The Station Is the Edge: Offline Clinical Decision Support for Remote Communities
📋 Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-08-30). "The Station Is the Edge: Offline Clinical Decision Support for Remote Communities". SOMAsoft Research. Available at https://somasoft.ai/papers/the-station-is-the-edge. Licensed under SAGL-1.0.
The Station Is the Edge
Clinical decision support for remote communities, designed for the network that exists
Evidence gate: truthiness 1.000 · 13/13 load-bearing claims grounded · 21 sources
What this is, and emphatically is not
This is a design note published so that it can be found, criticised, or ignored. It is not an offer, a pitch, or an approach to any community, health authority or government. No one asked for it.
The author is not Indigenous, is not Canadian, and has no standing to propose anything to a First Nation. That is not modesty; it is the single most important design constraint in the document, and it shapes the architecture below rather than sitting in front of it as a disclaimer.
The system described has zero deployed users and no verified external clinical validation. Anyone reading this as evidence of a working product has misread it.
01 — The asymmetry that defines the problem
Almost every clinical AI system now being built assumes a network. In the places where clinical decision support would matter most, that assumption is false — and it is false in a way the national statistics actively conceal.
| Coverage | |
|---|---|
| Canadian homes with high-speed access (Dec 2023) | ~96% |
| Households in Indigenous communities meeting 50 down / 10 up | ~24% |
Every Nunavut community is served by geostationary satellite, which does not meet that standard at all. A system designed against the 96% is designed for somebody else.
Canada aims for 98% of households connected by end-2026 and 100% by 2030. The national average is real. So is the distribution underneath it.
Older satellite and microwave infrastructure produces high latency, slow speeds and frequent disruption. When the Telesat Anik F2 satellite failed, communities across northern Quebec and Nunavut lost service together. Low-earth-orbit capacity is arriving — Canada announced over $13 million in May 2026 for Telesat Lightspeed capacity reaching up to 869 Indigenous households in the Northwest Territories — and it genuinely helps. It does not make the link something a clinician can assume at 3 a.m.
Who is actually holding the decision
Nursing stations run by Indigenous Services Canada deliver primary care in remote and isolated First Nations communities, staffed by two to four community health nurses. Those nurses work an expanded scope: primary care, public health, health promotion, emergency care, and after-hours on-call. When no physician or nurse practitioner is on site, they consult one by telephone, and regional physicians support them by telephone and telehealth.
So the decision-maker is a nurse, often alone, often at night, on a line that may or may not be up, several hundred kilometres from the nearest colleague. That is the edge this system has to serve, and it is not a degraded version of a hospital. It is a different job.
What being wrong costs
Medical evacuation in northern Canada runs $1,500 to $22,000 per transport, and roughly $25,000 to $30,000 for an air ambulance from Yellowknife to Edmonton. Averaged across population it is $218 per person per year in the Northwest Territories and $700 in Nunavut.
This cuts in both directions, and the second direction is the one that matters. An evacuation that should have happened and did not is a catastrophe. An evacuation that need not have happened takes a person hundreds of kilometres from their family, their language, and their community, and that harm does not appear on any budget line. A tool that optimises the cost number and ignores the second harm is not a health tool.
02 — What "offline" has to mean
Offline-capable is usually a feature. Here it is the whole architecture, and it has to be literal: the system must be fully functional with the network unplugged — not degraded-but-usable, not cached, not sync-when-available for anything a clinician relies on in the moment.
That single requirement eliminates almost every clinical AI product currently sold, and it is not a philosophical preference. It follows from the table above.
| Model | Score | s/task | Licence |
|---|---|---|---|
| qwen2.5:7b-instruct | 100.0% | 1.1 | Apache 2.0 |
| gemma3:4b | 92.9% | 2.0 | Custom vendor licence |
| qwen3:8b (reasoning off) | 92.9% | 1.1 | Apache 2.0 |
| qwen2.5-coder:7b | 85.7% | 1.1 | Apache 2.0 |
Measured on an 8 GB consumer GPU, three runs per model, August 2026.
Two findings matter for a station. First, the licence and the capability point at the same model: Apache 2.0 imposes no field-of-use restriction and follows derivatives cleanly, while a custom vendor licence carries use restrictions that travel with any locally adapted weights. For a community that intends to own its own model, that difference outweighs a seven-point score gap.
Second, this hardware is unremarkable. An existing 8 GB GPU costs nothing to reuse; a used RTX 4090 with 24 GB runs about $900 to $1,100; adapting a model to local vocabulary, forms and referral pathways costs about $3 to $5 per training run. A vision-language model at 4B fits in roughly 6 GB, which puts wound photography and document reading inside the same box.
The unglamorous conclusion. The barrier to sovereign clinical capability at the edge is not cost, hardware, or model quality. All three are solved. The barrier is that nobody has specified the deployment, and the people who could specify it have not been asked.
03 — Data governance as architecture, not compliance
The First Nations principles of OCAP — ownership, control, access, possession — were articulated in 1998 and are held by the First Nations Information Governance Centre. They assert collective First Nations ownership of information, control over how it is collected and used, access to it, and physical possession of it. The CARE Principles for Indigenous Data Governance extend the same right to govern collection, ownership and application.
Most software treats these as a compliance layer bolted to a cloud product: agreements, assurances, data-processing addenda. That approach concedes the fourth principle immediately. Possession cannot be delegated by contract. Either the data is physically in the community or it is not.
An offline box is therefore not merely compatible with OCAP — it is the only common architecture in which the fourth principle is literally true rather than promised. The alignment is not a marketing claim; it is a consequence of the network constraint in Section 01. The design that works technically and the design that respects the governance regime turn out to be the same design.
The part outsiders get wrong. OCAP is not a checklist to satisfy. It is a statement about who decides. A system designed by outsiders that is technically offline but whose ontology, alert thresholds, escalation rules and language were all chosen elsewhere still concedes control and ownership while satisfying possession.
The design must be refusable in parts. A community that wants the drug-interaction checking and not the triage suggestions, or the translation and none of the clinical content, must be able to have exactly that without the system degrading into uselessness. If refusal breaks it, it was never offered.
04 — The design
Five layers. The ordering is a dependency: each presupposes the one above, and the topmost is the one that cannot be automated.
L1 — The nurse decides. Always. Nothing in this system issues a determination. It prepares one for a named clinician who can reach a different conclusion and has the time to do it. This is also the line that keeps the tool outside medical-device regulation — and the moment it is crossed, the regulatory character changes entirely.
L2 — Retrieval with provenance, never generation. Every clinical assertion carries its source and the clinician can open it. The system's job is to find the right page of the right protocol in eight seconds at 3 a.m., not to compose novel medical prose. Composition is where fabrication lives.
L3 — Calibrated refusal. "I don't have grounded information on that" must be a first-class, frequent output, and outside the local corpus it must be the only output. A tool that always answers is worse than no tool, because it spends the clinician's trust on the cases where it is guessing.
L4 — The local corpus, locally governed. Formularies, protocols, referral criteria, evacuation thresholds — held in the community, editable by it, versioned. This is where local clinical judgement is encoded, and it is the layer an outsider must not write.
L5 — The ledger. Every query, source and decision appended to a tamper-evident local log, inspectable by the community. Not telemetry — none of it leaves. It exists so the community can audit the tool, and so a nurse can show why they did what they did.
Two of these already exist and are measured. The governance gate hash-pins an immutable policy, resolves verdicts most-restrictive-wins, writes a hash-chained attestation and detects a tampered ledger record — 15 of 15 tests passing. A disclosure gate that distinguishes the speaker as object of harm from the speaker as agent passes 22 probes with zero missed alarms and zero false alarms, and the full 29-probe battery against the live engine reports no critical flags.
The rest is unbuilt.
05 — What would falsify this
The argument fails, and should be abandoned, if any of the following is true:
- Connectivity closes faster than the design. If low-earth-orbit capacity reaches these communities reliably enough that a clinician can assume the link at 3 a.m., the offline requirement loses most of its force and a well-governed cloud service becomes competitive. Canada's target is 100% by 2030.
- Retrieval-with-refusal is not clinically useful. If nurses find that a system which frequently says "I don't know" costs more time than it saves, the design is wrong at the root. This is measurable and has not been measured.
- The corpus problem dominates. If assembling and maintaining a local clinical corpus is harder than the software, then this is a librarianship project wearing an AI costume, and should be funded and staffed as one.
- Communities do not want it. The most likely and most legitimate failure mode. Nothing here entitles anyone to a hearing.
06 — Limits
- No community has been consulted. Every design choice is therefore provisional, and the ones about clinical content are the least trustworthy.
- The author is not a clinician, and no clinician has reviewed this. The five-layer design has never met a patient.
- Zero deployed users, no verified external clinical validation. An independent capability probe of the underlying system found analogy, theory of mind, counterfactual reasoning and abstraction absent — which is precisely why the design confines it to retrieval with provenance and gives it no judgement to exercise.
- Jurisdiction is unexamined. Canadian provincial health-privacy law, cross-border considerations and medical-device regulation each require qualified legal advice that has not been obtained.
- Benchmarks measure coding, not medicine. The model scores in Section 02 come from single-function programming tasks. They establish that a small local model runs fast and follows specifications — not that it is safe near a patient.
- The equity evidence is borrowed. Documented harms elsewhere — the Optum algorithm's racial bias affecting roughly 200 million Americans, medication error rates of 17.7% versus 9.6% for children in limited-English-proficiency families, 96.7% of FDA-cleared AI devices cleared by substantial equivalence rather than outcome proof — motivate caution here. They are not evidence about this design.
Standing conclusion. The technically correct architecture for clinical decision support in a remote community and the architecture that respects Indigenous data governance are the same architecture: everything local, nothing leaving, the community holding possession because possession was never given away. That convergence is the only claim in this paper I would defend strongly. Everything downstream of it needs people who actually live there, and they have not been asked.
Sources
- Indigenous Services Canada — nursing in First Nations communities; Northern Health Region nursing stations.
- Government of Nunavut medical travel; CBC North; Patient transportation in Canada's northern territories (PubMed 31128577).
- Innovation, Science and Economic Development Canada — connectivity strategy and May 2026 Northwestel/Telesat Lightspeed announcement.
- HillNotes, Broadband Internet in Indigenous Communities; CRTC basic service objective; IRPP.
- First Nations Information Governance Centre — the First Nations principles of OCAP®.
- Global Indigenous Data Alliance — CARE Principles for Indigenous Data Governance.
- Local model benchmarks, three runs per model, August 2026; model licences via
ollama show --license. - AURI project — governance gate (15/15), disclosure gate (22 probes, 0/0), 29-probe review battery; AGI Capability Probe, 24 July 2026.
- AURIV health equity and AI fairness report.
Method paper. Not a proposal to any community, health authority or government. Not clinical, legal or regulatory advice.