Thesis
Every company’s best judgment should run as a model.
“The first principle is that you must not fool yourself—and you are the easiest person to fool.”
Every company runs on judgment calls. An analyst decides whether an alert is a real attack. A customs broker picks the tariff code for a shipment. A compliance officer clears or stops a payment. The best of that judgment lives in a few people’s heads and in the record of what they decided. We think it should run as an expert model: a small AI model trained on a company’s own data and the outcomes of real cases, which runs on the company’s own servers and which its software calls whenever that judgment is needed.
Frontier models miss these calls for a plain reason: the right answers were never public. No model read a firm’s case files or saw how its analysts ruled. General decision models like TypeSafe’s Jev are no exception: they do well on general knowledge but fall behind in medicine, finance and law. Train on a firm’s own records and the order flips: a model trained on financial tasks labeled by Bridgewater’s investors scored 84.7% against 78.2% for the best frontier model. It costs less, too. A frontier model meets every case cold: it reads the whole thing and reasons its way to an answer, and the company pays for every token it thinks. An expert model has learned the work, so it answers in one pass.
Building one takes a model research team, and most companies don’t have one. Thomson Reuters spent about $40M and two years on its own legal model. Datadog bought a model-training startup. Everyone else buys: 76% of enterprise AI use cases are now bought, not built, and what they buy was never trained on their work.
Why now? Two things changed in 2026. First, AI now builds AI. On PostTrainBench, the best AI agent that post-trains a model scores 44.6%, against 48.4% for the model makers’ own post-trained versions. In March it was 23.2% against 51.1%. Much of a model research team’s work can now be done by agents, at the pace of compute rather than of hiring.
Second, software needs models built for it. Most model calls now come from software: in GitHub Copilot’s coding agent, 87% are made by the agent, not the person. A program needs one of a few set answers it can act on, which is what decision models like TypeSafe’s Jev return, and Jev reached 35% of Vercel’s paid AI Gateway teams within two weeks of launch. When software asks, the answer should come from a model that knows the company’s work.
Pensieve is an experiment platform for building those models. Its autonomous research agents do a model research team’s work: they plan training experiments on a company’s data, run each one on the machines it needs, and keep only what wins on the company’s own test.
We build it around experiments because of Feynman’s principle. A model trained on a company’s past can fool everyone, its makers first: it can learn a shortcut in the data, or score well on a test that doesn’t match the real work. So every change to how a model is trained is tested against what came before, and an expert model ships only when it beats the best frontier model on the company’s own test, a test the company’s experts set. It then runs beside the live system, and nothing it says reaches users until it has proved itself on real outcomes. After launch it keeps learning on the job from real outcomes: agents catch drift and retrain it, and each new frontier model is scored on the same test. A new version ships only if it beats the one before.
People stay in charge of the job and of the test. What changes is how much of their time a model takes to build and keep sharp. If we are right, each new expert model should take fewer human hours than the last, and we will publish that number whichever way it goes. Our ten-year goal goes further: models that make original discoveries, each one checked by experiment.
Work with us
Start with a paid test.
Bring us one judgment call your experts make every day. We test an expert model on your own data, scored on your own metric against the best frontier model.
Review build — contact address not yet connected.