Governance / Clinical-grade
Phase 1 / Pilot intake open
The engine

One governed engine.
Three trust zones.

The defensibility of VETA is not the language model. It is the hidden control layer that decides what outputs are permitted, for whom, under what evidence threshold, and with what audit trail.

EngineVETA
Evidence storeV-EviCore
Audit gradePer-output trail
Request pipeline

Every request passes through six gates.

Gate 01
Identity router

DVM credentialing, role mapping, channel routing. Outputs differ by audience by design.

Gate 02
Species & condition gates

Anatomical and physiological constraints. Dogs and cats at MVP; each new species clears an evidence and governance gate.

Gate 03
Risk engine

Deterministic triage classification, dose ceilings, toxicology, and interaction screening. Rule-driven from V-EviCore.

Gate 04
Evidence governance

V-EviCore object store. Tier, source, last reviewed, version, lineage — every object addressable and auditable.

Gate 05
Output contracts

Audience- and risk-class-specific contracts. Sections fixed by rule, prose generated under guardrails.

Gate 06
Audit trail

Rules fired, evidence used, items blocked, template rendered. Stored with every output, queryable by clinical board.

V-EviCore / Evidence data model

Evidence is the moat.

Built in-house with a clinical content team — 2–3 DVMs on retainer and a PhD toxicologist. Reference texts are licensed, never the runtime source of truth.

evidence_id
Stable identifier, never reused.
tier
Strength of recommendation — guideline / consensus / expert / clinical use.
species_scope
Applicable species and signalment constraints.
condition_scope
Condition families, presentations, contraindications.
source
Citation, edition, version, last reviewed by named DVM.
version
Monotonic version number — every change is a new row, nothing mutates.
Determinism boundary
Rule-driven

Every safety-critical output. Triage classification, dose ceilings, species toxicology, drug interactions, output-contract section structure.

Model-assisted

Intake parsing, summarization, prose generation under guardrails. Never the safety decision itself.

The same evidence store serves all three channels. The outputs never cross.