Adverse-event capture
Potential adverse-event mentions are flagged as they occur and routed into a pharmacovigilance review queue, with per-tenant configuration for where those flags go and who receives them.
Safety & governance
Five rules that run on every response.
The guardrails are set at the platform level: written into how every answer is generated, and checked against each one, with every result recorded. No deployment can switch them off, and production refuses to start without them.
The guardrails
Enforced after generation, not requested before it.
A prompt asking a model to behave is a preference. These run on the generated text itself, which is why they hold when a question is phrased in a way nobody anticipated.
- G1
No directive language
Answers are written not to tell a patient what to do: they explain what sources say and route the decision back to a clinician.
- G2
No diagnosis or assessment
Answers are written not to diagnose, stage or assess an individual’s condition, and every generated answer is checked for language that does.
- G3
No unattributed superlatives
Comparative and superlative claims about a therapy are written with an attributed source, and every generated answer is checked for ones without.
- G4
No personal health information
Every generated answer is checked for personal health information, in addition to redaction on the way in.
- G5
Mandatory citations
A section without citations fails the check and is removed from the answer, and the failure is recorded.
Review workflows
What happens around the answer.
Deferral and follow-up
Questions it should not or cannot answer get a deferral, not an approximate answer. Where a deployment enables it, they are also captured — as a redacted topic, never patient details — in a human-response queue for a source-grounded follow-up.
Medical, legal and regulatory review
Content changes move through pull-request review, which produces a reviewer, a timestamp and a diff for every published claim — the evidence trail an MLR process needs.
Auditability
Operational logs cover system activity, usage and deployment behaviour, including guardrail outcomes, so a review can ask what the system did and get an answer.
Questions we get asked
Straight answers.
- Does OpenSupport give medical advice?
- No. It provides health information with citations and directs clinical decisions back to a healthcare professional. Validator rules block directive language and any diagnosis or assessment of an individual.
- How are adverse events handled?
- Potential adverse-event mentions are flagged and routed into a pharmacovigilance review workflow, configurable per tenant. The platform supports capture and routing; it does not replace your pharmacovigilance obligations or your reporting timelines.
- Does the platform store personal health information?
- It is built not to. Inbound text is redacted for personal identifiers before retrieval or generation, and every answer is checked for personal health information after it is generated. The patient experience does not require a stored medical history to answer a general information question.
- Can a sponsor influence which sources an answer uses?
- No. Retrieval and ranking logic is identical for every tenant. A deployment can be scoped to a defined portfolio or to its own private content, and that scope is enforced during retrieval — but within scope, nothing can be paid into a higher position.
- What happens when the system does not know?
- It says so rather than approximating an answer, and where a deployment enables it, the question is captured in a human-response queue for follow-up.
Bring your hardest compliance question.
The fastest way to evaluate the boundaries is to try to cross them. We will give you a deployment to do exactly that.