Sheet 09 · Research
Checking Every Reference Twice
Status: Pilot Used on a literature review. As of 30.09.2026.
In plain words
A second AI reviewer checks each reference: does it exist, is it this paper, does it support the claim. Every change stays marked.
Title block
- Status
- Status: Pilot Used on a literature review.
- Engine
- Engine: an external AI service For this method, the input is a reference list, the sentences that cite it and the published papers.
- Human check
- Every change is marked in the document, so a person can see it.
Data touched · Reviewed
- Data touched
- A reference list, the sentences that cite it, and the published papers.
- Reviewed
- 30.09.2026 · Rev A
Uses an external AI service. As of 30.09.2026. Built for my own workflow with AI coding assistants.
Step 1 Three questions
Step 2 The marks
Text version
Step 1: a sample reference to the Journal of Sample Studies, checked against the questions: exists, this paper, supports the claim. Step 2: the marks a reference can carry: confirmed, corrected, claim adjusted, unverified, or withdrawn with a marker left in its place.
Spot the bad reference
Invented references for an invented claim. Which would the checks stop? Open each one to see its mark.
Invented claim: a sample rinse changes a sample outcome.
Sample, A. A sample trial of a sample rinse. Journal of Sample Studies.
Confirmed. It exists, it is this paper, and it supports the claim.
Sample, B. Sample rinses and sample habits. Journal of Sample Studies.
Claim adjusted. The paper is real but says something narrower, so the sentence is rewritten to match it.
Sample, C. A sample review that no search can confirm.
Marked unverified, in the text and in the list, with what failed.
What it does
A working rule for AI-assisted research: a reference that was not retrieved in the same session counts as a guess. Each reference is then checked by a second AI reviewer that did not write it, on separate questions: does it exist, is it this paper, does it support the claim.
A bad reference is marked, never silently deleted, so the claim it carried is rewritten or softened too.
Mistakes leaned toward overstating a positive result, and a single check once let a fabricated quotation through.
How it works
- Cite only what was retrieved in the same session.
- A second reviewer, which did not write the list, checks each reference.
- The questions: does it exist, is it this paper, does it support the claim.
- A bad reference is marked, never silently deleted.
- Every change stays marked, so a person can see it.
What it will not do
- Trust a single pass. Every reference gets a second, separate check.
- Delete a bad reference silently. It is marked in place.
- Accept a reference recalled from memory. It must be retrieved in the same session.
For colleagues
The third question, whether the paper supports the claim, can need the full text, not only the abstract. A quotation counts as verbatim only if it was read as raw text, not through a summary.