Editorial policy

Every plain-English claim has to survive the source record.

Search engines penalize medical copy that a model invented. This page is the contract: what the rewriter may do, what it must cite, and what happens when it fails.

Pipeline

1. Fetch recruiting studies from the ClinicalTrials.gov Data API v2.
2. Store registry fields unchanged (NCTId, BriefTitle, BriefSummary, Phase, InterventionName, EligibilityCriteria, locations, contacts).
3. Run a rules engine: split inclusion vs exclusion, simplify stock phrases (placebo, double-blind), explain the Phase field using FDA/ICH labels.
4. Optionally pass the draft to Llama at temperature 0 (no sampling creativity) with a system prompt that forbids new drugs, sites, statistics, or eligibility.
5. Grounding check: if the model output contains a 6+ letter token that is not in the source blob and not on a closed allow-list, discard the model output and keep the rules draft.

Field map

Shown on every trial page:

• headline → BriefTitle + Phase + InterventionName + Condition
• inPlainEnglish → BriefSummary only
• whoItsFor → Inclusion criteria + age + sex
• whoShouldSkip → Exclusion criteria
• whatHappens → InterventionName + StudyType + DesignAllocation + BriefSummary
• effort → EnrollmentCount + site count + InterventionType

What we will not write

• “This treatment works” or any efficacy claim not quoted from the registry.
• Eligibility guesses (no SOC-2-style inference from missing pages).
• Contacts that were not published on the NCT record.

Transparency on the page

Sitewide medical disclaimer. ClinicalTrials.gov named as source. Direct NCT hyperlink. Organization and MedicalStudy / MedicalCondition / Dataset / FAQPage / BreadcrumbList JSON-LD. Operator identity on /about. Guides on /guides.