figure-table-quality

Figure & Table Quality Audit

Pipeline position: Phase 2.5 (between Grounding/Polish and Submission). Runs after figure-rhetoric (content) and before arxiv-preflight (compliance). See /manuscript-pipeline for full execution order.

Purpose

Verify that every figure and table in a manuscript renders at readable size in the compiled PDF. figure-rhetoric checks whether figures communicate the right message. This skill checks whether the reader can physically read them.

Execution

Step 1 — Build the display-scale map

For every \includegraphics in the .tex source:

1. Extract the display width (e.g., \textwidth, 0.7\textwidth, 0.55\textwidth)
2. Compute the effective display width in inches using the document geometry
3. Read the figure generation script to find the figsize for each figure
4. Compute: scale = display_width / figsize_width

Also extract from the generation script or config:

• Base font sizes: title, label, tick, legend, annotation
• Marker sizes, line widths
• Bar widths (for grouped bar charts)

Step 2 — Per-figure audit (all 9 checks)

For each figure, compute effective values at display scale and check:

2a. Font size at render

effective_font = script_font × scale

Element	Minimum	Warning
Axis title	7pt	8pt
Tick labels	6pt	7pt
Legend text	6pt	7pt
Annotations	6pt	7pt
Panel titles	7pt	8pt

FAIL if any element falls below minimum. WARN if any element falls below warning threshold.

2b. Label collision

Check for overlapping text in:

• X-axis tick labels (especially with rotation < 45° and > 3 labels)
• Y-axis tick labels (long text strings)
• Data annotations near each other
• Legend entries overlapping data

For rotated labels: compute horizontal footprint as len(label) × char_width × cos(rotation). If footprint > tick spacing, FAIL.

2c. Marker and line visibility

effective_marker = script_marker × scale
effective_linewidth = script_linewidth × scale

• Markers below 4pt effective: WARN
• Line widths below 0.5pt effective: WARN

2d. Bar chart readability

For grouped bar charts:

• Compute effective bar width in inches
• Check if bars are distinguishable (minimum 3pt effective width)
• Check if hatch patterns render at effective size
• Check for bar-label alignment

2e. Color and hatch accessibility

• Are all series distinguishable in grayscale?
• Do hatch patterns provide redundant encoding for color?
• Are there more than 5 colors without hatching? WARN
• Are similar colors used for unrelated series?

2f. Axis range efficiency

• Compute data range vs axis range
• If less than 40% of axis range contains data: WARN (wasted space)
• If data touches axis boundary: WARN (clipped data)

2g. Annotation readability

• Do any data annotations overlap each other?
• Are annotations positioned to avoid occluding data?
• Do annotations use consistent formatting (fontsize, weight)?

2h. Legend placement

• Does the legend overlap any data points or bars?
• Is the legend in a consistent position across similar figures?
• For multi-panel figures: is the legend in the first panel only (not repeated)?

2i. Whitespace and margins

• Does tight_layout() or equivalent handle margins?
• Are panel titles cut off?
• Is there excessive whitespace (> 30% of figure area empty)?

Step 3 — Table audit

For each \begin{tabular} or \begin{table}:

1. Column alignment — Are numeric columns right-aligned? Text left-aligned?
2. Rule style — Uses booktabs (\toprule, \midrule, \bottomrule)? No vertical rules?
3. Caption position — Table captions above, figure captions below?
4. Width — Does the table overflow margins? Check for \resizebox or \small hacks.
5. Number formatting — Consistent decimal places? Aligned decimal points?
6. Header clarity — Are column headers unambiguous?

Step 4 — Cross-figure consistency

1. Shared elements — Do figures that share axes use the same scale?
2. Color scheme — Is the same color used for the same condition across all figures?
3. Label vocabulary — Are condition/model labels identical across all figures?
4. Font family — Same font across all figures?

Output format

## Figure & Table Quality Report

### Display Scale Map
| Figure | figsize | display | scale | verdict |
|--------|---------|---------|-------|---------|
| fig1   | 7×4     | 6.27"   | 90%   | OK      |

### Per-Figure Findings
#### Figure 1 (fig1_resolve_rates.pdf)
- [PASS] Font sizes: tick 8.1pt, legend 8.1pt
- [PASS] No label collisions
- [WARN] Bar width 3.2pt — borderline at print size
...

### Table Findings
#### Table 1 (table1_resolve.tex)
- [PASS] Booktabs rules
- [PASS] Caption above tabular
...

### Cross-Figure Consistency
- [PASS] Color scheme consistent
- [FAIL] Label mismatch: fig2 uses "cmd", other figures use "Yuj"

### Summary
- [count] FAIL (must fix)
- [count] WARN (should fix)
- [count] PASS

Auto-Fix Rules

AUTO-FIX (apply directly):

• Rotation increase for overlapping x-labels (20°/15° → 45°)
• figsize reduction to match display context (eliminate >25% downscaling)
• Missing tight_layout() calls
• Inconsistent label text across figures (align to config/source of truth)

HUMAN-REQUIRED (present and wait):

• Figure redesign (different layout, panel arrangement)
• Axis range changes
• Color scheme changes
• Font size increases that affect layout
• Table restructuring

Integration

This skill reads:

• .tex source (\includegraphics directives, \geometry settings)
• Figure generation scripts (figsize, font sizes, rotations, annotations)
• Figure config files (shared settings)
• Generated figure files (visual spot-check if PDF readable)
• Table .tex files

It does NOT:

• Evaluate whether figures communicate the right message (that's figure-rhetoric)
• Check arXiv format compliance (that's arxiv-preflight)
• Audit prose or claims (that's manuscript-review)