Figure and Caption Audit

Check what the figure tells a reader before polishing how it looks. A visually clear diagram can still describe the wrong experiment.

Set the scope

Identify the requested deliverable: findings, caption repair, diagram instructions, or an actual asset edit. Preserve the author's sound structure and source assets. Use the current manuscript and relevant protocol as evidence; an older figure approval or review memo does not establish agreement with today's source.

Inventory the available representation: textual specification, editable source, exported image, or rendered manuscript page. These support different checks. A text specification supports a semantic review, not conclusions about pixels, legibility, cropping, color, or page layout. A source edit does not establish that an exported or embedded figure changed.

Audit the scientific meaning

1. Map the elements. For each box, curve, panel, symbol, or legend label, identify the real object it represents and its source definition. Distinguish collected records, eligible training examples, diagnostics, predictions, and measured outcomes. Keep names and units consistent across the figure, caption, Methods, and results.
2. Read every arrow as a claim. Determine whether it means data flow, temporal order, optimization, selection, evaluation, or a hypothesized relationship. Verify the endpoints and direction. Training, validation, and test feedback are different operations. Changing an unsupported feedback arrow to a dotted line does not establish a supported meaning; remove it or supply an accurate label and connection.
3. Check populations and comparisons. Show which examples each method can use. Shared architecture, nominal update budgets, or a common archive do not imply identical training support. Keep diagnostic-only records outside training paths. For charts, check denominators, conditions, uncertainty definitions, and whether panels describe the same experiment before allowing direct comparisons.
4. Match the measured endpoint. Trace the final stage to the actual outcome criterion. Preserve all required conditions, including evaluation stopping points. A system that detects a target has not thereby completed the downstream intervention. A schematic output icon should not imply a measured capability absent from the study.
5. Repair the caption and figure together. A corrected caption cannot reliably cancel a contradictory arrow or label. Put interpretation-critical conditions where a reader can find them; leave detailed settings to the paper when they are not needed to understand the figure. Distinguish an illustrative workflow from observed experimental results.

For example, a dataset may contain accepted images and excluded images kept for an audit. A single “all images → training” arrow misstates support even if the caption calls the archive comprehensive. Route the audit separately while preserving the useful overall organization.

Prose when drafting or revising

When drafting or revising prose, including audit explanations, read prose editing rules. Preserve scientific meaning, necessary reasoning and valid technical terms. Keep sound passages, do not add unreported properties, and limit edits to the requested scope. The local reference covers concept accuracy, clear sentences, repetition and punctuation, including the required exceptions.

Verify and report

Return a short verdict, then each consequential mismatch as element/location → implied claim → source evidence → smallest repair. Provide a replacement caption or edit instructions when requested. State which valid elements should remain. Separate scientific repairs from optional visual preferences.

If image or page evidence is available and visual review is in scope, inspect the actual artifact at a useful size. After authorized edits, check the new export and relevant manuscript placement when accessible. Report the exact representation and coverage inspected; do not reuse a stale preview as verification. When only text is available, label the outcome a specification review and identify the remaining visual check.

Stop when the requested figure and caption accurately communicate the supplied evidence. Do not redesign a sound figure, manufacture results, alter the scientific protocol to fit the artwork, or claim a render was checked without observing it. No justified change is a valid result.