TensorTest (Ritivel Labs, Inc.)

AI labs building agents for creative work lack rigorous ways to measure whether those agents actually finish real jobs. On TimelineBench, a benchmark of 56 real video editing jobs run across 16 agent setups, 93% of runs ended with the agent claiming success, but only 14.0% actually resolved when checked against hard delivery, content and brief criteria plus a quality test calibrated on blind judgments from 43 professional editors. This gap between claimed success and verified success means labs cannot trust self-reported completion and need an external, standardized way to evaluate agents on whole creative jobs rather than narrow subtasks.

TensorTest offers private evals where it runs a lab's unreleased checkpoint or agent on its editing jobs and sends back a failure report, with results kept private. It also packages its jobs as RL environments for a lab's own sandbox, exposing its hard checks as the reward signal, and commissions new training and eval data on commissioned footage, including training rights, professional editors' trajectories, preference pairs and rubrics. The current focus is video editing, measured through TimelineBench: each job includes raw footage, audio and a brief in, with a finished cut expected out, and a run only counts as resolved after passing delivery, content and brief checks plus the calibrated quality test.