Question 1

How often are predictions evaluated and how long does it take to get a score?

Accepted Answer

Each prediction is scored exactly once when its outcome settles: 7-day predictions settle approximately 10 days after the action (7 days + 1 + 2-day settle window), 14-day at approximately 17 days, and 28-day at approximately 31 days. Scores are published the same day they settle; the leaderboard updates daily. A model entering the network gets its first scores 10 days after its first submitted basket.

Question 2

Can I verify a prediction score myself, or do I have to trust the operator?

Accepted Answer

You can recompute any score yourself. Every daily receipt—outcomes, predictions, score components, and the exact formula—is published and hash-chained. The tool `scripts/verify_day.py` downloads the receipt and reproduces your scores locally. Receipts are optionally anchored on-chain; you can read the rolling Merkle root from the chain and verify the feed linkage independently.

Question 3

What does it cost to start running a prediction model on SN21?

Accepted Answer

Registration burns a small amount of TAO (~0.0005 on testnet; mainnet pricing varies). You must hold SN21 alpha per a published ramp (starting at 150 on 10 August 2026, stepping to 1,000 by 15 September). Emissions can flow toward meeting the hold as your model earns; there is no upfront buy-in. Hardware: 2 CPU cores @ 2 GHz, 4 GB RAM, 5 GB SSD minimum for a miner.

Question 4

How do I ensure my model doesn't get copied and misattributed to another miner?

Accepted Answer

Your model identity is tied to its digest (SHA256 of the container image). Commit your digest on-chain *before* making the image publicly pullable; this establishes precedence. If another miner runs an identical or behaviorally similar model later, SN21's copy detection groups them and only the first commitment earns. Rebuilding your own image is safe: identity follows the behavior in the published daily record, not the hotkey.

Question 5

What prevents a miner from predicting only on easy episodes and skipping hard ones?

Accepted Answer

Missing an episode costs a full-weight zero in your moving-average standing—the same standing mass a covered episode contributes. Covering an episode at any honest score always leaves your standing at least as high as skipping it. Additionally, a participation gate requires 75% coverage on each day's basket for bridge eligibility; sustained low coverage decays your weight to zero.

Question 6

How are emissions allocated—is there a winner-take-all cliff?

Accepted Answer

Emissions follow a published weight curve: rank 1 receives 50% of miner emissions, rank 2 gets 25%, rank 3 gets 10%, then a geometric tail (each subsequent rank receives 80% of the previous, capped at 20 earners). Below the score threshold you earn nothing. There is no cliff: as your standing rises through the curve, weights move gradually. Standings are per-prediction moving averages with age decay; thinner days simply contribute fewer rows.

AdTAO

AdTAO

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup