Problem
Pharma and biotech content teams go through MLR review before anything is published. Tracing evidence sentence by sentence is slow, risky claims slip through, and when a source changes it is hard to find every affected piece of content.
How it works
Every sentence gets one of five verdicts: supported, needs_review, unsupported, prohibited_for_audience or superseded_source.
Deterministic rules have final authority over approval status, prohibited claims, audience, source version and the audit log. Amazon Bedrock only handles semantic matching, review reasons and suggested edits. If Bedrock fails, a local reviewer takes over automatically. Nothing is approved or published automatically, and every decision is logged.
Demo
Using a fictional product, KTX-201: supported sentences pass, unsupported comparisons and prohibited claims are flagged, and when the evidence for a serious infection rate changes from 7.4% to 8.1%, the tool points to every affected document and sentence.
Limits and next
This is a prototype tested on fictional data. There are no real-world usage results yet.