Skip to main content

Task Configuration

A task is a single AI job inside a Neuronet, such as serving one model for inference or training one model. Each task carries its own configuration that tells the network exactly what miners must run and how their work is verified. An Owner sets this when adding a task, and most of it is filled in automatically when the model is chosen from the registry.

This page covers the settings that belong to a task. For settings that apply to the Neuronet as a whole, see Neuronet Configuration.

The Model and How It Runs​

These settings pin down the exact computation so that any miner's result can be checked against another's.

  • Model. The specific model the task serves, identified by its hash.
  • Canonical execution settings. The exact runtime the task requires: precision (such as bf16 or fp16), attention implementation, GPU architecture where it matters, compile mode, and the seed. Miners must reproduce these settings, otherwise their results will not match on verification. See Execution Spec for the full field list and why each one is pinned.
  • Task type. Whether the task is inference, training, or diffusion. This shapes how results are verified: see Inference Verification for inference and diffusion, and Proof of Training for training runs.

Verification​

These settings decide how strict and how frequent the checks are.

  • Verification tier. Tier A requires a bit-exact match (used within a single hardware class). Tier B accepts the tiny numeric differences that appear across different hardware, within calibrated limits. See Proof of Intelligence for what each tier means and Heterogeneous Mining for the measured thresholds behind Tier B.
  • Replay rate. The share of work that is re-checked by other miners. A higher rate means more frequent verification. See Inference Verification for what a replay compares.
  • Minimum proof of compute. The Pulse score a miner must clear to take on the task, which proves it has hardware capable of the work.

Access and Privacy​

  • Privacy level. How far user data is shielded, from transport encryption up to hardware-isolated execution. See Privacy for what each level protects against.
  • Security mode. Standard for open models and public data, or Confidential for private models and sensitive data that must run inside a trusted execution environment (TEE).
  • Miner approval mode. How miners join the task. In standard public tasks miners join on their own once they qualify. For Confidential (TEE) tasks the Owner approves, invites, or removes miners directly.
  • Trust level. How strongly the miner's code and environment are attested, from inferred verification up to environment fingerprinting.

Economics​

  • Owner fee. The Owner's cut of the task's emission. It can be lowered but never raised.
  • Task weight. How a Neuronet's emission is divided among its tasks when it runs more than one.
  • Cost settings. How much capacity each request consumes, expressed in slots. See Staking for how capacity works.

How These Are Set​

Most of a task's configuration is filled in automatically when an Owner picks a model from the registry, so creating a task is a matter of a few choices rather than dozens of raw fields. Advanced users can still set the underlying values directly.

Some of these settings can be changed after a task is running, such as lowering the Owner fee or adjusting task weight. Others, like the model and its canonical execution settings, are fixed for the life of the task so that verification stays consistent.

Last updated on