gaming or fails●●●●● 2
a MINER SCOREBOARD section still reads Loading after five seconds in a browser, a W&B table of validator runs is linked from the docs, no dated benchmark or demo
Q4Identity and documentation●●●●● 3
subnet_name CliqueAI matches the site, description is a title not a sentence, url resolves to a JS shell, no about page, discord is a channel in the shared server
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent c 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 24 live, 0 unreachable, 0 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.