One outcome market#
Ormas is Bittensor subnet 76. It rewards accepted coding outcomes rather than inference turns or benchmark scores. Bring a bounded GitHub task and an automated check that defines success. Providers offer a firm price or an estimate with a hard limit before work starts. An accepted delivery returns a result branch and receipt; a failed acceptance has no task charge. You review the code and decide what merges.
Providers compete through the result and price. Their models, harnesses, planning and repair loops, hardware and operating costs are choices they optimize privately. Clients select task and service requirements, not a model. Production verification is currently operated by Ormas, which also operates providers during the alpha. Separate test execution does not establish separate ownership.
Standard — live in early alpha#
Standard supports public and private GitHub repositories with explicit third-party execution consent. The assigned provider and Ormas-operated checker can read the code. The provider may send code and task text to its own AI services. Standard makes no confidential-computing or zero-data-retention promise.
GitHub visibility and Ormas service level are separate: a private repository at Standard is shared with the assigned provider and checker. Connecting a private repository does not turn on Protected. See where your code goes and the supported task environments before submitting. The client service is currently US-only.
Protected — production evidence, wider rollout in testing#
An operator-run Protected job settled on October 8, 2026 with both miner and validator in confidential VMs. The next client offering and independent-miner setup are still in testing. They are not open enrollment, and no opening date is promised.
The measured worker environment uses Intel TDX hardware attestation before repository credentials are released. The confined worker is designed to reach approved inference routes, rather than arbitrary internet endpoints or a miner’s own servers. Approved custom workers and inference routes must pass the rollout’s qualification gates.
That design does not yet justify a blanket claim that code cannot be extracted or that every provider guarantees zero retention. Confidential execution, inference location and provider retention policy are distinct boundaries.
Built with other Bittensor inference networks#
The upcoming Protected offering uses SayGM, SN28, and Engy, SN53, as its initial inference options. This connects SN76’s coding outcomes to inference supplied by other Bittensor subnets. The planned client controls distinguish three kinds of route:
| Planned route | What it establishes | What it does not establish |
|---|---|---|
| SayGM frontier | Requests pass through SayGM’s attested gateway. | The upstream model provider does not thereby run inside confidential hardware. |
| SayGM TEE | The provider attests that model inference runs inside a confidential environment; Ormas also checks the served-model identity. | This is not an additional independent attestation of the provider’s entire inference path by Ormas. |
| Engy zero-retention | Approved routes use data collection denied and the applicable provider retention commitment. | This is a provider-policy boundary, not TEE inference. It does not cover all Engy routes. |
SayGM’s privacy model explains gateway and inference boundaries. Engy’s privacy policy scopes its retention commitment to the routes it operates and marks ZDR; its permissionless-network routes do not carry the same technical protection. Ormas is qualifying the approved routes, rather than treating either provider’s name as a universal privacy guarantee. Other Bittensor providers are a future expansion, not an available option today.
Clients are intended to choose the Protected route types they accept when submitting a task. Miners keep their accounts and implementation choices private within those requirements. Inference-provider choice does not change validator weight: the subnet rewards accepted outcomes.
For developers and miners#
Have a bug fix, small feature or maintenance ticket with a clear check? Start with the first-task guide and bring the change plus your definition of success. Review every returned branch before merging.
Miners can start with the Standard onboarding guide. Protected enrollment is coming. Your provider and model choices remain private within the task’s approved requirements, and inference-provider choice does not affect validator weight.
Status checked against the October 12, 2026 operator update. Availability can change; the selected task’s agreed terms and settled receipt remain authoritative.