Participate

For Validators

Round structure, the 300-clip draw, the sandbox, abstention rules, and validator cost.

A validator is the only party that touches private data, and the only party whose result counts. Its guarantee to the network is that it ran the exact artifact, on a legal draw, with no way for the artifact to reach the outside world.


Round structure

Rounds run on a 24-hour cadence. Commitments freeze at the start of the round. Late submissions roll to the next round.

Every scoring round draws 300 clips:

SliceClipsPurpose
Studio120Core translation quality
Broadcast90Realistic framing and media conditions
Conversational60Natural register and signer variation
Canary30Unseen conditions, grounding and generalization

Draw rules:

  • No single source recording may dominate a slice.
  • Active clips obey a 30-day reuse cooldown.
  • If the corpus cannot produce a legal draw, the validator abstains rather than weakening the evaluation.

The sandbox

The validator's container runtime pulls the image for the runtime profile the manifest names, currently torch2.4-transformers4.44. The manifest carries the profile name and never a digest; the network publishes each profile as an image pinned by digest when it releases that runtime. Submissions never bring their own image. The network's runtime image runs the model under all of the following:

  • No network interface
  • Read-only root filesystem
  • Read-only clip mount
  • Read-only model repository mount at /model
  • The validator's own runner mounted read-only; the submission contributes weights and a loader class, never a program
  • Dropped Linux capabilities
  • No privilege escalation
  • CPU / memory / GPU / wall-clock limits
  • One writable output mount
Central anti-proxy control: the design does not try to detect unauthorized API calls after the fact. It removes the interface required to make them.

Abstention

A validator that cannot execute a round honestly abstains. It does not fabricate a score and it does not reuse a stale one.

Infrastructure failure is separated from miner failure throughout the mechanism. A miner is never penalized for a validator's outage, and a validator never covers for a broken round.


Reproducibility

  • Model repositories are pinned by commit SHA and per-file SHA-256, the loader module included.
  • The manifest names a runtime profile, never a digest; the network publishes each profile as an image pinned by digest when it releases that runtime, so every validator executes the same environment and the same runner. The current profile is torch2.4-transformers4.44.
  • Evaluation seeds are deterministic.
  • Derangement seeds and bootstrap procedures are fixed by round identity.

Any round can be replayed with the same artifacts and corpus state and will produce the same scores.

The validator publishes the scoring record without publishing the active private clips.


Cost of running a validator

The V1 parameter and runtime caps are set to keep validation inside one-GPU economics. Project baseline estimates, at maximum planned scoring load:

ItemEstimate
GPU-hours per daily round10–16
GPU cost per round (reference cloud rate)$18–29
Monthly GPU cost$550–875
Total monthly operating cost (incl. artifact storage and bandwidth)$600–925

These are planning estimates from the August 2026 baseline, not quoted prices.