For AI labs and model builders

Your situation

You train and ship models that act: coding agents, computer-use agents, tool-using assistants. Every release needs numbers, and the numbers that matter most are about agentic work, where the model does something on a system and then reports what it did. You want those numbers to hold up outside the lab. You also want training data that moves the model rather than filling GPU hours.

Three problems

Public benchmarks are contaminated. Once a test set is public, it finds its way into training data, and a score on it stops measuring the capability it was designed for. ARC Prize keeps evaluation sets that are never published and runs the tests itself for this reason (ARC Prize, 2024). A lab that evaluates only on public sets cannot tell improvement from memorisation. See benchmark contamination.

Self-grading isn’t credible, and the usual judge is weak. When the organisation that built the model also scores it, outsiders discount the result. The standard tool, an LLM judge reading the agent’s transcript, does not rescue it: the best judge measured reached an AUROC of 0.65 on τ²-bench and 0.54 on AppWorld at detecting false success (Advani, 2026). And frontier agents are now measured for their propensity to overclaim what they did (OverclaimBench, 2026), so the behaviour a judge has to catch is a documented property of current models, not an edge case.

Most RL data teaches nothing. Under group-based methods such as GRPO, a task that all samples pass or all samples fail gives zero advantage and no gradient. In one 78K-problem maths corpus, about 88% of problems fell into that bucket, and calibrated synthesis in the same study gained 8.6 points against 5.8 from the best filtered subset of real data (MagicBeanStalk, 2026). Environments that are not calibrated to the model’s current ability waste most of the compute spent on them. See the zero-advantage problem.

What Inoku offers

• Private evaluations. Agentic tasks on real systems, never published, run by an independent party. False “done” and false “can’t” reported separately, by task family, with confidence intervals. No model weights needed.
• Custom task environments. Task suites with a ground truth the agent cannot reach, held-out splits, tasks that cannot be completed, and difficulty calibrated to the model you are training. Exclusive or shared.
• Inoku Verified™. A public mark backed by a score, under rules we will publish. Every major model will be ranked on the free public leaderboard whether or not its maker pays, and payment never changes a score.

Labs are already the main buyers of RL environments, with contracts reported at six to seven figures a quarter (Sapphire Ventures, 2026). What is scarce is not environments. It is environments whose ground truth is independent of the agent, and results an outsider will believe.

Status

Inoku Verified™ is in development. Private evaluations and custom environments go to design partners first, at no cost during the programme. Each page says what is built and what is not.

Ready to talk? Become a design partner, or contact us about an evaluation.