Research
Machine learning, measured on legal work.
Bernima’s research is machine learning for legal work: training specialized models on the units legal work is made of, evaluating them on legal tasks rather than on fluency, and engineering how they run inside the infrastructure a firm controls.
The premise
Fit the model to the task, then prove it on the task.
Most legal AI is evaluated on fluency. Legal work is mostly structured: extracting a clause, coding a document, dating an event, checking a citation, clearing a conflict. Only a fraction of it is open-ended reasoning.
So the research starts from the task. Fit a model to each kind of work, then measure it on that work, against a known answer.
The result is a stack of specialized models rather than one general one: smart models for the structured work, and a frontier model only where the task earns one and the firm’s policy permits it.
| Task | Model |
|---|---|
| Finding clause boundaries | Legal language model |
| Checking citation form | Legal language model |
| Coding a document | Specialized task model |
| Resolving entities | Specialized task model |
| Calculating a date | Specialized task model |
| Open-ended analysis | Frontier model, by policy |
- Runs on the customer’s hardware
- Only where policy permits, de-identified first
Research areas
Five areas of research.
Where Bernima’s machine learning research is concentrated, from the models themselves to the architecture they run in.
- 01
Specialized models for legal tasks
Training and evaluating compact models on the units legal work is made of (clause boundaries, citation form, date precision, custodian and privilege state) rather than on general instruction following.
- 02
Evaluation that survives a partner
Most legal AI benchmarks measure fluency. We measure whether a system refuses correctly, cites something that exists, and carries uncertainty forward instead of resolving it silently.
- 03
Isolation and inference architecture
How per-matter isolation is enforced beneath the application, what it costs at the runtime level, and how to demonstrate to a reviewer that two matters cannot reach each other.
- 04
Provenance and reproducibility
Signing, append-only records, and the harder problem beneath them: reproducing an output months later, when the model that produced it would otherwise have been retired.
- 05
Inside learning systems
What training builds inside a model, measured from the inside rather than inferred from its outputs alone: the science beneath models that behave reliably by construction.
How we measure
The discipline: measured on the task.
The standards Bernima holds its models to, so that what a model is trusted with matches what it has been shown to do.
- 01
Measured against a known answer
Each model is evaluated on the task it is built for, against a known answer, rather than on general fluency.
- 02
Refusal, citation, uncertainty
We test whether a system refuses correctly, whether what it cites exists, and whether it carries uncertainty forward instead of resolving it silently.
- 03
Compared task by task
Our models are compared with frontier models on each task. Routing is per task precisely because no single model wins everywhere.
- 04
Evaluated on your matters
During an evaluation, the same measurements are computed on the customer’s own matters, inside the customer’s own deployment.
- 05
Reproducible by construction
Models are version-pinned and retained, so an output can be re-run and explained long after it was produced.
- 06
No client data in training
Nothing from a customer matter enters training, fine-tuning or evaluation. Research corpora are sourced separately, with recorded provenance.
From research to products