Score

The Score Predict README describes a football prediction market, but the repository’s earliest August 8, 2023 commit belongs to the inherited Bittensor Subnet Template. That generic example rewarded numerical responses, so its date does not establish when Score’s football product began. In the later football design, miners had twelve seconds to choose home win, away win or draw. Validators waited for the completed match, compared predictions with the result and set weights for accurate miners.

A fan app pooled human picks into one miner and promised to recycle its TAO rewards into points and weekly prizes. That consumer promotion belonged to the app; it did not give later SN44 alpha holders a claim on prizes or company revenue.

The next repository shifted the reward from a final score to game-state recognition. Miners had to locate and track players, the ball, referees and goalkeepers across football video; validators sampled frames and checked geometry and semantic labels.

That output could support analytics from the footage itself, but the cost changed too: miners needed suitable GPUs, while validators depended on selected frames and external model-assisted checks. The token stayed on subnet 44 even though the purchased work was no longer a forecast.

TurboVision broadened the pitch from football tracking to live video and imagery. Instead of hard-coding one contest, its task list defines separate elements with their own schedule, look-back window, reference method and portion of the final weight.

This is how one subnet can seek a new market without replacing its alpha. It is also a concentration of editorial power: whoever publishes that task list decides which models are useful this week and can make equipment optimized for yesterday’s problem less valuable.

Eligible miners come from the on-chain registry, but their outputs are not rewarded by an autonomous chain vote. A runner creates or fetches reference answers, saves evaluations in R2, and a central validation process sends final weights through a signer; a blacklist can remove hotkeys.

Those controls can make incident response practical, but they place task selection, data availability and weight submission in a small operational path. Alpha holders receive exposure to the subnet pool and its emissions, not ownership of the task list, footage, models, stored results or customer agreements.

At review, Bittensor marked SN44 health 31/100 and “abandoned,” and its metagraph summary showed zero miners. The same page showed thousands of holders, stake, trading, eight occupied validator slots and ongoing emissions.

TurboVision also merged a security change on August 26, 2026. Code maintenance and a live alpha market do not prove that customers are buying the video output; the abandoned label does not by itself prove a permanent shutdown. The defensible conclusion is that public operational evidence conflicts.