figure-integrity

name	figure-integrity
description	Rules for maintaining figure and table data integrity. All visual data must trace to result files — never invented or modified for convenience.
tags	["writing","figures","tables","data-integrity","latex"]

Figure and Table Integrity

Core Rule

Every number displayed in a figure or table must be traceable to a file in results/ or data/. If the source file doesn't exist, the figure should not exist either.

When Creating or Updating Figures

1. Read the source data first. Before writing any plotting script, read the actual result file (results/*.json, results/*.csv) and extract the exact values. Do NOT hardcode numbers from memory or from the paper text.
2. Load data from files, not literals. Plotting scripts must json.load() or pandas.read_csv() from the result file. Acceptable:

   with open("results/exp3_repair_n100.json") as f:
       data = json.load(f)
   drr = data["mean_drr"]
   

   Not acceptable:

   drr = 0.942  # from the paper
   

3. Handle missing data honestly. If a result file doesn't exist or is incomplete:
   • Leave the figure cell/bar/point blank or mark it "N/A"
   • Add a % TODO comment in the LaTeX caption
   • Do NOT invent a plausible value

Where Figures Belong (Initial Placement)

This applies when you first write the paper, not only during page adjustment.

• AI concept figures (PaperBanana / Nano-Banana: overview, method, mechanism schematics) belong in the BODY, next to where the concept is introduced (Introduction or Method) — never in \appendix. They illustrate the core contribution; that is why they were generated.
• Do not create an \appendix solely to hold figures. An appendix is for supporting text (proofs, full tables, prompt templates, hyperparameter grids, extra qualitative samples beyond the main results) — not for the paper's main figures.
• This holds regardless of page count. Even when the body is well under the page limit (e.g. a short paper using 1.6 of 2 allowed pages), a concept figure still goes in the body — being short is not a reason to appendix it.
• A concept figure whose \includegraphics sits after \appendix is a placement bug. Move it into the body.

When the Paper Text Cites a Figure

The flow is always: experiment → result file → figure → paper text. Never the reverse.

• If the paper says "F1=0.891 (Figure 4)" — verify that Figure 4's plotting script reads a file that contains 0.891.
• If the paper text and the figure disagree, the result file is the authority. Fix whichever is wrong.

During Page Adjustment

Figures and tables are protected zones during page compression or expansion:

• You may resize a figure (change width/height in \includegraphics)
• You may move a figure to a different position within the body
• Keep figures in the body. Do NOT move a method / result / concept figure to the appendix to meet the page limit. A figure that illustrates the work earns its place in the main text — banishing it to the appendix to save space defeats the reason it was generated. To save space, condense PROSE or move detailed text/tables to the appendix instead; under "relaxed" page-fitting, tolerating a slightly short/over last page is preferable to appendixing a figure. Only genuinely supplementary figures (extra qualitative examples beyond the main results) may live in the appendix.
• You may change float placement ([t], [!htbp], [p])
• You must NOT change the data shown in the figure
• You must NOT regenerate the figure with different values
• You must NOT remove data points, bars, or table rows to make it smaller

After Any Figure Modification

1. Spot-check values. Pick 2-3 numbers from the figure and verify them against the source file.
2. Check captions. If the figure was moved or its context changed, update the caption to remain accurate.
3. Check references. Ensure all \ref{fig:...} in the text still point to the correct figure.

Common Mistakes to Avoid

• Regenerating a plotting script from scratch using numbers "remembered" from the paper text, when the original script loaded from a result file
• Changing figure data during page compression to make the figure smaller
• Copying a figure caption's numbers into a new version of the figure without reading the actual data
• Creating a figure for an experiment that was blocked or incomplete — an honest "not evaluated" note is better than a fabricated chart