mathews-tom/armory/figure-rhetoric

Figure Rhetoric Audit

Pipeline position: Phase 1b (content audit). Runs in parallel with manuscript-review. Requires compiled PDF. No prior dependencies. See /manuscript-pipeline for full execution order.

Purpose

Evaluate every figure in a manuscript as a rhetorical act — a visual argument that must land with the reader. Each figure exists to communicate a specific claim. This skill audits whether the figure actually achieves that communication, or whether it undermines, obscures, or contradicts the author's intent.

A figure that is technically correct but rhetorically ineffective is a wasted opportunity. Reviewers form judgments from figures before reading the methodology. A figure that fails to show what the text claims creates doubt even when the underlying data supports the claim.

Relationship to Other Skills

Boundary: manuscript-review §24 checks "does the prose match the figure?" This skill checks "does the figure communicate what the prose needs it to?" §24 catches factual mismatches (text says 14.3%, figure shows 13.8%). This skill catches rhetorical failures (text says "dramatic improvement," figure shows bars that look identical because the y-axis starts at 0).

Workflow

1. Ingest

CRITICAL: This skill requires visual inspection. LaTeX source alone is insufficient. The entire point of this audit is what the reader sees.

Obtain the rendered figures by one of:

1. Compiled PDF (preferred) — use the Read tool on the PDF file to inspect

each page containing a figure at actual rendered size

1. Individual figure files — use the Read tool on each .pdf, .png,

.jpg figure file referenced by \includegraphics

1. 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.

For each figure:

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?

1. 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.

1. Read the caption. Does the caption tell the reader what to see, or

does it just describe the axes?

1. 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

For each figure, evaluate across 8 dimensions:

Dimension 1 — Claim-Figure Alignment

The central question: does a reader who looks at this figure see the claim the text makes?

• 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

Failure modes:

• 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.

Dimension 2 — Chart Type Appropriateness

Is this the right type of visualization for this claim?

Flag when a more effective chart type exists for the claim being made.

Dimension 3 — Axis and Scale Design

Axes can make or break the visual argument.

• 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.

Dimension 4 — Visual Hierarchy and Emphasis

Does the figure guide the reader's eye to the right element?

• 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.

Failure modes:

• 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

Dimension 5 — Data Density and Simplification

Is the figure showing the right amount of data for its claim?

• 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.

Dimension 6 — Caption as Interpretation Guide

The caption should tell the reader what to see, not just what the axes are.

Levels of caption quality:

1. Descriptive only (weak): "Training loss over epochs for five methods."

The reader must figure out the takeaway.

1. Directive (adequate): "Training loss over epochs. Method A (red)

converges in ~50 epochs while baselines require 150+." Points the reader to the claim.

1. 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.

For each caption, identify its level and recommend upgrading if Level 1. Level 2 is the minimum for effective communication. Level 3 is ideal for key figures supporting core claims.

Dimension 7 — Perceptual Accuracy

Does the visual encoding accurately represent the underlying data?

• 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).

Dimension 8 — Redundancy and Narrative Arc

Do the figures as a set tell a coherent story?

• 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.