For teams shipping health AI
Your AI cites a page.
That page cites a review.
The review rests on a retracted trial.
MoolTrace builds a citation graph in FalkorDB from the answer, through each page and review, down to the trials. It blocks the answer if a source it relies on was retracted, warns on an Expression of Concern, shows you the exact path, and drafts a fix.
A citation check passes all 35 Harvard nutrition pages.
MoolTrace flags 2, and I checked both by hand.
Not in the graph? MoolTrace says “not checked”, never “safe”.
Page
Review
Trial
Live · Harvard vitamin D page
Watch it catch one.
click to look deeper
Page
Harvard vitamin D
checking…
Following citations…
At 1 hop the page looks clean: nothing it cites directly is flagged.
2 of 35
A citation check passes all 35 — MoolTrace flags 2.
Harvard nutrition pages that rest on withdrawn or retracted research. I checked both by hand.
Zinc for the common cold
Review · Cochrane 2011 (Singh) · cited by Harvard zinc
Withdrawn by Cochrane in April 2015 over concerns about the calculation and analysis of data.
Withdrawn by Cochrane · 2015Every page linked from Harvard's Vitamins & Minerals index · checked 17 Oct 2026 · each hit verified by hand
How it works
What goes in, what the graph does, what comes out.
In
Your answer
A draft, its links, or a call from your agent.
The graph · FalkorDB
Page
Review
Trial
Notice
strongest chain: 0.55 × 0.85 = 0.47, found with algo.SPpaths
Out
✕Don't ship
Rests on research retracted in May 2025, 2 hops down. The path and the sentence behind each hop come with it.
The fix
Harvard reviews the evidence at nutritionsource.hsph.harvard.edu/vitamin-d.
An earlier version of this answer cited the Harvard Nutrition Source page “Vitamin D”, which cites Avenell et al. 2014, which uses data from Ishida et al. 2004, a paper that was retracted in May 2025, so that source was removed.
- 1
You give it an answer
Paste a draft or the links it cites. Or let your agent call it over MCP.
- 2
It builds the graph
Every page gets resolved to the papers behind it (Crossref, OpenAlex, PubMed) and stored in FalkorDB as page, review and trial. Retractions and concerns hang off the papers they flag.
- 3
It weighs every link
A trial the review actually included counts fully. A passing mention barely counts. FalkorDB's shortest path search finds the strongest chain to a flagged paper.
- 4
You get a verdict
Don't ship, Review, Clear or Not checked. Always with the path and the sentence behind each hop.
- 5
It fixes the answer
Only the sentence that leaned on the bad source changes. Then it checks the answer again.
Track 01 · Agents that act on connected data
Works where answers get made.
Your agent checks itself.
The model picks the tools. The graph decides. Only tool output ships.
check_answer · don't ship
Walks page → review → trial in FalkorDB.
explain_path · Harvard → Cochrane → Ishida
Shows the exact chain and the sentence behind each hop.
rewrite_answer · fixed, clear
Removes the retracted source, then checks again.
Without MoolTrace · ships as written
Vitamin D helps prevent fractures in older adults. Harvard reviews the evidence at nutritionsource.hsph.harvard.edu/vitamin-d.With MoolTrace · checked again: clear
Vitamin D helps prevent fractures in older adults (no checked source remains for this claim). An earlier version of this answer cited the Harvard Nutrition Source page “Vitamin D”, which cites Avenell et al. 2014, which uses data from Ishida et al. 2004, a paper that was retracted in May 2025, so that source was removed.Any agent, over MCP.
Point Claude, Cursor or your own agent at the remote endpoint.
https://mooltrace.h9g8gwz2b559t.ap-south-1.cs.amazonlightsail.com/mcpCI blocks the PR.
Fail the build when an answer rests on a retracted source.
mooltrace batch --fail-on block
exit 1Retractions arrive later.
When a paper is retracted, find every shipped answer that rests on it.
Trial · Ishida 2004 · Retracted
← Review · Cochrane 2014 (Avenell)
← Page · Harvard vitamin D
← stored answers citing it
0 → 31 of 40
reviews distorted by retracted trials (BMJ 2025 answer key): a DOI check finds 0, MoolTrace reaches the trial in 31
0.43 → 1.00 recall
on 52 labelled answers, DOI-only check vs MoolTrace
2 of 35
Harvard nutrition pages rest on withdrawn or retracted research

