The sourced clinical scribe: AI that doesn't fabricate

The most dangerous AI in a clinical setting is not the one that is wrong — it is the one that sounds right but has no basis in the actual case data. A confidently-written diagnostic draft that references vitals the patient never had, or a treatment plan that cites a consultation that didn't happen, is worse than no AI at all.

Vitary's AI layer is built to make this impossible.

The doctrine: display, insert, trace — never fabricate

Vitary's AI doctrine is explicit and enforced in code:

• The AI never becomes a source of clinical truth. It produces drafts, not decisions.
• Every AI-proposed block must be tied to an explicit source already known in the workspace — a consultation, a waiting-room entry, an exam result, a treatment line, a similar case.
• The view displays, inserts, and traces. It does not fabricate raw reasoning.
• Proposed blocks stay as drafts requiring human action. No invisible writes to the medical record.
• If source evidence is insufficient, the AI refuses — it returns blocked rather than guessing.

How it works in practice

Clinical scribe

When a veterinarian opens a consultation, the AI assistant can generate draft blocks — a diagnostic draft, owner-instruction draft, or clinical review. Each block carries mandatory citations to the source data it was built from.

If the consultation context doesn't contain enough information to support a draft, the scribe does not produce one. It returns blocked with an explanation of what is missing. The veterinarian then knows the AI is not hiding behind a confident-sounding hallucination.

Clinical summary

The AI can produce a cited clinical summary — case summary, clinical evaluation, care plan, safety review. The provider contract is strict: it may only rewrite summaries from provided source-backed sections, and it must reuse the citation IDs. It is explicitly forbidden from inventing diagnoses, treatments, values, owners, animals, dates, prognosis, or legal conclusions.

Knowledge review

For medication and treatment safety, the AI performs a source-backed knowledge review using an approved document index. The prompts explicitly forbid general memory, free web search, diagnostics, dosages, treatments, or legal conclusions unsupported by approved sources.

Natural-language analytics

A vet or manager can ask a business question in plain language: "How many dogs were treated this week?" or "Which procedures generate the most unpaid invoices this month?"

The AI does not write SQL or executable code. It parses the question into strict JSON, choosing only from certified query definitions and allowed filters. If the question is ambiguous, it returns clarification-required. If it is outside the supported set, it returns unsupported-query.

An independent AI verifier then re-reads the plan, the results, and the proofs — and confirms or blocks without inventing checks of its own.

The architecture that makes this possible

Two key architectural decisions enable this doctrine:

No AI provider is ever called from the client

Providers are server-owned and rate-limited by what we call "6D contracts." The client/renderer never holds API keys. This means:

• No provider can be abused or bypassed
• Rate limiting and metering are enforced centrally
• The AI contracts (JSON-only responses, mandatory citations, refusal rules) are enforced at the server level, not trusted to the client

Strict JSON-only responses

Every provider response must be valid JSON with mandatory citations and an explicit blocked refusal when source evidence is insufficient. This is not a "please try to cite your sources" prompt — it is a contract enforced by the server adapter.

Why this matters for veterinary medicine

Veterinary clinics are not hospitals. They are small businesses with limited IT resources, often in areas with unreliable internet. They need software that works offline, keeps their data locally, and helps them work faster without replacing their clinical judgment.

Vitary's AI layer is designed to be a constrained, source-backed, human-validated assistant — not an autonomous clinical authority. The veterinarian remains responsible. The AI speeds up the drafting and summarizing, but every word it produces is tied to something real, and every word requires a human click before it enters the record.

That is the difference between an AI assistant and an AI liability.

---

This article describes the AI architecture behind Vitary by Amari, Uptech LLC's offline-first veterinary practice software, currently in beta.