Develop external consistency checking for visually undetectable figure tampering

Develop an external consistency-checking method that compares rendered scientific figures against structured data to detect internally consistent legend and category-label manipulations that leave no in-image evidence of tampering.

Background

The failure analysis finds that legend and category swaps can produce figures that are geometrically unchanged and internally coherent, making the tampering visually undetectable from the image alone. The paper argues that overcoming this limitation requires checking the rendered figure against an external representation such as structured data, and explicitly leaves that concrete development task for future work.

References

We conclude that the remaining errors are not a perception problem solvable by higher resolution or better prompting, but would require an external consistency check (e.g. comparing the rendered figure against the structured data), which we leave to future work.

— SciTrue: Reliable Scientific Claim Validation with Frontier and Open Language Models at the NTCIR SciClaimEval Task  (2609.00654 - Bao et al., 1 Sep 2026) in Section 6, Failure Analysis