Subnet 37: Aurelius
Aurelius is Bittensor Subnet 37, a subnet focused on moral-reasoning alignment. Its goal is to produce training data that improves how language models reason about ethical dilemmas. Miners contribute structured ethical-dilemma scenarios, validators check and simulate the accepted ones, and the resulting simulation transcripts become a dataset aimed at raising model performance on moral-reasoning benchmarks. The incentive-mechanism code is maintained in the Aurelius-Protocol/Aurelius-Protocol repository.
What the Subnet Produces
The subnet’s product is not a live service or a model but a corpus of simulated moral-reasoning transcripts. A miner supplies a scenario — a structured description of an ethical dilemma — and, once it passes validation, the scenario is played out as a multi-agent simulation. The transcript of that simulation is scored and kept as training data. Because the value lies in novel, well-formed scenarios that survive simulation, the mechanism rewards useful new dilemmas rather than volume.
Scenario Design Context
The Aurelius Protocol repository describes miner work as a library of operator-authored scenario configurations. That is an important boundary: miners are not submitting the final transcript or a model weight update directly. They are rewarded for supplying structured dilemma settings that can be checked, simulated, and turned into training examples.
This makes the scenario itself the unit of contribution. A useful scenario needs enough structure to define the moral tension, the participating agents, and the decision or reflection setting that will be simulated. It also needs to be novel enough to add value to the dataset after validation. The repository’s validation pipeline exists because malformed, repetitive, or low-quality scenarios do not create useful moral-reasoning data.
For readers, this explains why the miner contribution is upstream of the transcript: a scenario configuration can generate a simulation only if it passes the validator pipeline.
Concordia Simulation Context
Aurelius uses Concordia for accepted scenario simulations. Concordia describes itself as a library for generative agent-based models in grounded physical, social, or digital environments. Its “Game Master” pattern resolves natural-language actions from simulated entities into outcomes inside the environment.
That source context matters because Aurelius does not merely store a scenario prompt. Once a scenario is accepted, it becomes an interaction among simulated agents. The resulting transcript is the artifact that can be used as moral-reasoning training data, because it records how agents act and respond inside the dilemma rather than only listing the dilemma’s premise.
The Concordia connection also clarifies the validator role. Validators are not only filtering text; they decide which scenario configurations are good enough to spend simulation work on. The pipeline therefore protects the quality of the transcript corpus before simulations are run.
Miner and Validator Roles
Miners run an axon that serves a library of operator-authored scenario configurations. When a validator requests one, the miner returns the next scenario from its library, signed by its hotkey so the work can be attributed to it. Miners do not generate scenarios on demand; they curate a library ahead of time.
Validators put each submitted scenario through a multi-stage acceptance pipeline before it can earn anything. The pipeline checks the scenario’s format, rejects ones that are too similar to scenarios already seen, applies a quality gate, and only then runs the scenario as a sandboxed generative-agent simulation using Google DeepMind’s Concordia framework. A scenario that fails any stage is dropped; one that passes all of them produces a transcript that is scored and used to set the miner’s on-chain weight. Those weights feed into Yuma Consensus, which converts them into the incentive split across miners and validators.
The subnet also uses a work-token balance to rate-limit submissions: a validator spends a token for each scenario it accepts for simulation, so a miner has to maintain a balance to keep earning. This ties the cost of evaluation to real on-chain accounting rather than letting miners flood validators for free.
On-Chain Identity
Live SN37 data, including metagraph state and alpha token pool information, is available on TaoStats. The live Finney identity for netuid 37 registers the subnet name as Aurelius, with the description “Decentralized Alignment of Artificial Intelligence,” and points to the Aurelius-Protocol/Aurelius-Protocol repository, which is the source of truth for the validation pipeline and scoring described above.
Relationship to Yuma Consensus
Subnet 37 uses Yuma Consensus to convert the simulation-quality weight vectors that validators submit into the emission shares distributed to miners and validators within the subnet each tempo. The linked documentation describes how validator weight submissions are aggregated into consensus weights for each miner registered on the subnet.
In Aurelius’s context, validators run each submitted moral-reasoning scenario through a multi-stage acceptance pipeline — format checks, novelty filters, a quality gate, and a Concordia generative-agent simulation — and translate the resulting transcript scores 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 Aurelius (SN37), that sequence changes how readers should interpret moral-reasoning simulation examples and alignment dataset evaluation outcomes.
In localnet, Aurelius-compatible miners and validators can be developed and tested in an isolated environment. Localnet moral scenario evaluations and emission outcomes do not represent production subnet performance.
On testnet, Aurelius-compatible Concordia simulations can be exercised in a shared, non-production network. Testnet scenario quality assessments and validator scores are separate from mainnet subnet state.
On mainnet, Aurelius (SN37) is the live production subnet where miners produce moral-reasoning scenarios for Concordia simulation and validators evaluate those contributions to determine real Bittensor emissions. The Aurelius Protocol repository describes the mechanism that applies on the production network.
The Bittensor Networks reference separates mainnet, testnet, and localnet. A moral scenario result or emission outcome from one environment should not be read as representing production subnet performance in another environment.
Reader Boundary
Subnet 37 Aurelius should not be read as generic Bittensor subnet documentation, a chatbot subnet, or proof that one transcript alone is the miner submission. It names one subnet’s moral-reasoning scenario and simulation pipeline on netuid 37 (Understanding Subnets, Glossary: Netuid).
Scenario Configurations Are the Miner Unit
The Aurelius Protocol repository describes miners as supplying a library of operator-authored scenario configurations rather than final transcripts or model weights (Aurelius Protocol repository).
Validators evaluate whether those structured dilemma settings are fit to simulate and convert into training examples.
Concordia Simulates Accepted Scenarios
Accepted scenarios are run through Concordia’s generative agent-based simulation model (Concordia repository).
The resulting transcript records agent behavior inside the dilemma rather than only the scenario premise.
Validator Weights Still Flow Through Yuma Consensus
Subnet 37 uses Yuma Consensus to convert validator weight submissions into emission shares each tempo (Yuma Consensus, Emission).