Figure Rhetoric Audit

/manuscript-pipeline

Purpose

Relationship to Other Skills

Workflow

1. Ingest

1. Compiled PDF (preferred) — use the Read tool on the PDF file to inspect each page containing a figure at actual rendered size
2. Individual figure files — use the Read tool on each

   .pdf

   ,

   .png

   ,

   .jpg

   figure file referenced by

   \includegraphics

3. If neither is available — ask the user to compile the PDF or provide figure files. Do NOT proceed with source-only analysis. A rhetoric audit without seeing the figures is meaningless.

1. Visually inspect the figure. Read the figure file or the PDF page containing it. Before reading any prose, record what the figure communicates at first glance — the immediate visual takeaway. What pattern, trend, comparison, or relationship does a naive reader see?
2. Extract the claim context. Read the 2-3 paragraphs surrounding the first

   \ref{fig:X}

   reference. Identify the specific claim the figure is supposed to support. Write it down as a one-sentence assertion.
3. Read the caption. Does the caption tell the reader what to see, or does it just describe the axes?
4. Compare. Does the visual takeaway (step 1) match the claimed assertion (step 2)? The figure must be evaluated through the reader's eyes, not the author's intent.

2. Per-Figure Analysis

• Identify the prose claim (e.g., "Method A converges faster than B")
• Identify the visual impression (e.g., "Two nearly identical curves")
• If these differ: the figure fails its rhetorical purpose regardless of whether the data technically supports the claim

• Invisible difference: The text claims a meaningful difference but the figure's scale, aspect ratio, or data density makes the difference imperceptible. The data is there; the visual isn't.
• Wrong emphasis: The figure shows many things, the text discusses one. The reader doesn't know where to look.
• Contradictory impression: The visual impression actively suggests the opposite of the claim (e.g., "convergence" but the curve is still trending; "improvement" but the bars look equal).
• Unstated context: The figure requires domain knowledge to interpret that the surrounding text doesn't provide.

• Y-axis origin: Starting at 0 when differences are small makes them invisible. Starting above 0 when showing absolute quantities is misleading. The choice must serve the claim:
  • Showing "our method is better" → zoom in on the relevant range
  • Showing "the effect is small relative to the total" → start at 0
  • Always: clearly label the axis range; broken axes with

    //

    if needed
• Axis scale: Linear vs logarithmic. Log scale appropriate when data spans orders of magnitude or when relative differences matter more than absolute. Flag linear scale with data spanning 3+ orders of magnitude.
• Axis range: Does the range include all relevant data? Does it extend far beyond the data (wasting space)? Is the range chosen to exaggerate or minimize visual differences?
• Aspect ratio manipulation: A very wide or very narrow plot can exaggerate or flatten trends. The slope of a trend line should be perceptible but not exaggerated.
• Dual axes: Almost always confusing. Two different y-axes invite incorrect visual comparisons. Prefer: two separate panels with aligned x-axes.

• Primary element: The data series, region, or comparison that the text discusses should be visually dominant (thicker line, bolder color, larger markers, foreground position).
• Secondary elements: Context, baselines, and reference data should be visually recessive (thinner lines, gray, smaller markers, background).
• Annotations: If the text references a specific point, region, or threshold, it should be annotated in the figure (arrow, callout, shaded region, dashed reference line). The reader should not have to decode coordinates to find what the text describes.
• Cognitive load: Count the number of distinct visual elements the reader must track. More than 5-7 series/groups in one plot exceeds working memory. Split into panels or highlight the comparison of interest.

• All lines/bars have equal visual weight — reader doesn't know what's important
• The "our method" line is the same width and style as 10 baseline lines
• Text says "note the divergence at epoch 100" but nothing in the figure marks epoch 100
• Legend has 15 entries — reader must constantly reference it

• Overloaded: Too many series, categories, or data points for a single plot. The claim is about the relationship between 2 methods but the plot shows 12. Simplify: show the comparison of interest prominently; relegate others to supplementary material or a secondary panel.
• Underloaded: The plot shows so little data that the claim isn't convincing. A single point where a trend is claimed. Two bars where a distribution is relevant. Three epochs where convergence behavior matters.
• Summarized when raw matters: Showing only means when the variance is the story (or when it would undermine the story — flag both). Confidence intervals, error bars, or violin plots for stochastic results.
• Raw when summary matters: Individual data points where the aggregate pattern is the claim. A scatter plot of 10,000 points where a density plot or binned heatmap would show the structure.

1. Descriptive only (weak): "Training loss over epochs for five methods." The reader must figure out the takeaway.
2. Directive (adequate): "Training loss over epochs. Method A (red) converges in ~50 epochs while baselines require 150+." Points the reader to the claim.
3. Interpretive (strong): "Training loss over epochs. Method A (red) converges 3x faster than the nearest baseline (blue), supporting the claim that architectural change X reduces optimization difficulty." Connects the visual to the argument.

• Area encoding for quantities: If using bar width, bubble size, or area to encode values, the mapping must be proportional to area (not radius or diameter). A value 2x larger should have 2x the area, not 2x the radius (which is 4x the area).
• 3D effects: 3D bar charts, 3D pie charts, perspective effects — these distort perception of values through foreshortening. Flag any 3D visualization of 2D data.
• Color scale linearity: Sequential color maps (viridis, plasma) have perceptually uniform steps. Rainbow/jet color maps have perceptual discontinuities that create artificial boundaries in continuous data. Flag rainbow/jet for continuous data.
• Truncated axes without indication: Y-axis not starting at 0 without a visible break (

  //

  notation) can mislead readers about relative magnitudes.
• Aspect ratio distortion: Pie charts not circular. Bar widths inconsistent. Maps with incorrect projections (rare in ML papers but common in spatial analysis).

• Redundant figures: Two figures showing essentially the same information in different forms. Unless each adds distinct insight, merge or choose the more effective one.
• Missing figures: Is there a key claim in the text that has no visual support but would benefit from one? A figure that isn't there is a missed opportunity if the claim is central.
• Figure ordering: Do the figures appear in the order of the paper's argument? Architecture → training → results → analysis is a natural arc. Figures out of narrative order confuse the reader's mental model.
• Visual consistency across figures: Same method → same color/marker across all figures. Same data → same axis scale when comparison is relevant. Cross-reference with manuscript-review §12 (visual language consistency).

3. Common Antipatterns

4. Generate Report

## Figure [N]: [brief description]

**Prose claim:** [one-sentence claim the figure is supposed to support]

**Visual takeaway:** [what a reader actually sees at first glance]

**Alignment:** [STRONG | ADEQUATE | WEAK | CONTRADICTORY]

**Issues:**

1. [Dimension]: [specific problem]
   - **Impact:** [how this affects the reader's understanding]
   - **Fix:** [specific recommendation with concrete changes]

**Caption assessment:** [Level 1/2/3] — [recommendation if upgrade needed]

**Recommended changes:** [prioritized list]

## Figure Set Assessment

**Overall narrative coherence:** [figures tell a coherent story / gaps exist / redundancies]

**Strongest figure:** Figure [N] — [why it works]
**Weakest figure:** Figure [N] — [primary issue]

**Priority fixes:**
1. [Most impactful change across all figures]
2. [Second most impactful]
3. [Third most impactful]

**Missing figures:** [claims that need visual support but lack it]
**Redundant figures:** [figures that could be merged or cut]

5. Output

[manuscript-name]-figure-rhetoric-report.md

• Count of figures by alignment rating (STRONG / ADEQUATE / WEAK / CONTRADICTORY)
• Top 3 fixes by impact
• Any CONTRADICTORY figures (highest priority — these actively hurt the paper)

Core Principles

• The reader is naive. Do not evaluate figures through the author's eyes. The author knows what the figure "should" show. The reader sees only what is visually present. Every judgment in this audit is from the reader's perspective.
• Claim first, then figure. Read the prose claim before looking at the figure. The figure's job is to support that specific claim. If the figure is beautiful but doesn't support the claim, it fails.
• One figure, one message. A figure trying to show three things shows none of them clearly. If the text makes three claims about one figure, the figure is overloaded or the text should point to three figures.
• Visual perception trumps data accuracy. A figure can be numerically correct but perceptually wrong (e.g., differences invisible due to scale). The reader's visual impression IS the communication. If the impression doesn't match the claim, the figure has failed.
• Concreteness over abstraction. Recommendations specify the exact change: "change y-axis range from 0-100 to 85-95," not "consider adjusting the axis." Include suggested chart types, axis ranges, color choices, and annotation text.
• Severity scales with claim importance. A weak figure supporting a minor methodological point is LOW priority. A weak figure supporting the paper's core result is CRITICAL — it's the first thing a reviewer scrutinizes.