Lor-1 clinical reasoning.
A clinical foundation model calibrated to reason inside South African clinical conventions — SA drug names, tiered formulary, national treatment guidelines, and local referral pathways — served behind an OpenAI-compatible API.
/lorbench
Areas improved
Stronger where SA clinical practice lives.
- Drug dosing
- Substantial
- Differential management
- Strong
- Clinical scenarios
- Solid
- Referral & triage
- Moderate
Qualitative bands — see Benchmarks for the approach.
- Surface
- Lorraine app + Platform API
- API
- OpenAI-compatible
- Clinical frame
- South Africa
- App data
- af-south-1 (Cape Town)
- Inference
- GPU capacity in the US
- Status
- Production
Model card sections
A model for one healthcare system.
General-purpose language models have become remarkably capable medical reasoners. They can work through differential diagnoses, explain pharmacology, and synthesise complex clinical scenarios. They share one failure mode though: geographic bias. Trained on a corpus dominated by US and European medical literature, they develop statistical priors that skew heavily toward US and European practice patterns. Ask one for first-line therapy for an uncomplicated UTI and it will most likely suggest trimethoprim-sulfamethoxazole or ciprofloxacin — at odds with SA primary-care guidance, which directs clinicians to nitrofurantoin. Ask about a cardiac arrest and it will talk about epinephrine rather than adrenaline.
These aren’t surface errors of tone. In clinical practice they are errors.
Five shapes of drift
- 01
Formulary drift
Reaches for drugs that aren’t on the SA Essential Medicines List, or that aren’t stocked at the level of care the question actually concerns.
- 02
Protocol mismatch
Falls back to international clinical pathways — NICE, UpToDate, CDC — instead of the corresponding local protocol: IMCI, VTP, BANC+, the national TB algorithm.
- 03
Naming conventions
Surfaces US drug names and abbreviations — epinephrine, acetaminophen — in place of their SA equivalents, adrenaline and paracetamol.
- 04
Referral-pathway ignorance
No working concept of PHC → district → regional → tertiary, or of the specific criteria that govern movement between levels of the SA public system.
- 05
False confidence
Produces authoritative-sounding answers exactly where SA guidance is absent or ambiguous, instead of flagging the gap.
These failures compound. A clinician who receives a fluent, confident, internationally-biased answer has no easy way to detect the drift without already knowing the correct local answer — which defeats the purpose of the tool in the first place.
Lor-1 is calibrated at the weights, not the prompt. System prompts and retrieval wrappers can paper over some of this drift, but the model underneath still reaches for the wrong defaults the moment context is thin. Training against a South African clinical corpus shifts the underlying priors so that SA drug names, the tiered Essential Medicines List, the SA Standard Treatment Guidelines, and the PHC → district → regional → tertiary referral ladder are the model’s native reference points — not a layer applied after the fact.
Lor-1 powers Lorraine's clinical decision support for South African clinicians, registrars, and nurses, launched in March 2026. The model is also available directly through the Lorraine Platform API for teams building SA-aware clinical software. In both surfaces, Lor-1 is a reference tool with a qualified user in the loop — not an autonomous decision system.
What Lor-1 learned
- 01SA drug names and dosing conventions — paracetamol (not acetaminophen), adrenaline (not epinephrine), amoxycillin, and SA-standard formulations.
- 02Tiered formulary awareness — which drugs sit at PHC, district, regional, and tertiary levels, with the correct escalation criteria.
- 03SA protocol adherence — VTP, IMCI, BANC+, and the national TB algorithms, with the right thresholds and referral triggers.
- 04Tool-grounded answering — calling structured clinical lookups to cite SA guideline values rather than guessing from parametric memory.
- 05Calibrated uncertainty — explicit abstention when SA guidance is insufficient, instead of drifting into international defaults.
Where frontier models drift
Swipe to scroll
Native reference frame
How we measure SA clinical reasoning.
Evaluating a geographically specialised clinical model needs a benchmark that tests geographic specificity, not just medical knowledge. No such benchmark existed for South African clinical AI, so we built one. The central hypothesis behind it is simple: general-purpose models fail on SA clinical questions not because they lack medical knowledge, but because they default to the wrong geographic frame. LorBench tests that directly, by including questions where the SA answer diverges from the international answer and tracking which one the model picks.
LorBench is our internal benchmark for SA clinical reasoning — a curated set of source-grounded questions covering drug dosing, differential management, clinical scenarios, protocols, and referral triage. Each question has a handful of “elements” — specific clinical facts that a correct answer must surface — and scoring is binary per element, judged by a third-party LLM against the source guideline. We look at two things: how much of the required SA clinical content a model surfaces, and how consistently it uses SA-specific terminology instead of international substitutes.
Geographic bias is worth a second look. When the model gets a question wrong, does it fail toward international defaults, or does it fail randomly? A model that fails internationally — quietly replacing adrenaline with epinephrine, PHC protocols with US pathways — is more dangerous than one that fails randomly. We measure geographic bias directly, not just accuracy.
Rather than publish percentage deltas here, we describe where Lor-1 has moved the needle. Every release shifts the balance a little further toward SA practice across the categories below, with the biggest gains landing where SA guidance diverges most sharply from international defaults.
Areas strengthened
Swipe to scroll
- Legend
- Substantial
- Strong
- Solid
- Moderate
- Modest
Disclosure
Full methodology, per-question breakdowns, and numeric benchmark results are available to enterprise customers under NDA.
What we train on, and why it’s shaped the way it is.
Models inherit the shape of their data. For a clinical model, that’s not a metaphor — it’s the line between a system a clinician can trust in the SA frame and one that sounds confident in the wrong one. The dataset behind Lor-1 is its own artefact: curated, guideline-grounded, and built around a specific thesis about where training signal should go.
A smaller, more expensive corpus that targets the gaps — instead of inflating coverage.
Most clinical fine-tune datasets are produced chunk-first: chunk the source, ask a generator model to write a question–answer pair from each chunk. That produces shape, not improvement — examples that resemble training data without correcting any particular behaviour. We invert the direction. We start from clinical task families where the base model measurably drifts on SA practice, retrieve the source material, draft a guideline-aligned answer, and only keep the example if it represents a useful delta over what the base would already have said.
Composition at a glance
Swipe to scroll
Three ways we concentrate signal
- 01
/reject-sample
Rejection sampling
For every candidate training example we generate several variants, score each against the source guideline on a structured rubric, and only the top-scoring survivors make it into the corpus. The quality floor matters more than the volume ceiling.
- 02
/gap-mine
Adversarial gap mining
Rejection sampling improves examples we can already produce; it does not find gaps we don’t know about. Gap mining probes the boundary between local and international knowledge — surfacing cases where a foreign-trained model would give a different answer than SA guidance — and concentrates training signal there.
- 03
/cross-doc
Cross-document synthesis
Real SA clinical reasoning rarely lives inside a single guideline. An HIV-positive pregnant patient spans VTP, maternal care, and ART. A child with TB and malnutrition spans paediatric TB, IMCI, and nutritional support. We generate cross-source examples explicitly — single-source examples teach recall, cross-source examples teach synthesis.
Primary sources
The training corpus is drawn from the same documents practising South African clinicians already reference. Breadth of medical knowledge is not the objective — the base model has that. Fidelity within the SA clinical frame is.
On scale
We intentionally do not publish corpus size. “How many examples” is a misleading number for a failure-driven dataset — an example that closes a real drift is worth several thousand chunk-generated lookalikes. What we do publish is movement on LorBench release over release — see the Benchmarks section.
How Lor-1 learns South Africa.
The base model already contains broad medical reasoning. The goal of adaptation is not to teach it medicine from scratch, but to shift a narrow set of behaviours — the ones where it reliably gets South African clinical practice wrong — without disturbing the rest. Every design choice below is in service of that constraint.
Training Lor-1 is an exercise in restraint.
The training corpus and the data strategy live in their own section — see Dataset for how the corpus is constructed and how signal is concentrated. This section is about method: the pipeline that takes a curated corpus through staged adaptation, and the rules that decide what gets promoted.
Pipeline
Data pipeline
Authoritative SA guidelines are chunked and turned into task-shaped training examples, with adversarial gap mining and rejection sampling concentrating signal where the base model drifts.
Staged adaptation
A broad grounding phase teaches clinical coverage; a smaller calibration phase reinforces SA preference and explicit uncertainty. Calibration is reinforced last so it resists being overwritten.
LorBench evaluation
Every candidate checkpoint is scored on LorBench across multiple decoding configurations. Element recall measures clinical completeness; SA preference measures geographic bias directly.
Inference
The promoted checkpoint is served behind an OpenAI-compatible endpoint. Reasoning traces can be surfaced to clinicians for transparency or suppressed for downstream integrations. The serving topology behind that endpoint is not published.
Design principles
- 01
Adapt, don’t overwrite
End-to-end retraining of a model at this scale on a corpus of our size caused measurable forgetting of general medical capability in early experiments. Targeted adaptation leaves the base weights intact and encodes SA-specific behaviour as a small, additive update — localised, rollback-able, retargetable.
- 02
Base architecture chosen empirically
We evaluated multiple base architectures before committing. The one selected held calibration best across decoding temperatures and responded most cleanly to SA-specific training signal. This is a measured choice, not a brand preference.
- 03
Failure-driven, not volume-driven
Chunk-first fine-tune datasets produce shape, not improvement. We invert the direction: start from task families where the base measurably drifts on SA practice, draft a guideline-aligned answer, and only keep the example if it closes a specific gap. Smaller and more expensive, but every example earns its place.
- 04
Tool use as first-class behaviour
A large share of SA clinical questions hinge on a single numeric fact — a dose, a threshold, an age cutoff. That belongs in structured lookups, not in parametric memory. Lor-1 is trained to reach for those lookups and cite retrieved values, not wrapped in a retrieval layer around an unchanged base.
- 05
Calibrated abstention
Models naturally favour confident answers. Training Lor-1 to say “the available SA guidance does not clearly support an answer here” requires dedicated examples where that response is correct — and we over-represent them in a late calibration phase so the behaviour resists being overwritten.
- 06
Greedy decoding by default
Temperature-zero sampling gave the strongest element recall on every configuration we tested. Higher temperatures trade deterministic clinical accuracy for diversity we don’t need in a reference tool.
- 07
Benchmark-gated promotion
No candidate checkpoint is promoted without a measured lift on LorBench across multiple decoding configurations. A lower training loss without a corresponding LorBench lift is treated as noise, not progress.
Promotion gate
Nothing ships on loss alone.
Every candidate checkpoint is evaluated against LorBench across multiple decoding configurations before it can be promoted. A lower training loss without a corresponding LorBench lift is treated as noise, not progress.
Disclosure
Specific hyperparameters, dataset composition, and pipeline internals are held privately and shared with qualifying enterprise customers under NDA.
A reference tool with a clinician in the loop.
Lor-1 is designed for clinician-supervised workflows. Deployments must keep a qualified user responsible for every decision the model contributes to. It is appropriate where a clinician can review the output, weigh it against the patient in front of them, and accept or reject it on their own authority. It is not appropriate anywhere that expects autonomous clinical action on the model’s output alone.
The boundaries below are not legalese — they reflect where the model is measurably useful versus where it is measurably unreliable. We would rather be narrow and honest than broad and misleading.
Intended for
- Clinical reasoning support for differential diagnosis, management planning, and guideline-informed next steps.
- Structured summarisation and extraction from clinical notes, handovers, and supporting context.
- Integration into clinician-facing workflows where a qualified user remains responsible for every decision.
Not for
- Sole authority on emergencies, prescribing, referral decisions, or diagnosis.
- Clinical record storage or processing of patient-identifiable data.
- Use outside the SA clinical frame, or as a substitute for clinical judgement.
Should I use Lor-1 for this workflow?
- Fit
A qualified clinician reviews every output before it contributes to a decision.
- Fit
The question sits inside the South African clinical frame — SA drugs, SA guidelines, SA referral ladder.
- Fit
You need differential support, guideline-informed next steps, or structured clinical summarisation.
- Out
The workflow expects autonomous clinical action without a qualified user in the loop.
- Out
Inputs include patient-identifiable data, or outputs must be stored as the clinical record.
- Out
The decision is an emergency, a prescribing commitment, or a referral dispatched without human sign-off.
If the workflow fails any of the last three, it’s out of scope
Known limitations
Swipe to scroll
Every limitation is actively tracked and re-evaluated release to release
A note on time
The model’s knowledge reflects the guideline versions available at training time, and clinical guidelines update periodically. Integrations should surface that caveat where it matters — particularly for dosing, regimen changes, and any guidance known to be revised since the last published training cutoff.
One endpoint, and residency stated in full.
What we publish about serving is the part an integrator has to build against: the API contract, the controls in front of it, and where data actually sits. The topology behind that endpoint is not published here — it changes, and a public card that pins it down would be wrong within a release. Enterprise customers get it under review instead.
A contract your SDK already speaks, and a residency statement that names the country.
What we will commit to in public
- 01
OpenAI-compatible contract
/contract
The public surface mirrors OpenAI’s chat-completions contract. Streaming, tool-use, and structured output behave as expected. Pointing an existing SDK at the Lor-1 endpoint is a base-URL change, not a rewrite.
- 02
Authenticated proxy
/proxy
Requests reach the model through an authenticated proxy holding per-tenant keys, per-key rate limits, and tenant-scoped request and log boundaries. Integrators address one endpoint; what sits behind it is not part of the contract.
- 03
Residency, stated in full
/residency
Application data and backups are hosted in AWS af-south-1 (Cape Town). Model inference currently runs on GPU capacity in the United States. Every provider and the country it operates from is listed in our sub-processor register.
- 04
Reasoning traces are opt-in
/traces
Reasoning can be surfaced to a clinician for transparency or suppressed for a downstream integration. That is a per-request choice, not a property of the deployment.
OpenAI-compatible
A base-URL change. Not a rewrite.
Teams already building on a frontier API can A/B Lor-1 on SA queries without touching their integration. Streaming, tool-use, and chat completions behave as expected.
- chat.completions (streaming + non-streaming)
- Tool / function calling
- Structured output & JSON mode
- Reasoning traces (opt-in)
from openai import OpenAIclient = OpenAI( base_url="https://api.uselorraine.co.za/v1", api_key=LOR_API_KEY,)response = client.chat.completions.create( model="lor-1", messages=[ {"role": "user", "content": "First-line Rx for uncomplicated UTI, SA PHC setting."}, ],)Controls and residency
- Proxy
- Authenticated, per-tenant API keys
- Rate limiting
- Per-tenant, per-key, burst-aware
- Isolation
- Tenant-scoped request & log boundaries
- Observability
- Request logging · latency · recall sampling
- App data & backups
- af-south-1 (Cape Town)
- Model inference
- GPU capacity in the United States
- Contract
- OpenAI chat completions · tool-use · streaming
- Serving topology
- Not published
Request routing is transparent to integrators — one endpoint, same contract
Enterprise deployments
POPIA-aligned data handling, tenant isolation, uptime SLAs, and the serving topology this card leaves out are covered under NDA in enterprise review.
Regulatory status
Lor-1 is not a medical device.
It augments clinician reasoning; it does not replace it. Every integration is responsible for clinician oversight and safety controls appropriate to its clinical context.