Lorraine logo

Models

Clinical AI,
calibrated for SA.

A small catalogue of purpose-built clinical models. Each card describes what the model is for, how it should be used, and the numbers that back up the claim.

/systemlive
region
ZA-1
model.lor1
online
bench
LorBench · SA
strengths
dosing · ddx
api
openai-compatible
/flagshipProduction

Lor-1

The clinical reasoning model for South Africa.

A clinical foundation model calibrated to reason inside South African clinical conventions — SA drug names, the tiered Essential Medicines List, national treatment guidelines, and local referral pathways. Served behind an OpenAI-compatible API.

  • Differential reasoning grounded in SA Standard Treatment Guidelines.
  • Drug lookup aware of EML tier and care-level availability.
  • Structured summarisation and extraction from clinical notes.
  • Explicit abstention when SA guidance is insufficient.
Open model card

/benchmark

Dosing

Strongest lifts

SA drug names, dosing schedules, and formulary conventions.

On the roadmap

What’s next.

Work in progress, published so customers can plan around it. Nothing ships until it clears the same benchmark bar as Lor-1.

In research

Lor-1 Vision

Multimodal extension of Lor-1 for clinical imaging in SA settings — chest X-ray triage, paediatric growth charts, and diagnostic photograph review with SA-grounded interpretation.

In research

Lor-1 Small

A smaller, edge-friendly variant of Lor-1 for low-bandwidth and on-prem deployments. Same SA calibration, a fraction of the infrastructure footprint.

How we publish

Design commitments.

  • 01

    Clinical scope

    Each card states intended use, user assumptions, and the clinical boundaries where a qualified clinician must remain responsible for the decision.

  • 02

    Measured, not asserted

    Every production model reports its lift on LorBench — an internal 203-question SA clinical benchmark — before and after adaptation, by question type.

  • 03

    Honest limitations

    We publish known failure modes, recall ceilings, and category weaknesses alongside the headline numbers. Silence about limitations is not a marketing feature.

  • 04

    Controlled disclosure

    Deployment-specific evidence packs, evaluation reports, and architecture notes are shared through enterprise review under NDA — not by default.

Public cards intentionally omit training recipes, dataset composition, serving topology, and internal codenames. Enterprise customers get a deeper technical appendix under NDA.

Request enterprise review