Score
SN44Turns any camera into a smart camera using AI-powered computer vision
Score turns raw camera feeds into structured analytics in real time, and the network has pushed beyond a sports-only focus into retail security, venue ops, and live broadcast tooling.
// Vision AI on every camera, decentralized.
Score (subnet 44) is a Bittensor subnet for computer vision. The network takes video, runs models contributed by miners, and outputs structured insights like player positions, ball trajectories, or detected objects.
The simple version: It is like having a stadium full of analysts watching every camera you point at it, except the analysts are AI models competing for the cleanest output.
Centralized equivalent: Closest analogues are Stats Perform, Hawk-Eye, and Genius Sports for sports vision, plus enterprise platforms like Verkada for venue and retail vision.
How it works:
- Miners submit computer vision models that solve specific tasks ("Elements") defined in an on-chain manifest, from player tracking to object detection.
- Validators score those models against ground truth, either real datasets or pseudo-ground-truth generated locally with SAM3, then submit per-element weights on chain.
- The problem it solves: Specialised computer vision is expensive, brittle, and locked behind a few centralised providers. Sports leagues, retailers, and venue operators pay heavily for tooling that only handles narrow scenarios.
- The opportunity: Every camera is a potential data source. Score's Manifest can grow to cover any vertical where structured video matters, and the team is already shipping into multiple ones.
- The Bittensor advantage: Open competition between miner models forces quality up and price down per task. The same incentive flywheel that pushed LLM subnets toward commodity pricing applies to vision.
- Traction signals: Score's product layer, branded Manako AI, has shipped vision modules covering sports analytics, retail security ("Detect Crime"), and venue analytics ("Detect Beverage"). Public posts reference work with PwC France and Reading FC, plus a recent cricket ball-tracking challenge run on the subnet.
Category: Other (Computer Vision) | Centralized Competitor: Stats Perform, Hawk-Eye, Verkada
Computer vision sat behind a handful of vendors for a decade. Per-camera licensing, per-sport bespoke pipelines, and slow integration cycles meant only the largest leagues, retailers, and venue operators could deploy it at scale. Score is betting that an open, model-versus-model marketplace can collapse those costs and broaden the reach of structured video.
Mechanism:
According to the active TurboVision repository README, Score's design centres on a Manifest, a document that lists Elements. Each Element is a self-contained vision task with its own evaluation window and weight share. Miners register models per Element. The runner pulls the active Manifest each block window, builds ground truth (a real labelled dataset where available, otherwise a pseudo-truth generated locally with SAM3), and scores eligible miner outputs against that target. Validators aggregate scores across the lookback window, select per-Element winners, normalise weights to one, and submit on chain.
This means new verticals can be added by extending the Manifest rather than forking the codebase. The team has used that flexibility publicly: a cricket ball-tracking challenge launched on Bittensor, plus retail security and venue analytics modules under the Manako AI product brand.
The active codebase, score-technologies/turbovision, shows continuous commit activity through May 2026, with five contributors visible on GitHub. The older score-technologies/score-vision repository has not received commits since December 2025, but it is no longer the working codebase. Anyone tracking Score from the on-chain identity field will see the older repo and miss the live work, which is worth flagging for risk monitoring.
The team is publicly named: Max Sebti (Co-founder & CEO), Tim Kalic (Co-founder & CTO), Nigel Grant (Co-founder & CBO), and Dr. Peter Cotton (Co-founder), per the TAO.app subnet about page.
Score's market footprint at snapshot is a 195,418 TAO market cap, 0.03926 TAO price, 61,728 TAO of root depth in the pool, an emission EMA of 8.48%, and 10 active miners. Net TAO flows over the past 7 days are positive at roughly 2,811 TAO, which under the Taoflow model means the subnet is on the receiving end of emissions. Price is up roughly 29% over the last 30 days and around 9% over the past week.
- Concentration: Gini coefficient of 0.53 across the top 100 stake distribution suggests moderately concentrated ownership or stake distribution. Large positions could significantly impact pool dynamics.
- Brand Layer: Score and Manako AI sit on top of subnet 44. Many public partnerships and product launches are framed under the Manako brand, which means subnet performance and product traction are not always visible from on-chain data alone. Track product activity in addition to chain metrics.
- Repo Drift: The on-chain GitHub identity still points to score-vision, while active development is in turbovision. Anyone evaluating Score purely from the on-chain field will see a stale picture. Worth verifying when monitoring.
- Competition: Sports vision is a crowded space with well-capitalised incumbents. Score's edge depends on staying ahead on per-element model quality and broadening the Manifest faster than centralised providers can re-price their offerings.
