Subnet 23: Trishool
Trishool is Bittensor Subnet 23. Its on-chain identity registers the subnet as an AI alignment protocol built on Bittensor.
What Trishool Provides
The registered identity describes Trishool as “the AI alignment protocol built on Bittensor” — work aimed at making AI systems behave in line with intended goals. Miners contribute to that objective and the network rewards the strongest contributions. The registered project website is trishool.ai and the codebase is the TrishoolAI/trishool-phase2 repository, which defines the task in more detail than the on-chain identity alone.
Miner and Validator Roles
Subnet 23 operates under the standard Bittensor two-role structure. Miners supply a capability to the network, and validators evaluate those contributions and set weights. Reward distribution follows Yuma Consensus.
Adversarial Prompt Evaluation
The Trishool Phase 2 repository describes the subnet as building an AI guard model through adversarial prompt competition. In that framing, miners submit adversarial prompts or prompt items, and validators evaluate those submissions through an agent-and-judge process before weights are written back on chain.
That task shape makes Trishool different from a general chatbot subnet. The miner contribution is not just answering a user query; it is producing challenging alignment material that can expose unsafe or weak behavior in an AI guard setting. The validator side then measures whether the submitted material produces the kind of outcome the subnet is designed to detect.
The repository also describes a platform layer that validates submission format, checks for duplicates, and provides evaluation data to validators. For readers, that means Trishool’s scoring context combines miner-provided adversarial material, platform-mediated evaluation inputs, and validator-side assessment before the final weight-setting step.
The same source frames the judging output around safety outcomes such as safe, partial, and jailbreak classifications. Those labels are useful because they show what the subnet is trying to measure: whether a submitted prompt reveals a stronger or weaker guard behavior. Trishool’s market therefore rewards adversarial alignment inputs that matter to guard-model evaluation, not generic conversation quality.
The agent names in the repository, including OpenClaw and Judge, are project-specific tooling around that evaluation loop. At Taopedia level, the important concept is the loop itself: miners create alignment-test inputs, validators evaluate them through the subnet’s judging process, and the resulting scores become weights for Bittensor emission distribution. That is why the article treats Trishool as an alignment-evaluation subnet, not only as a registered project name.
References: Trishool Phase 2 repository, Subnet 23 on TaoStats, Yuma Consensus
On-Chain Identity
The live Finney identity for netuid 23 registers the subnet name as Trishool, with the description “Trishool is the AI alignment protocol built on Bittensor.” The registered project website is trishool.ai and the GitHub repository is TrishoolAI/trishool-phase2; a Discord channel is also recorded in the identity. Live subnet data is available on TaoStats.
Relationship to Multiple Mechanisms
Trishool has validators evaluate alignment contributions and set subnet weights. The Glossary and Multiple Incentive Mechanisms Within Subnets docs note that validators must evaluate miners separately for each mechanism.
For readers, this article documents one subnet market. If that netuid runs more than one incentive mechanism, validator scores and weights should be read per mechanism rather than as one combined path.
Relationship to Yuma Consensus
Subnet 23 uses Yuma Consensus to convert validator weight submissions into the incentive distribution for the subnet. For Trishool, those weights are downstream of validator evaluation of adversarial alignment material rather than a generic chat-quality score.
The Emission documentation describes how subnet emissions are allocated from validator weights. That makes the project-specific judging loop and the Bittensor emission mechanism related but separate: Trishool defines what validators evaluate, while Yuma Consensus and emissions describe how accepted weights affect rewards.
Development Stage Context
The Introduction to Bittensor describes subnet development as moving from localnet to testnet and then mainnet. For Trishool (SN23), that sequence changes how readers should interpret adversarial alignment evaluation examples and scoring outcomes.
In localnet, Trishool-compatible miners and validators can be developed and tested in an isolated environment. Localnet adversarial prompt evaluations and emission outcomes do not represent production subnet performance.
On testnet, Trishool-compatible components can be exercised in a shared, non-production network. Testnet alignment evaluations and validator scores are separate from mainnet subnet state.
On mainnet, Trishool (SN23) is the live production subnet where miners submit adversarial alignment material and validators evaluate those contributions to determine real Bittensor emissions. The Trishool repository describes the evaluation mechanism that applies on the production network.
The Bittensor Networks reference separates mainnet, testnet, and localnet. An alignment evaluation example or emission outcome from one environment should not be read as representing production subnet performance in another environment.
Reader Boundary
Subnet 23 Trishool should not be read as generic Bittensor subnet documentation, a general chatbot subnet, or proof that conversation quality alone defines emissions. It names one subnet’s adversarial AI-alignment evaluation market on netuid 23 (Understanding Subnets, Glossary: Netuid).
Platform Layer Validates Submissions Before Judging
The Trishool Phase 2 repository describes a platform layer that checks submission format, filters duplicates, and supplies evaluation inputs to validators (Trishool Phase 2 repository).
Miner prompts therefore pass structured validation before the agent-and-judge scoring step runs.
Alignment Labels Drive the Scoring Outcome
The repository frames judging output around safety classifications such as safe, partial, and jailbreak rather than generic response fluency (Trishool Phase 2 repository).
Those labels define what validator weights are meant to reward in guard-model testing.
Validator Weights Still Flow Through Yuma Consensus
Subnet 23 uses Yuma Consensus to convert validator weight submissions into emission shares each tempo (Yuma Consensus, Emission).