// we test agents, train open models with RLHF and RLVR, and build the verifiers that keep both honest.
/ what we do
Most work starts with evaluation: you cannot improve what you cannot measure honestly.
/ who we work with
/ the learning loop
A model’s progress depends on the tasks it attempts, the feedback it receives and the way improvement is measured. We are interested in the connections between those pieces, and in the engineering that makes them practical to study.
Define a problem and the conditions for success.
Use human judgment or executable verification where each fits the task.
Study how the chosen feedback changes model behavior.
Test beyond the training signal and examine the failures that remain.
/ focus
/01
Preference data, reward signals and post-training approaches that connect model behavior to informed human judgment.
[ read the approach -> ] for human feedbackexplanatory Pairwise preferences, and the disagreement between raters, become the signal.
/02
Tasks and execution environments that let us check outcomes, inspect failures and study the limits of a verifier.
[ read the approach -> ] for verifiable rewardsexplanatory A verifier is a component to question, not an oracle.
/03
Held-out tasks, counterexamples and reproducible comparisons that help distinguish useful improvement from a misleading score.
[ read the approach -> ] for independent evaluationexplanatory Held-out tasks answer a different question than the training reward.
/04
Inference, rollout execution and controlled infrastructure that make experiments easier to run and understand.
[ read the approach -> ] for systems behind learningexplanatory Rollouts run side by side, each inside explicit resource bounds.
/ try it
A small, honest demonstration of why we test beyond the training signal. Pick a candidate, run the weak verifier, then run the held-out tests and compare.
task Return the largest number in a list.
function largest(xs) { const s = [...xs].sort() return s[s.length - 1]}[2, 5, 9]. Pass it and the reward is 1.0.toy example Runs in your browser. Every result is computed from the code shown, not from a model.
/ how it works
You describe the task, the model or agent, and what is not working. No confidential data needed yet.
We agree what success means, which data and access are involved, and what you receive at the end.
A small first slice, usually an evaluation, so both sides see real results before going further.
The main work, with regular check-ins and the evaluation running alongside it.
Code, data, evaluation suites and a report you can rerun without us.
/ pricing
Scoped per project. Agreed up front. Most teams start with a fixed-scope evaluation sprint.
/ founders hub
Shipping a product on a model or an agent, usually without a research team? These are the moments we see most.
/ how we work
We want the assumptions behind an experiment to be visible: the task, model, feedback, evaluation conditions and limitations. Our engineering interests support that goal through explicit boundaries, observable execution and repeatable workflows.
/ open questions
A program can pass inadequate tests, and a query can match one example database without being semantically correct. We want to find those gaps before a training run learns to exploit them.
A rising reward curve says the optimized signal went up. Comparing it with held-out task success, under stated conditions, is how we want to judge whether training actually helped.
Disagreement can come from an unclear rubric, a genuinely hard case or a real difference in judgment. Each means something different for how the feedback is used in post-training.
Some tasks admit executable checks, such as running SQL against a database or tests against code. Others depend on informed judgment. A training workflow should keep that difference visible instead of treating subjective ratings as objective correctness.
The task, base model, feedback, evaluation conditions and limitations, written down where a reader can check them. Results without that context are hard to assess.
/ faq
/ contact
Working on model post-training, verification or evaluation? We are interested in conversations grounded in real tasks, practical constraints and questions that can be investigated.
Please share a short description first. Avoid sending confidential data in an initial enquiry.