Training environments
Containerised, multi-step tasks with automatic graders, built for reinforcement learning. Agents will earn credit for sound method, not just a lucky final answer.
We're building a library of expert-written biology tasks with locked, verified answers, packaged as training environments and private evaluations for the teams building AI scientists.
Every task will be re-solved by a second expert, tested against frontier agents, and cleared by a biosecurity review before release.
Agents can now run whole analyses: read the literature, pull the data, write the code, draw a conclusion. Teaching them to do it well takes tasks that are hard, realistic and checkable, and biology has very few.
What we're building
Each task will be a complete research problem: a question a scientist would actually ask, real data, a sandbox with the standard tools, and an answer that has been independently confirmed.
Containerised, multi-step tasks with automatic graders, built for reinforcement learning. Agents will earn credit for sound method, not just a lucky final answer.
Held-out suites that will never be published or reused for training, refreshed every quarter, so scores show real capability, not memorisation.
A planned annual, neutral ranking of biology agents on a shared subset.
Verification
| Criterion | Typical sources of biology tasks | Cognita Bio |
|---|---|---|
| Task size | Single-answer questions | Study-scale, multi-tool analyses |
| Ground truth | One author's answer | Blind re-solve, experiment-backed where needed |
| Grading | Final answer only | Final answer plus checks on each step |
| Neutrality | Often built by an agent vendor | Independent: we don't build or sell an agent |
| Safety | Varies | Every task cleared by a biosecurity review |
Our first release will focus on two areas:
We're building study-scale tasks across both areas, released in batches as each one passes verification. We won't rush a task to hit a number. Imaging and structural biology will follow.
Design partners
We're working with a small number of design partners: model builders and biology agent platforms who want tasks built around their hardest problems.
Design partners get:
The first batch goes to design partners. General access follows once the first release passes verification.
We're planning Docker-based environments and a hosted evaluation API, and we'll confirm formats with design partners first.
We plan to offer time-limited exclusive tasks. Tasks in the public scorecard set will never be exclusive.
A vetted network of PhD and postdoctoral scientists, with each task reviewed by a second expert in the same field. If you're a scientist interested in writing tasks, email us at hello@cognita.bio.
Every task will go through dual-use screening with an outside biosecurity adviser before release. Tasks that fail won't be built.
Not at first. We're focused on the teams building AI for science.
Join as a design partner and be among the first to run your models on verified biology tasks.
Talk to the team: hello@cognita.bio