Pay Only For Verified Model Lift

This model turns data procurement from a sales promise into a measurable performance contract. Instead of asking a frontier lab to commit budget before seeing any gain, Surge lets the lab run post training on a real dataset, check whether benchmark scores move, and only pay when the lift clears an agreed bar. That is especially powerful in coding and RL data, where a few points on hard evals can justify a large spend.1 2 3 4

• The mechanism is simple. A lab trains on the full dataset, measures before and after performance on a target eval, then pays if the result improves. Surge shows this model alongside concrete benchmark deltas, including GLM 4.7 moving from 37.2% to 44.3% on SWE-Bench Pro after training on its coding dataset.1 4
• This is a sharper wedge than standard labeling vendors use. Scale bundles labeling with broader model infrastructure and has used packaged offerings to hold share, while self serve tools like Labelbox shift recruiting and quality control back onto the customer. Surge is using proof of model lift as the sales motion instead of seats, workflow software, or general managed service claims.2 5 6
• It also fits the economics of this market. A small set of frontier labs drives most spend, vendor switching can be fast, and many providers are trying to climb from labor marketplaces into benchmarks, evals, and RL environments. In that context, train first and pay on improvement is a way to win budget quickly and make proprietary datasets feel less interchangeable.2 3 6 7

The next step is for data vendors to price more work on verified outcome, not hours or headcount. As benchmarks, eval APIs, and RL environments become the way labs decide where to spend, vendors that can repeatedly show lift on hard tasks will pull ahead of firms selling labor without a visible link to model performance.2 4 6 7