Where we use AI, and where we refuse to.

What models do well

Frontier language models are better at answering medical questions than we are, and we are not going to pretend otherwise — Nature Medicine published exactly that result in June 2026, finding general-purpose models outperformed dedicated clinical AI products on medical benchmarks and were preferred by clinicians. Conceding a true thing you cannot win is what buys you the right to the next paragraph.

Vishwanath K, et al. General-purpose large language models outperform specialized clinical AI tools on medical benchmarks. Nature Medicine. 2026. doi:10.1038/s41591-026-04431-5.

What ARKA produces

ARKA does not answer questions. It produces a decision and a transaction — a rating, a protocol, an authorisation packet, a scanner code — and it produces the same one every time. A payer cannot contract against an answer that changes.

Permitted vs forbidden

The table below is generated from docs/arka-ire/firewall.json (v1.0.0) — the same file scripts/lint-ire-firewall.ts and the firewall tests enforce.

Enforcement

This is not a policy. It is four mechanisms:

1. an import-boundary test that fails the build if the model proposer is imported outside the extraction module.
2. a decision-surface test that feeds an all-model-derived indication into every emitter and asserts none of it survives.
3. a runtime assertion on every decision surface.
4. `npm run lint:ire` in CI.

The table above is generated from the same file those four mechanisms read. The claim and the build artefact cannot diverge.

Grounding rule

A model may propose a coded element with a verbatim quote. We locate that quote in the source text ourselves. If it is not there character-for-character, the element is discarded — not flagged, not reviewed. Discarded. There is no fuzzy match, because fuzzy matching is how hallucinations get through.

Build artefact

An engineer evaluating you will check. Finding that the JSON below is the same file CI and the import-boundary tests enforce is worth more than any paragraph we could write.

{
  "version": "1.0.0",
  "permitted": [
    "extract candidate clinical elements from free-text notes",
    "draft the human-readable indication paragraph",
    "draft the medical-necessity narrative",
    "draft the patient-facing shared-decision script"
  ],
  "forbidden": [
    "the appropriateness rating",
    "the approve/deny decision",
    "the protocol selection",
    "any safety gate",
    "anything written to an audit trail or a payer contract"
  ],
  "enforcement": [
    "import-boundary test",
    "decision-surface test",
    "runtime assertNoModelDerivedInput",
    "npm run lint:ire"
  ]
}

Related: Indication reconstruction, Evidence, Trust.