Score SN44

Most security cameras are excellent witnesses and rather poor employees.

They record warehouses, shops, factories, stadiums and public spaces for thousands of hours, yet usually contribute little while events are actually unfolding. A camera may capture a worker entering a dangerous area, a vehicle blocking a loading bay or a supermarket shelf remaining empty for half a day. The footage exists. The useful response often arrives later, after someone has found the correct recording and watched it.

Score, Bittensor Subnet 44, is building a decentralised network for computer vision: AI systems that can identify and track what is happening inside images and video.

Its commercial platform, Manako, turns those capabilities into products that businesses can connect to their existing cameras. A warehouse might ask to be alerted when someone enters a restricted zone. A retailer could monitor shelves or unusual behaviour at a self-checkout. A factory could check whether safety procedures are being followed and create a record when they are not.

The relationship between the two matters. Score is the competitive network through which specialist vision models are developed and improved. Manako is the product layer that allows a company to describe a real operational problem and deploy the resulting intelligence without needing to understand the subnet underneath it.

The problem with passive cameras

There are already billions of cameras installed throughout the physical economy. They watch entrances, production lines, roads, storage facilities and retail floors. Most were designed primarily to record evidence rather than interpret events.

That creates a peculiar imbalance. Companies already own the sensors, networking and video feeds required to observe their operations, but much of the information remains trapped inside footage that nobody is watching. A supermarket camera can see that a shelf is empty, but it does not tell the employee responsible for restocking it. A factory camera may record unsafe behaviour, but the footage becomes useful only after an incident has occurred. A security guard can monitor several screens, although human attention becomes a scarce resource long before the number of cameras does.

Computer vision can turn those cameras into active systems. Instead of merely saving video, a model can recognise objects, follow movement and detect events as they happen.

The difficulty is that conventional computer-vision projects can become expensive surprisingly quickly. A company may need to collect and label footage, hire machine-learning specialists, train a custom model, purchase suitable hardware and integrate the output into its internal systems. A request that sounds simple in ordinary language—“tell me when a vehicle blocks this loading bay for more than five minutes”—can become a substantial software project once edge cases, camera angles and deployment constraints are introduced.

Score and Manako are trying to make that process more repeatable. The ambition is to let a business define what matters in normal language, test the resulting system on its own footage and deploy it on cameras it already owns.

Football as a proving ground

Score initially focused on analysing football video.

That may sound like a narrow starting point, but football presents a difficult and useful computer-vision challenge. A model must distinguish players, referees, goalkeepers and the ball; understand their positions on the pitch; and continue tracking the same objects while they move, overlap or temporarily disappear from view. The camera itself is also moving, which removes some of the convenient assumptions available in fixed industrial footage.

Miners on Score received football video and returned standardised information describing what appeared in each frame. Validators then assessed whether the objects had been identified correctly and whether the model maintained a consistent track as each player moved through the sequence.

The validation process does not simply repeat every miner’s entire workload. It samples selected frames and checks several forms of plausibility: the geometry of the football pitch, the positions of detected objects and the accuracy of labels such as player, referee, goalkeeper and ball. Detection accuracy matters, but so does continuity. A model that repeatedly assigns a new identity to the same player may recognise a human figure without actually understanding the movement taking place.

Football was therefore less the final product than a demanding test environment. It gave Score a measurable competition in which miners could improve object detection, tracking and model efficiency before those capabilities were applied to less theatrical surroundings such as warehouses and checkout lanes.

From subnet models to Vision Agents

Manako is the interface through which Score’s computer-vision capabilities become usable by businesses.

A customer begins by describing what a camera should detect. The request might concern a person crossing a boundary, a vehicle remaining stationary, an item disappearing from a shelf or a worker failing to wear required equipment. Manako then selects and combines specialist vision capabilities suited to the task.

The customer can test the resulting Vision Agent on existing footage before connecting it to a live feed. Once deployed, the agent watches for the specified event and triggers an action when it occurs. That action might be a Slack message, an email, a WhatsApp alert, a timestamped record or a call to another system through an API.

This product design is important because the customer is not expected to purchase a raw model. A warehouse manager does not generally want to compare detection benchmarks or investigate which miner produced a particular capability. The manager wants the correct person to receive an alert before someone drives a forklift into the wrong area.

A Vision Agent also does not need to be one enormous general-purpose model. It can combine several smaller capabilities: one model detects a person, another tracks movement, while a separate rule determines whether the person has remained inside a restricted zone for long enough to justify an alert. This modular approach can make the system cheaper and easier to adapt than running a large model over every frame.

Manako says processing can take place locally on the customer’s own device. In that configuration, the continuous camera feed does not need to be uploaded to Manako’s servers. The model observes the footage where it is produced and sends only the resulting event or alert.

For workplaces and public-facing businesses, that distinction is meaningful. Video may reveal employee behaviour, customer movements, production methods and the layout of sensitive facilities. Local processing does not eliminate every privacy concern associated with automated surveillance, but it can reduce the amount of raw footage leaving the organisation.

Why build this on Bittensor?

Computer vision improves through experimentation. Different teams can test model architectures, training data, tracking methods, compression techniques and deployment strategies. The strongest model in a laboratory may still be unsuitable for a shop or warehouse if it requires expensive GPUs beside every camera.

Score uses Bittensor to turn this search process into an open competition.

Miners develop models and methods for interpreting video. Validators issue tasks, assess the resulting outputs and determine which miners are contributing the most useful capabilities. Better performance leads to greater rewards, giving independent developers an economic reason to improve the system.

The network can search across several objectives at once. Accuracy remains important, but commercial computer vision also needs to be fast, compact and inexpensive enough to run at the edge. A model that improves detection by one percentage point while requiring ten times the hardware may be scientifically interesting and commercially inconvenient.

According to a benchmark published by Manako, its current model is approximately 19 megabytes, can run on a standard CPU and achieved average accuracy of around 79 percent in its selected person- and vehicle-detection tests. Those figures should be understood within the benchmark Manako chose rather than treated as universal performance across every camera and environment. Their significance lies in the design priority: the team is explicitly optimising the relationship between accuracy, model size and deployment cost.

Bittensor gives Score a mechanism for continually searching for better trade-offs. Miners can attempt to produce models that are more accurate, smaller, faster or better suited to difficult real-world footage. Validators then decide which improvements are sufficiently useful to reward.

A conventional computer-vision company could run the same research process internally. Score’s thesis is that a competitive network of independent contributors may explore more approaches and improve more continuously than a single fixed team.

That depends on the quality of the incentive mechanism. Miners optimise for what validators measure, not for the broader intentions described on a website. If the benchmark captures commercially useful performance, the competition can produce valuable intelligence. If it rewards shortcuts or narrow optimisation, miners will become extremely good at the wrong task. Bittensor makes the research process open; it does not make the scoring problem disappear.

Making existing cameras more useful

The practical attraction of Manako is that businesses may not need to replace their existing camera infrastructure. The intelligence layer can sit on top of equipment already installed for security or monitoring.

This creates a broad range of possible applications. Industrial facilities could detect unsafe behaviour, spills or blocked access routes. Retailers could monitor stock levels or unusual checkout activity. Logistics companies could track vehicles and loading operations. Sports organisations could extract structured data from ordinary video. Compliance teams could automatically create records when specified events occur.

The economic value will often come from fairly unglamorous interventions. A loading bay remaining blocked for twenty minutes, a machine sitting idle or a worker repeatedly entering the wrong area may each seem minor. Across many sites and thousands of operating hours, small delays and preventable incidents accumulate into meaningful costs.

This is where specialised Vision Agents may be more useful than a general system claiming to understand everything inside a camera feed. A business does not necessarily need an AI with a philosophical grasp of the warehouse. It needs a dependable alert when pallet number three has been left in front of the fire exit.

Smaller companies may benefit most from the reduction in deployment complexity. A large industrial group can hire its own computer-vision team. A regional warehouse operator is unlikely to do so, even when one or two camera-based automations could materially improve safety or efficiency. Manako is trying to turn those isolated use cases into configurable products rather than bespoke research projects.

Score and Manako

The separation between Score and Manako is one of the more important aspects of the project.

Score operates as the intelligence market. It gives miners a reason to develop better vision models and validators a framework through which to compare them. Manako translates those models into a commercial service with an interface, deployment process and integrations that customers can understand.

This is a useful structure for Bittensor more broadly. A subnet may be technically sophisticated and still have no obvious relationship with a paying customer. Businesses do not usually want to interact directly with incentive mechanisms, wallets and validator networks. They want a product that solves a recognisable problem, comes with support and fits into the software they already use.

Manako announced an alliance with PwC France and Maghreb in April 2026. Under the agreement described by the team, PwC can incorporate Manako’s physical-AI infrastructure into advisory and deployment work for enterprise clients in sectors including retail, logistics, manufacturing, energy and infrastructure. Manako identifies Score Subnet 44 as the decentralised network supporting the product.

That relationship could help address a common obstacle for young infrastructure companies: reaching conservative enterprise customers. Large organisations may be interested in computer vision but prefer to purchase it through firms they already trust rather than through an unfamiliar crypto-native company.

Manako was also selected as the overall winner of the Start in Block 2026 competition at Paris Blockchain Week, following a process that reportedly involved more than 1,000 applicants. Awards are evidence of attention, not of durable revenue or product-market fit. Even so, the result suggests that outside evaluators are treating Manako as a serious business product rather than merely a demonstration of subnet technology.

Why this subnet is interesting

Score is interesting because it presents a relatively credible path from a Bittensor competition to a product used in the physical economy.

The subnet began with a narrowly measurable challenge in football video. Miners competed on detection and tracking. Validators evaluated the results. The underlying capabilities are now being packaged for warehouses, shops, factories, sports organisations and other environments where existing cameras capture more information than people can reasonably inspect.

Manako provides the crucial bridge. It turns a model capability into something closer to an operational instruction: watch this area, recognise this event and notify this person when it occurs.

That bridge is where many Bittensor projects will either become commercially useful or remain technically interesting. Incentivising model improvement is only the first half of the system. Someone must still package the output, understand the customer, handle deployment and make the economics work outside the subnet.

Score is also a useful example of physical AI on Bittensor. The network is not primarily generating text, images or answers for people sitting behind laptops. It is trying to help machines interpret events taking place in factories, shops, stadiums and logistics facilities.

The strongest evidence will eventually come from deployment rather than benchmarks or awards: businesses using the agents, measurable improvements in operations and revenue flowing back from customers who find the product valuable.

For now, the central idea is convincing. Companies have already installed the eyes. Score and Manako are trying to provide the intelligence that tells those eyes what is worth noticing.

More info: https://www.manako.ai/