figure-verifier — Figure Claim Verifier

The failure this skill prevents: a paper states a result in the abstract or results section, but the figure tells a different story — different magnitude, different trend, different units, or data that doesn't reach the claimed threshold. These discrepancies are easy to miss when reading quickly and impossible to catch without actually reading the figure data.

The core rule, parallel to citation verification: every numerical value you report from a figure must be read from the actual image you can see — not recalled from training data, not inferred from the paper's reputation, not estimated from the figure caption alone.

Maturity note: The reading-precision of the image (VLM) path depends heavily on figure resolution and on the host model's vision capability. Reading a value off a dense plot is inherently approximate — prefer APPROXIMATE with a stated precision over MATCH unless the value is unambiguous. When a figure is too low-resolution or cluttered to read confidently, UNREADABLE is the correct, honest answer.

What you need to run a check

Required:

• The figure image (user pastes it into the conversation, or you fetch it from an open-access source)
• The specific claim to check — either provided by the user or extracted from the abstract/results

Optional but helpful:

• The paper DOI (lets you fetch the abstract to find relevant claims automatically)
• The figure caption (helps identify what axes and conditions the figure represents)