Investigation Report: Severely Inconsistent Correlation Coefficients and Fabricated Statistical Significance in an AI-Driven Teacher Rotation Policy Simulation Paper

Academic fraud report · Geng Detector

Summary

This report assesses a 2026 paper published in the Journal of East China Normal University (Educational Sciences), DOI 10.16382/j.cnki.1000-5560.2026.08.003, which uses a multi-agent simulation to study teacher rotation policy. Verdict: Highly Suspicious (orange). Two findings are confirmed as red-severity issues. (1) The text states Pearson r = 0.073 for Scenario A and r = -0.184 for Scenario B (p < 0.05), but Figure 4 labels show r = 0.191 (Scenario A) and r = -0.148 (Scenario B) — the in-text and figure coefficients do not match, a classic AI hallucination symptom. (2) At n = 50 (df = 48), the threshold for p < 0.05 two-tailed is |r| > 0.273; r = -0.184 yields p ≈ 0.19 and r = -0.148 is likewise non-significant, so the claimed p < 0.05 is statistically impossible. A third finding (yellow) notes the suspiciously neat 82.5% comparison (0.265 vs 0.145) but is marked insufficient evidence. The authors' disclosure claims of manual verification and independent Python re-calculation directly contradict these elementary errors. Limitations: judgments are based on the paper text and figures only; raw simulation code and logs were not examined.

Verdict

Highly Suspicious (🟠). The paper contains internally contradictory correlation coefficients and a statistically impossible significance claim, both confirmed. The authors' explicit statements of rigorous human quality control are inconsistent with these errors.

Key findings

Text–figure inconsistency on correlation coefficients (Severity 🔴, Confirmed). The narrative reports Scenario A r = 0.073 and Scenario B r = −0.184 (p < 0.05), whereas the Figure 4 caption shows Scenario A r = 0.191 and Scenario B r = −0.148. The in-text and graphical values for the same underlying data do not agree.
Impossible p-value at n = 50 (Severity 🔴, Confirmed). With n = 50, df = 48, the two-tailed p < 0.05 threshold requires |r| > 0.273 (p < 0.01 requires |r| > 0.349). Both reported coefficients (r = −0.184 and r = −0.148) fall below this threshold; r = −0.184 corresponds to p ≈ 0.19. The stated significance is therefore not supported by the numbers given.
Disclosure inconsistency (Severity 🟠, Confirmed). The authors' appendix declares manual verification, paired-t re-calculation in Python, and one-by-one figure–text reconciliation. The errors above are precisely the kind that such checks should have caught, undermining the credibility of the disclosure.
Suspiciously clean comparative figure (Severity 🟡, Insufficient evidence). Scenario B vs Scenario A teacher attrition is given as 0.265 vs 0.145, reported as "82.5% higher." Direct calculation yields (0.265 − 0.145) / 0.145 = 82.76%, close but not identical to the stated 82.5%. The neatness of the contrast is noted but not, on its own, conclusive evidence of fabrication.

Evidence highlights

DOI: 10.16382/j.cnki.1000-5560.2026.08.003
In-text Scenario A r = 0.073; Scenario B r = −0.184, claimed p < 0.05.
Figure 4 caption: Scenario A r = 0.191; Scenario B r = −0.148.
Sample size per scenario: 50 independent replications, so df = 48.
Critical |r| at df = 48: 0.273 for p < 0.05 (two-tailed); 0.349 for p < 0.01.
Implied p-value for r = −0.184, n = 50: approximately 0.19 (not significant).
Attrition figures cited: 0.265 (Scenario B) vs 0.145 (Scenario A); reported gap 82.5%; computed gap 82.76%.

Notes

All numeric values are transcribed exactly from the report; no new measurements were introduced.
Findings 1 and 2 rest on basic arithmetic and the standard Pearson t-test formula; no specialized statistical methods are required to verify them.
The authors have not yet been contacted; this is a desk-based review.
Recommended follow-up: request raw Python simulation code, random seeds, and full statistical output logs; raise the r-value and p-value discrepancies on PubPeer; notify the editorial office of the Journal of East China Normal University (Educational Sciences).
Final determination of misconduct requires investigation by the relevant institution; this report is an evidence compilation, not a ruling.