Start here

How It Works

The one fixed architecture principle, the three roles, and the 24-hour round lifecycle.

The one fixed principle

Miners submit frozen artifacts. Validators execute those artifacts themselves, offline, on private data.

Everything else follows from this. The principle:

removes miner endpoint availability from scoring;
prevents model swapping after submission;
blocks proxy calls to frontier APIs;
makes every result reproducible;
lets any dispute be settled by re-execution.

The three roles

Role	Responsibility
Miner	Trains the best ASL to English system it can. Publishes open weights and a loader class in one public repository, and commits a manifest that pins them by commit SHA and names a network runtime profile. Builds no container image. Performs no work during scoring.
Validator	Freezes commitments, verifies artifacts, draws a private challenge set, executes every valid artifact offline, scores it, applies gates, sets weights, publishes the record.
Subnet owner	Maintains the private corpus, the reference miner, the scoring specification, the protocol constants, and the commercial deployment path for the champion.

A round, end to end

Rounds run on a 24-hour cadence. Commitments freeze at the start of a round; anything late moves to the next one.

Miner commits manifest on chain→Round opens — commitments freeze→Validator reads frozen commitments→Pulls the named network runtime; fetches the repository at its commit SHA→Verifies hashes, loads the model, counts parameters, checks determinism→Draws a legal 300-clip challenge set→Drives each model with its own runner — no network interface→Scores translation, grounding, generalization, pose→Sets miner weights on chain→Publishes the scoring record

What the validator does, in order

Freeze on-chain commitments.
Fetch the repository at the pinned commit SHA and verify every declared file, the loader module included, against its declared SHA-256 hash.
Pull the runtime profile the manifest names, mount the repository read-only at /model, import the entry class, call load() with no network, and count the loaded parameters by traversal. The manifest is checked, not trusted.
Run the determinism fixture: the same clips executed twice at the declared seed must produce byte-identical output.
Draw a legal private challenge set.
Extract shared reference pose keypoints once.
Execute every valid artifact with networking disabled.
Parse outputs.
Compute translation, grounding, generalization, and pose scores.
Apply hard rejection gates.
Compute miner weights.
Publish the scoring record without publishing the active private clips.

What makes a round replayable

Model repositories are pinned by commit SHA and per-file hash. The runtime each one executes in is a profile the manifest names, never a digest; the network publishes that profile as an image pinned by digest when it releases the runtime. Evaluation seeds are deterministic. Derangement seeds and bootstrap procedures are fixed by round identity.

The same round can therefore be replayed later, with the same artifacts and the same corpus state, and reproduce the same scores. That property is what turns a score into evidence.