Score

SN44 Score turns sports footage into a play-by-play data feed, the way a press-box statistician logs every player's position, except the statisticians are competing computer-vision models and only the most accurate one per task is paid.

The commodity, explained from zero

The product is structured data pulled from video: where each player is in every frame, where the ball is, where the pitch lines fall, and who plays which role. The README calls it "structured, decision-ready data in real time", made from raw footage, "with early deployments focused on professional sports".

It has a price because the human version is slow and expensive. The older Score Vision repository, which the site still links as its documentation, puts manual annotation at $10 to $55 per minute of footage and says the goal is to cut that cost "by 10x to 100x". That page names four kinds of customer: clubs and leagues, broadcasters, betting operators and analytics providers.

Who buys today is not stated. The front page shows customer logos as images and demo clips of football, NFL and NBA footage, but names no customer in text and gives no numbers or dates.

Why it is on Bittensor at all

The stated change is cost. The older repository says distributing the computation across miners, with a lightweight check on their output, is what makes the 10x to 100x reduction possible. The current README says only that the network "pairs expert models with a global community of validators and miners". Neither page compares this with a company running the same models on its own servers, so whether decentralization, rather than the models, makes it cheaper is not settled.

How the work gets done

Miners, the statisticians, train a model, host it on Chutes (a separate Bittensor subnet that rents model hosting) and record its fingerprint on chain. Validators, the editors, pull a video challenge, build an answer key (real, or made by a segmentation model called SAM3), send frames to every eligible miner and score the replies. Each task, called an Element, has weighted sub-scores such as overlap with the answer key, player count, smoothness and role, gated by a minimum threshold and multiplied by a difficulty factor. The highest average per Element wins its share of the emission, the subnet's share of newly minted TAO; a copied model loses the tie to whichever was committed on chain first. Here the press-box comparison breaks: a club pays its statistician whether or not they are the league's best, while Score pays one winner per Element and nobody else.

How you would know it works

No dated artifact exists on the subnet's site. The front page carries demo videos with no dates or figures, the Console is an empty shell, and the evaluation reports and 2025 draft whitepaper sit in a repository folder nothing links to.

What is missing

There is no price list and no named customer in text. The README's link to the miner guide returns a 404; the guide and the scoring note with its worked table exist, in folders the README never points to. The site's Documentation link goes to the older score-vision repository, not the one the chain identity names. The chain contact hello@wearescore.com appears nowhere on the site, the discord field is empty, and there is no status page or support path beyond a contact form and GitHub issues.

Go deeper

• The front page and demos
• Current TurboVision README
• Scoring note with worked table

Metaphor: a press-box statistician. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.