result-table

# 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. 2. 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. 3. 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...