Hone
SN5Trains AI models to make better predictions using a hierarchical learning approach
Hone is an ARC-AGI-2 reasoning benchmark subnet on Bittensor. Miners do not run solvers directly during evaluation, they point to a git repo containing their solution. Validators clone the repo, run the solver in a secure GPU sandbox, and score miners by exact match rate on ARC problems. Only the top 5 miners above a 20% accuracy floor receive rewards.
// Reasoning benchmarks, executed in a sandbox.
Hone is a Bittensor subnet built around running the ARC-AGI-2 reasoning benchmark. The key idea is that miners provide solutions in git repositories rather than running solvers directly at query time. Validators clone a miners repo, build it inside a secure GPU sandbox, and execute evaluation tasks with network access during prep and no network access during inference.
The simple version: Like a Kaggle competition where you submit code that our judges execute in a clean box.
Centralized equivalent: Multi-party ML evaluation services like benchmarks with reproducible leaderboards.
How it works:
- Miners expose an HTTP endpoint that points to their solution repository.
- Validators query miner info, submit jobs to a Sandbox Runner service, which clones repos, builds container images, and runs ARC problems with controlled environment.
- Scoring uses exact match rate, with a 20% accuracy floor and exponential decay rewards for the top 5 above the floor.
-
The problem it solves: Running sophisticated reasoning benchmarks fairly is hard. Submissions can be fragile, environment dependent, and hard to reproduce. Hone forces participants to provide reproducible Git based solutions and evaluates them in an isolated GPU sandbox.
-
The opportunity: ARC-AGI-2 is widely recognized as a challenging test for general reasoning. By turning it into a live Bittensor subnet, Hone incentivizes progress on tasks that go beyond pattern matching or simple completion.
-
The Bittensor advantage: Decentralized evaluation adds resistance to collusion and reduces trust in any single judge. Multiple validators can independently run the same Sandbox Runner and weight miners based on consistent performance metrics.
-
Traction signals: Market data shows price around 0.01906 TAO, market cap near 89,750 TAO, and negative 7 day net flow of -677 TAO. Emission share is about 328.89 percent. Root proportion is about 0.168, which points to moderate protocol subsidy in the current price.
Category: Code Generation and Development Tools | Centralized Competitor: ARC challenge infra
Hone is not about training models or running inference on demand. It is a standardized evaluation framework for ARC-AGI-2 reasoning tasks. The important design choice is that miners submit git repos as solutions and the subnet evaluates them in a Sandbox Runner instead of expecting live on-chain calls.
Mechanism:
Miners expose an HTTP info endpoint pointing to repo info. Validators call the Sandbox Runner API to clone the repo, build the solution as a Docker image, run prep steps with network access, and then execute inference tasks without network. This two phase execution isolates the environment effectively and reduces dependency on live network services or external tooling during the actual scoring run.
Scoring is based on exact match rate on ARC problems. There is a 20% accuracy floor, meaning miners below that threshold do not qualify for rewards. Only the top 5 miners above the qualify floor receive rewards, distributed with an exponential decay factor of 0.8 per rank. If no miners meet the floor, 100% of the rewards are burned.
Submission caching is also built in. Identical repo, branch, and commit combinations reuse cached scores instead of re-running evaluation, which reduces duplicate work and incentives constant meaningless resubmission.
On market structure, Hone shows mixed signals. Root proportion is elevated at about 0.168, meaning a significant portion of the current price is coming from protocol subsidy. Gini is about 0.754, which points to fairly concentrated ownership or stake distribution. Recent 7 day net flow is negative, suggesting some capital is rotating out.
The subnet is technically sound. The combination of git based submission, Sandbox Runner isolation, submission caching, and exact match scoring is a reasonable way to run a long form reasoning benchmark at scale. The real test over time will be whether enough high quality ARC-AGI-2 solutions exist to keep the top 5 above the floor without burning rewards.
-
Talent bottleneck. ARC-AGI-2 is a hard benchmark. If only a few miners can consistently solve problems and meet the accuracy floor, the top 5 distribution may become stable and reduce competitive pressure.
-
Evaluation cost. Running full repo builds and GPU inference for every submission is expensive. The Sandbox Runner architecture needs solid compute and storage backing to scale without slowing down the subnet.
-
Root subsidy reliance. Current root proportion above 0.5 means price is supported partly by protocol emissions rather than only organic demand. That is normal for new subnets but not an ideal state long term.
-
Gini concentration. High concentration means ownership or stake is not broadly distributed yet. This can make the subnet look like an insiders club if it persists.
