coreSR.AIcoreSR
Usage Guide

INSPECT-SR

How to screen trials for trustworthiness with the integrity suite

1

What INSPECT-SR is and when to use it

  1. 1INSPECT-SR (INveStigating the Potential for Error in Clinical Trials, for Systematic Reviews) — Wilkinson et al.; Delphi-derived checklist, 21 items in 4 domains
  2. 2It asks whether a trial's data can be believed at all (fabrication, falsification, catastrophic error). It is NOT risk of bias: run RoB 2 / ROBINS-I separately and report both
  3. 3Unit of assessment: one assessment per trial publication. Companion reports of the same trial can be attached as supplementary files
  4. 4"Yes" is the concerning answer on every item. Four escalator items (1.1 retraction, 1.2 expression of concern, 3.2 figure manipulation, 4.8 impossible integer statistics) are each individually sufficient for serious concerns when confirmed
Tip: Domain and overall judgements are holistic reviewer decisions. The CoreSR default rollup (0 yes → no concerns; 1 non-escalator yes → some; ≥2 yes or an escalator → serious) is a suggestion, and the report labels it as such.
2

What runs automatically for every trial

  1. 1Retraction Watch / Crossref notices (1.1, 1.2) and PubPeer commentary (1.2, when a key is configured)
  2. 2Trial registry retrieval (ClinicalTrials.gov, ICTRP, ISRCTN, ANZCTR), prospective-registration timing (2.2) and registration-versus-publication diff (2.3)
  3. 3Structured facts extraction from the full text (design, flow, baseline table, ethics, registration) — schema-validated, verbatim numbers, null when unsure
  4. 4statcheck p-value recomputation for every t, F, χ², z and r reported with a p (4.9)
  5. 5GRIM/GRIMMER on integer-scale baseline cells (4.8); the Carlisle baseline p-value distribution method (4.3); participant-flow arithmetic (4.6); loss-to-follow-up, allocation-split and recruitment-rate rules (4.5, 4.2, 2.4)
  6. 6Ethics-reporting completeness (2.1), tortured-phrase screening (3.1), citation integrity — every reference resolved against Crossref, DataCite and PubMed with retracted references flagged (3.1) — and image forensics on the PDF (3.2)
Tip: Checks need the full text. Bulk-upload PDFs, or attach the PDF on the trial page; the panel says "no full text" when the text engines could not run.
3

Reading the integrity panel

  1. 1Flag = the engine found something; confidence high/medium/low says how much the reviewer must still verify. "Flag — verify" on GRIM/GRIMMER means the integer-scale label came from the extractor and needs confirming before 4.8 is answered "yes"
  2. 2Clear = the engine looked and found nothing. Not assessable = the inputs were missing (no registration ID, no baseline table, no test statistics). Neither is proof of trustworthiness
  3. 3Expand a row to see the evidence: the verbatim statistic and recomputed p, the unresolvable references, the per-variable Carlisle table, the extracted flow counts. Open the source link for registry records and notices
  4. 4Each item card shows badges for the checks that inform it; click a badge to jump to the row
4

Reviewing and completing a trial

  1. 1Work domain by domain: confirm or override each AI-suggested item response and record the evidence in the rationale (verbatim quotes are stripped if they cannot be located in the text)
  2. 2Record the domain judgements and the overall judgement with a rationale; the assessment cannot be marked complete without an overall judgement
  3. 3Use the audit trail drawer to see who changed what and when; export the per-trial Word report for the appendix
5

Project view, exports and sensitivity re-runs

  1. 1The trial × domain matrix marks unreviewed suggestions with an asterisk; filter by judgement or status, and use "Run checks (all)" after adding PDFs or changing project rules
  2. 2Export the project Word report (methods statement, matrix, item flag frequencies, check coverage, RAB, per-trial rationales) or the CSV matrix for analysis
  3. 3Link the pairwise or NMA project from the project page. The sensitivity page lists its outcomes, shows how each flagged trial maps to a study row (first author + year), and refuses a run that would exclude nothing
  4. 4Report the before/after effect, CI, I² and trial counts from the run history; the Word export of the run carries the AI executive summary
Tip: Trials that cannot be matched to a study row stay in the analysis and are listed; set a display name on the trial equal to the study label used in the analysis to fix a mismatch.
coreSR.AICoreSR

AI-powered evidence synthesis