Score (SN44) Ties Vision Miner Rewards to Sub-100ms Inference Performance

Score (SN44) shipped a speed update on its incentive layer, and vision models now earn rewards only when they combine accuracy with sub-100ms inference latency. The change enforces the subnet’s mission of edge-deployable AI through a dedicated referee machine that tests every leaderboard model and near-qualifying entry once per hour with 10 random challenges.

Miners can fail in two specific ways: the model is too slow, or the model gives different answers than the version that qualified in the first place. Rewards pause on the failed model until the miner submits a fix, though nothing else in their portfolio is affected.

The Change and Why It Matters

Score’s mission has always been vision AI that runs where it is needed rather than where centralized cloud compute lives.

Real-world deployment on cameras, drones, phones, and edge devices requires models that are fast and light in addition to being accurate, and the update makes that operational at the incentive layer.

1. The new reward condition. Vision models must combine accuracy with inference latency under 100ms to earn rewards.

2. The alignment with the mission. Edge-deployable AI needs to run on-device, on-site, and on the most efficient compute available, which requires fast and light models rather than only accurate ones.

3. The change sits at the incentive layer. Miners are rewarded only when the model meets both criteria, not one.

How the Referee Runs

The change is enforced through a dedicated machine that acts as an independent test authority sitting alongside the subnet.

1. 10 random challenges per model per round. The referee throws 10 challenges at each model in each testing cycle.

2. Every leaderboard model tested every hour. No exceptions.

3. Near-qualifying models tested too. Models almost good enough to make the leaderboard are also checked, so miners cannot game the queue.

4. Full transparency. Every test result is visible on the subnet’s conformity page for anyone to audit.

What Failure Looks Like

There are two specific ways a model can fail the referee, and both trigger the same immediate consequence.

1. Too slow. The model exceeds the 100ms latency threshold.

2. Inconsistent answers. The model returns different responses than the version that qualified, which prevents miners from swapping in a weaker model after passing the initial bar.

Rewards pause on the failed model, meaning that earnings stop until the miner submits a corrected version. A failed model is not a ban as a new submission puts the miner immediately back into the rewards pool.

Other submissions remain unaffected and a slip-up on one task does not impact any of the miner’s other models.

Fast Enough to Ship

The Score update turns edge deployability from a design principle into an incentive-enforced requirement. Vision models that hit accuracy but miss latency no longer earn, and models that qualify but silently degrade lose rewards until they fix the drift. The hourly referee makes both conditions live at all times rather than checked at submission and forgotten.

For a subnet building toward one small AI brain per camera running on the device itself, this is the moment the reward structure catches up with the deployment surface. Every test result is public on the conformity page for anyone tracking the leaderboard.

➛ Check Out the Conformity Pages Here and Here.