Results tables

By Stage 06 the run has machine-readable results under workspace/results/ and an experiment_manifest.json indexing them with schema metadata. The table in the paper must be derived from those files, not retyped. A number that appears in the manuscript and nowhere in workspace/results/ is unsupported, and it is the single easiest thing for a reviewer to catch.

Build the table from the artifacts

1. Read workspace/results/experiment_manifest.json for the result artifacts

and their schemas.

1. Generate the table with a script written to workspace/code/, not by hand.

A regenerable table survives a Stage 06 rerun; a hand-typed one silently goes stale the moment a result changes.

1. Write the generated table to a file the manuscript inputs

(workspace/writing/tables.tex, or an inline markdown table in the report), so the paper and the artifact cannot drift.

What goes in the main table

The main table answers the paper's one claim. Everything else goes to the appendix.

• Rows: the methods being compared, with the proposed method last so the

eye lands on it after the baselines.

• Columns: the metrics the claim is about. Adding metrics the claim does

not concern dilutes it and invites reviewers to find a column where you lose.

• Baselines: the strongest available, not the most convenient. A table

whose baselines are weak reads as a weak result regardless of the margin.

• Ablations: in the main table only if the claim is about the mechanism.

Reporting uncertainty

State what the spread means, every time. 0.74 ± 0.03 is meaningless without knowing whether that is a standard deviation, a standard error, or a confidence interval, and over how many seeds.

• Give n (seeds or folds) in the caption or a column.
• Say which dispersion measure you used. NeurIPS and ICML both ask for this

explicitly — see the venue-checklist skill.

• A single run has no error bar. Report it as a single run and say so; do not

present it in a format that implies replication.

• Bold the best result only if the margin exceeds the spread. Bolding a

within-noise win is the most common quiet overclaim in a results table.

Captions that stand alone

Reviewers read figures and tables before the body. The caption must carry: what is being compared, on what data, with what metric, over how many runs, and what the reader should conclude. "Table 2: Results." fails all five.

Good: "Table 2: Accuracy on the held-out split of BENCHMARK, mean ± standard deviation over 5 seeds. Retrieval recovers 12 points that long-context prompting loses when the relevant evidence is diffuse."

Figures

The same rules apply, plus:

• In markdown output mode, figures must be .png (or another renderable

raster format) under report/images/ and embedded with a report-relative path — ![Caption](images/name.png). A .pdf figure shows the report viewer nothing, and the Stage 07 gate rejects it.

• In LaTeX mode, figures live in workspace/figures/ and are included with a

path relative to main.tex.

• Every figure referenced in prose must exist, and every figure that exists

should be referenced. An unreferenced figure is either dead weight or a missing paragraph.

Before you finish

• Every number in the table traces to a file under workspace/results/.
• The dispersion measure and the number of runs are stated.
• Nothing is bolded that is within noise.
• Each caption is readable with the body text covered up.