Ritivel

Ritivel, which presents its product as TensorTest, builds private evaluations, reinforcement learning environments and training data for AI labs, drawn from real creative work. Its first domain is video editing, covered by a benchmark called TimelineBench. TimelineBench gives agents 56 real editing jobs, each a whole task with raw footage, audio and a brief going in and a finished cut coming out. A run counts as resolved only if it passes hard checks on delivery, content and the brief, plus a quality test calibrated on blind judgments by 43 professional editors. The site reports results for 16 agent setups: 93 percent of runs ended with the agent claiming success, while 14.0 percent resolved. Three offerings are listed. Private evals are available now, where the team runs an unreleased checkpoint or agent on its jobs and returns a failure report that stays private. RL environments package the jobs for a lab's own sandbox with the checks exposed as the reward. Training and eval data is commissioned, with new jobs on commissioned footage, training rights, and professional editors' trajectories, preference pairs and rubrics. The benchmark code is published under the Apache 2.0 license. The company is a Y Combinator W26 startup, and terms are scoped per engagement with no published prices.