How a report is built
An Erudio report describes what clinicians said in public about one clinical trial, why their readings differ, and the source behind each view. This page sets out what every report guarantees, how it is checked, and where its limits are.
Say what was said. Then say what the evidence shows.
A trial's headline can say more than its evidence, and commentary about a trial can say more than either. Our reports keep three things apart: what a clinician said, what the primary sources establish, and what remains unknown.
When a clinician's reading differs from the primary evidence, the report keeps their view as they said it and cites the primary source beside it. Comparisons that clinicians draw between trials stay attributed views, never a ranking.
A report leads with the conversation that actually happened. It adds neither enthusiasm nor skepticism for balance, and a well-supported favorable reading can lead when that is what clinicians said.
Reports never judge efficacy, never estimate how many clinicians hold a view, and never give treatment advice. Counts describe the sources we retained, not the profession.
What every report promises
Every claim has a source
Speech is kept apart from evidence
Numbers stay exact
Identity at its real confidence
Our own words are checked
Gaps are stated
Seven stages
- 01
Resolve the trial
Each report is bound to one registry record. We resolve the NCT ID, titles, acronym and aliases, interventions, conditions, sponsor and design from ClinicalTrials.gov, and exclude look-alike trials that share an acronym.
- 02
Gather sources
We gather public commentary that names the trial: podcast episodes, YouTube video, X threads and X videos. Registry results, protocols, publications and regulatory documents are gathered as context. Source versions and the retrieval time are recorded.
- 03
Screen and extract
Every source is screened for relevance. From each relevant passage we extract claims: a view, a question, an experience, a reason. Each keeps its expression type and the conditions the speaker stated, such as population, comparator, endpoint and timepoint.
- 04
Audit
An independent audit checks every claim against its original passage. Errors are corrected or removed, and omitted views are added. A report with an unresolved material defect stops here and is never written up.
- 05
Compose
Verified claims are grouped by trial aspect and composed into findings. Each finding leads with what clinicians said and why, with the primary facts needed to read it placed beside it. Aspects the commentary does not answer are marked incomplete, not filled in.
- 06
Review
An independent review checks every finding and summary against the evidence, for fidelity and for claims of certainty, efficacy, consensus or comparison that the sources do not support. A report that does not pass stays an unpublished draft.
- 07
Version and publish
Each accepted edition gets a release hash, an edition number and a semantic version. The report page, the catalog record, and the JSON and Markdown versions are generated from the same record.
How AI is used
Reports are produced with AI models working inside these checks. Models extract, audit, write and review; each step is checked against the original sources by a separate pass, and a report is published only when it passes. Nothing is accepted on a model's word alone: every claim has to trace to its source.
Three ladders on every report
Screened sources were retrieved and reviewed for the trial. Cited sources back at least one verified claim, and include primary sources used for context.
People found in screened sources, including name-only mentions; people identity-verified, matched to an Erudio clinician profile (with an NPI where available); and voices cited in the report.
Focused if the edition cites 1 to 12 commentary sources, Standard if 13 to 40, Deep if 41 or more. A focused report is still useful: it tells you the public conversation is quiet.
Editions, versions and corrections
Subscribed reports are checked weekly. Versions follow MAJOR.MINOR.PATCH: a major version changes the report's structure or contract, a minor version adds, changes or corrects a finding, and a patch adds evidence without a material change. A check that finds nothing new updates only the “last checked” date.
Subscribers are notified on minor and major versions. Corrections are listed on the report page with their date. Pages for withdrawn or superseded reports stay online with their status, so citations keep working.
What a report cannot tell you
- Public commentary is not a sample of the profession. Quiet clinicians are invisible to us.
- Retrieval finds sources that name the trial. References that do not name it, and transcription variants, can be missed.
- Opinions, hypothetical decisions and self-reported experience are preserved as such, not as established facts.
- Repeated posts and overlapping media versions are not independent observations.
- Source publication dates are not assumed to be readout or session dates.
- A report does not judge whether a treatment works, detect sponsor spin, or audit every protocol amendment.
- It is not medical advice and not for promotional use.
- Each report lists its own known gaps on its page.
Public sources only
Reports use publicly available podcasts, videos, posts, publications, registry records and regulatory documents. We do not use private data or content behind a login. Free previews paraphrase findings and refer to speakers by role; the full report links each claim to its source and names each speaker as the source names them. A speaker introduced only by a first name is marked as such; we never complete a name. Reports and the MCP return passages and quotes with a link to the original, never full transcripts.
Detailed method documentation is available to customers and qualified prospects under NDA.
Questions about method: contact us.