Question 1

How does Leoma ensure all validators grade the same model the same way?

Accepted Answer

All duel parameters—the corpus manifest digest, frame resolution, metric (LPIPS/MSE/SSIM), and confidence threshold—are pinned in chain.toml and hashed into a consensus_digest that is sent with every eval request and echoed back in the verdict. Held-out test clips are selected from a digest-pinned manifest (built offline), with each clip's window and ground-truth hash fixed in advance. The block hash seeds the duel deterministically, so no validator can change which clips are tested or produce different scores for the same model without the mismatch being detected.

Question 2

What prevents miners from overfitting to the test set?

Accepted Answer

The block hash is unpredictable until a block is mined, so miners cannot see which clips will be evaluated before they submit their model. The held-out clips are drawn from a pinned manifest that is versioned in git and rebuilt offline; live bucket listings are never used at duel time. This combination ensures the test set is reproducible across validators while remaining opaque to miners.

Question 3

How is model submission and inference decoupled?

Accepted Answer

Miners upload fine-tuned weights to Hippius Hub (a content-addressed registry) and commit an immutable repo@digest reveal on-chain. They do not host inference or an API. Validators independently download the weights by digest and run the model themselves on their own GPU eval servers. This removes the need for miner infrastructure, prevents model corruption in transit, and ensures every validator can verify the exact snapshot they are evaluating.

Question 4

What happens when a challenger model wins?

Accepted Answer

If the challenger's distance to ground truth is confidently better than the king's (bootstrapped lower-confidence bound exceeds delta_threshold), it is crowned. The deposed king is added to a bounded king chain (up to 4 prior kings), and the validator sets equal weights across all active kings. Emission is split equally across the active chain; when there is no king, 100% burns to UID 0.

Question 5

Can models be pre-screened before the full duel?

Accepted Answer

Yes. Validators reject models that do not match the pinned base architecture (config.json shape mismatch) in seconds without using GPU time. Models that repeat the conditioning frame (freeze cheat) are also caught by a deterministic gate. Only architecture-compliant candidates proceed to the full duel.

Question 6

What is a reference-based duel metric?

Accepted Answer

Leoma scores generated frames against known ground-truth video continuations using LPIPS (learned perceptual image patch similarity) by default, with MSE and SSIM as alternatives. Lower distance scores are better. Both the king and challenger generate continuations on the same held-out clips using an identical per-clip seed, and their distances to the true continuation are compared statistically.

Thirty Spokes

Thirty Spokes

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup