Skip to content
Decidendi.

RLHF · Evals · Red-teaming Legal domain

Expert legal-reasoning data for frontier AI

Legal reasoning is where models quietly fail.

Decidendi is a vetted network of Commonwealth- and African-jurisdiction lawyers who rank, correct, and red-team AI legal reasoning: the expert coverage your current vendors can't source, delivered as managed, quality-controlled capacity.

  • Ghana
  • Nigeria
  • Kenya
  • Commonwealth

01  Specimen: annotated model output

Criminal procedure · Common law

Model output

The defendant could not prove he was at home when the burglary occurred, so the court should convict him of the offence.

Unreviewed · Ranked 2 of 2

Expert note

Reverses the burden of proof. The prosecution must prove guilt beyond reasonable doubt; the accused proves nothing. A model that ships this reasoning is a liability in any common-law jurisdiction. That is exactly the failure our reviewers catch.

Category: Burden · Reviewer verified

02  Position

i

The gap

Frontier models ship worldwide, but legal-reasoning quality outside US and UK law is thin. Crowd workers can't evaluate it. You need practising lawyers, who are hard to source at scale and nearly absent for African and Commonwealth jurisdictions.

ii

The edge

Decidendi is built from inside the talent pipeline these jurisdictions produce, with direct access to the practising lawyers who qualify from it.

iii

Why now

Expert, high-consequence domain data is the scarce input. Jurisdiction-specific legal correctness is a rising liability, and for now an open lane.

03  Capabilities

Preference ranking

Two model answers, ranked on legal soundness, with the reasoning written out, not just a label.

Rubric grading

Single answers scored against your rubric, calibrated to hold consistently across reviewers.

Red-teaming

Probes that surface hallucinated holdings, misapplied tests, and invented authority before your users find them.

Ideal responses

Model-quality answers drafted by lawyers, to standard, for supervised fine-tuning.

You define the tasks and own the data. We supply the vetted experts, the managed QA, and quality that holds at volume.

04  Quality control

We report our agreement numbers.

Agreement
Inter-annotator agreement reported every batch (Krippendorff's α).
Gold
Gold items seeded invisibly at 5% of live volume.
Dual review
10% dual review with documented adjudication.
Failure report
Delivery includes a failure report by category: authority, jurisdiction, burden, application.

05  Contact

Clinton Tano · Founder