Academic Integrity Review: "Optimizing Human-Powered Energy Generation Using Gaussian Process Regression" (JoVE, DOI: 10.3791/69810)

Academic fraud report · Geng Detector

Summary

Verdict: Strong indicators of data fabrication and methodological incoherence. Four major findings collectively suggest the manuscript may be partly or wholly AI-generated, with experimental claims that violate basic hardware, physiological, and statistical principles. Key issues: (1) The reported Cholesky decomposition initialization time of 47 ± 3 ms on an STM32F334C8T6 (Cortex-M4, 72 MHz, 64 KB SRAM) is physically impossible for an n=112 double-precision covariance matrix, which requires ~100 KB of memory alone, far exceeding the MCU's SRAM. (3) The experimental timeline (40 minutes total) contradicts the Representative Results, which claim fatigue detection '>40 min'—after subjects had already entered cooldown and assessment. (2) Hardware description in Protocol 1.1.6 (customized fixed gear, 12.4:1 ratio) contradicts the human-subject section 2.2.4 (front-2/rear-9, 0.222 ratio). (4) 5-fold stratified cross-validation on n=7 participants is statistically incoherent. Confidence is moderate-high; these are internal consistency and physical-feasibility issues, not direct proof of misconduct.

Verdict

🔴 Strong indicators of fabricated experimental data and methodologically incoherent manuscript. Multiple findings involve physical impossibilities, timeline contradictions, hardware description inconsistencies, and statistical misuse. The pattern is highly consistent with AI-generated or copy-pasted content assembled without coherent experimental grounding.

Key findings

Hardware–physics violation: STM32F334C8T6 (72 MHz, 64 KB SRAM) cannot hold an n=112 double-precision covariance matrix (~100 KB) nor perform Cholesky decomposition in 47 ± 3 ms. Reported GPR initialization latency violates the device's memory and compute envelope.
Timeline–physiology contradiction: Protocol total session length is 40 minutes (5 min baseline + 5 warmup + 20 main + 5 cooldown + 5 evaluation), yet Representative Results report fatigue onset at '>40 min', when subjects should already be at rest.
Hardware description contradiction: Protocol 1.1.6 describes a fully customized fixed-gear bicycle with 48T/16T chainring/cog (12.4:1 reduction); section 2.2.4 then refers to 'front gear 2, rear gear 9 (ratio 0.222)'—a stock derailleur setup incompatible with the described modifications.
Statistical incoherence: 5-fold cross-validation on n=7 participants cannot be balanced; the claim of 'distribution integrity' with participant-level folds is methodologically unsound for such a small sample.

Evidence highlights

Reported GPR initialization: 47 ± 3 ms on STM32F334C8T6 with n=112 dataset.
SRAM budget: 64 KB total; required double-precision covariance matrix: ~100 KB.
Reported auxiliary timings (e.g., 15 ± 3 iterations, 8.3 ± 2.1 ms) appear suspiciously precise given the underlying hardware constraints.
Session protocol: 5 + 5 + 20 + 5 + 5 = 40 minutes total, vs. fatigue reported at >40 min.
Conflicting gear ratios: 12.4:1 (Protocol 1.1.6) vs. 0.222 (Experimental design 2.2.4).
Sample size: n=7 participants; cross-validation scheme: 5-fold.

Notes

DOI preserved as reported: 10.3791/69810.
Publication year listed as 2026 in the source report; this should be independently verified against the publisher record.
Institutional affiliation referenced: Westlake University (西湖大学); not independently confirmed from the source PDF.
These findings are based on internal logical and physical-feasibility analysis. No raw data, code, or participant logs were examined. Final determination of misconduct requires institutional investigation.