Skip to main content

​
Our starting assumption

Large language models make mistakes, and in published research they regularly cite sources that do not support their claims or do not exist. SynapseRx is engineered around that assumption rather than in denial of it. Three principles follow.

​
1. Citations are checked, not trusted

Every citation in a SynapseRx analysis passes through verification after generation:
  • Source policies. Each section of the analysis has rules about what kinds of sources are acceptable for its claims.
  • Link verification. Cited URLs are checked; dead, unreachable, or fabricated links are removed before the pharmacist sees the analysis.
  • Trusted-source screening. Citations are screened against a curated registry of trusted medical sources, and duplicates are collapsed.
A citation that fails any check is dropped, not displayed. We describe supported claims as source-backed; we deliberately avoid the word “verified” as a blanket promise.

​
2. Unsafe output is withheld, not shipped

Generated analyses pass readiness checks before delivery. Output that fails is held for pharmacist review instead of being presented as finished work. The pharmacist reviews, edits, and approves everything a patient or prescriber ultimately receives, and generation contracts prevent downstream documents (scripts, handouts) from introducing clinical findings that were not in the reviewed analysis.

​
3. Patient information is handled as PHI

  • Prescriber communications (SBAR reports) are sent over HIPAA-compliant email under a Business Associate Agreement with the email provider.
  • Lab attachments are encrypted at rest and purged on a schedule.
  • Each pharmacy’s data lives in tenant-isolated infrastructure with row-level security.
  • Non-PHI email (invitations, marketing) runs on separate, non-PHI infrastructure.

​
4. Public research is screened for direct identifiers before egress

The pharmacy coworker can research current public information through the approved AI Gateway. This is separate from the patient-facing clinical evidence pipeline.
  • The coworker is instructed to formulate a short general query rather than sending the open case to a search provider.
  • The server rejects definite patient, case, contact, account, and private-workspace identifiers before any public request.
  • Results return bounded excerpts, retrieval dates, and source links so the pharmacist can inspect where a current claim came from.
  • Public results are advisory. They cannot silently update an analysis, handout, SBAR, MTM packet, prescription, or pharmacy record.
  • A blocked, empty, or unavailable search does not disable chat. SynapseRx answers any stable, useful portion and states which current fact was not checked.

​
Continuous measurement

SynapseRx is evaluated against a synthetic clinical case corpus with expected-insight checklists and forbidden-content traps before changes ship, and production behavior is monitored daily for failures and quality drift. Telemetry-only shadow checks measure how well claims are grounded without altering clinical output.
Questions about our safety architecture, or want the detail behind any of this for your compliance review? Contact us at rxvipnetwork.com/contact.