Subnet 5: Hone
Hone is Bittensor Subnet 5, a subnet where miners compete on ARC-AGI-2 abstract reasoning tasks. The subnet’s repository frames the work as solver evaluation: miners point to solution repositories, and validators use a managed GPU sandbox to evaluate those solutions (Hone repository, Subnet 5 on TaoStats).
What the Subnet Evaluates
Hone evaluates whether a submitted solver can answer held-out ARC-AGI-2 reasoning problems. The task is not ordinary model hosting or prompt serving. It is benchmark-style reasoning evaluation where the useful output is a score for how many problems the solver answers correctly (Hone README).
This makes Subnet 5 easier to read as an evaluation subnet. The miner contribution is a solver, the validator-side work is scoring, and the subnet reward path depends on those scores.
Sandbox Evaluation
The sandbox design is central to Hone’s task. A sandbox runner executes submitted solver material, allows preparation before scoring, then runs inference in an isolated phase before calculating the score (Hone README).
That boundary keeps the score from being a simple self-report. A miner does not merely claim a result; the solver is evaluated in a controlled execution environment. The sandbox context is the reason Hone’s score can stand for benchmark performance rather than ordinary service uptime.
Scoring Model
Hone’s main metric is exact-match accuracy over ARC-AGI-2 problems. The scoring model also includes a minimum accuracy floor, a top-ranked reward focus, cached scoring for identical solution references, and burning when no solver qualifies (Hone README).
Those details show what the subnet is trying to reward. Hone is tuned toward correct reasoning on the held-out task, not toward broad participation for its own sake. The cache and submission limits also keep repeated identical solver material from being treated as new progress.
Sandbox exact-match scores shape the validator weights submitted on netuid 5; Yuma Consensus then settles those weights into emission shares each tempo—a high ARC-AGI-2 score is subnet evaluation input, not the emission payout itself, and it should not be read as general reasoning ability on other benchmarks (Hone README, Yuma Consensus, Emission).
On-Chain Identity
Live SN5 data is available on TaoStats. The Finney identity reports the subnet name as Hone. It lists the GitHub repository as manifold-inc/hone, which documents the ARC-AGI-2 evaluation behavior described above (btcli reference).
Concept Boundaries
Subnet 5 should be read as a Hone overview, not as miner onboarding material, a validator operations guide, or a hardware shopping list. The stable concept is ARC-AGI-2 reasoning evaluation through a managed sandbox, with Bittensor incentives tied to validator scoring (Hone README, Yuma Consensus).
Project-specific requirements, service endpoints, solver packaging details, and troubleshooting belong in Hone’s own repository. The stable distinction is that Hone rewards sandboxed reasoning performance, not generic compute availability.
Miner and Validator Roles
Subnet 5 operates under the standard Bittensor two-role structure. Miners supply the subnet’s capability and validators evaluate those contributions and set weights. Reward distribution follows Yuma Consensus.
Distinction from Yuma Consensus
Hone on netuid 5 runs an ARC-AGI-2 reasoning subnet that evaluates solver repositories in a secure GPU sandbox. That sandbox-scoring vocabulary names how Hone measures held-out reasoning performance on netuid 5, not the on-chain step that turns validator weight submissions into emission shares each tempo (Hone README, Yuma Consensus).
Yuma Consensus is the on-chain mechanism that runs at the epoch boundary on the selected netuid. It reads the weight matrix from eligible validators, applies clipping and bonding, and converts the result into miner incentives and validator dividend shares (Yuma Consensus, Emission).
Earlier sections describe ARC-AGI-2 solver evaluation, sandbox execution, and exact-match scoring. Those inputs shape the validator weights; they do not settle emissions themselves. Yuma aggregates the weight signals that flow from that sandbox-scoring path rather than running ARC-AGI-2 benchmarks itself (Hone README, Understanding Subnets).
Sandbox scoring therefore sits upstream of settlement. Yuma still credits emissions from included validator weights on netuid 5 once the tempo closes (manifold-inc/hone, Emission).
Readers should treat Hone as the ARC-AGI-2 reasoning evaluation subnet on netuid 5, and Yuma Consensus as the recurring mechanism that allocates rewards from validator weights each tempo.
- Subnet 5 (Hone) — ARC-AGI-2 sandbox reasoning evaluation subnet.
- Yuma Consensus — epoch-boundary settlement from validator weights to emission shares on a task subnet (Yuma Consensus: Validator emissions).
Reader Boundary
Subnet 5 Hone should not be read as generic Bittensor subnet documentation, a model-hosting or prompt-serving service, or a certificate that a solver is generally intelligent. It names one subnet’s ARC-AGI-2 reasoning evaluation — scoring submitted solver repositories on held-out problems inside a managed GPU sandbox — on netuid 5 (Hone repository, Understanding Subnets, Glossary: Netuid).