Protocol Mirror
WebMCP preview
Clinical-trial outcome checkClinicalTrials.gov registry vs PubMed publication

Did the trial publishwhat it registered?

Load a real ClinicalTrials.gov record and its PubMed report. Your agent quotes the exact registered and reported text through WebMCP and stages each match or discrepancy for you. AI assembles evidence. A human decides.

WebMCP previewTools appear when an agent connects

No agent is connected. Open this page in ChatGPT's in-app browser, or in Chrome 149+ at protocol-mirror.vercel.app (that origin carries a WebMCP origin-trial token, so no flag is needed). Loading a real trial and reviewing by hand work without one.
  1. 01Agent reads both recordsExact registry and abstract text, flagged as untrusted
  2. 02Agent proposes with citationsMatched, omitted, introduced, or uncertain, each with quoted spans
  3. 03You decideAccept or reject in the page; no agent tool can
  4. 04Exports only what you approvedThe receipt tool appears after your first decision

Ask your agent

Three prompts for the case on this page, for ChatGPT's in-app browser or Chrome with WebMCP. The agent reads and proposes; you decide in the review queue below.

Active demonstration case

A randomized study of Cardioluma for persistent hypertension

Deterministic fictional record · no clinical claim

RegistrationNCT-DEMO-042Updated 2025-02-14
PublicationPMID-DEMO-2042Published 2026-01-09
SponsorNorthstar TherapeuticsPhase 3 · deterministic demonstration record
Review progress0 / 40 awaiting a human decision

Load a real trial

Review a real registry-to-publication pair.

The same bounded ClinicalTrials.gov and PubMed readers the agent uses. Publication text arrives as abstract sections, not extracted outcomes; every proposal still waits for your decision.

Real-world stress test

Two models, the same real evidence, opposite errors.

Two local models were blinded to 24 published labels and restricted to the exact registry and abstract text these tools return. One over-called changes, the other missed them. That is why every proposal shows its quotes and why the final call stays with you.

24real NCT/PMID pairs12 labeled change · 12 labeled no change
48 / 48live WebMCP reads succeeded172 registry outcomes · 106 abstract sections
Over-calledqwen3:4bCalled every one of its 10 decided no-change cases a change
Under-calledornith-1.5:9bMissed 10 of its 11 decided change cases
Strict grounding resultqwen3 17.9% · ornith 69.0% unsupported claims0 authority attempts · 0 misconduct claimsRun-specific evidence—not a universal hallucination or clinical-accuracy claim.

Evidence table

Registered intent reported record

MatchedFlaggedUnreviewed
01

Registered outcomes

ClinicalTrials.gov protocol record

0proposed relationships
02

Reported outcomes

Journal publication record

Human checkpoint

Review queue 0

No human decisions recorded yet.

The queue is clear.

Ask an agent to inspect the case with WebMCP, or stage the guided demonstration.

This session

What the agent called, what you decided.

Kept in this browser tab only, so a reload does not lose the case. The receipt tool stays locked until your first decision.

  • get_live_clinical_trialread-only
  • get_live_pubmed_articleread-only
  • get_registry_historyread-only
  • get_audit_stateread-only
  • get_evidence_spansread-only
  • propose_outcome_mappingstages only
  • request_human_reviewfocuses only
  • export_review_receiptlocked until you decide

No tool calls yet this session. Ask your agent to call get_audit_state, or load a real pair above.

Inspectable reasoning

Evidence drawer

Select any outcome to inspect its exact source span, or select a staged proposal to inspect the agent rationale.