reportlisted

# report — measure it and write it up Answer a question **with numbers you can defend**, then hand back a single HTML file anyone can open. The fifth sibling of **rewrite**, **refactor**, **reskin** and **audit**. Those four act on code; this one acts on *data*. Same house discipline — evidence over vibes, stated limits, no pretending to be more certain than the measurement allows — but the artefact is a report rather than a diff. Domain-agnostic. Two outputs: **(1)** a verified set of findings, **(2)** a shareable single-file HTML report. The report goes to a mixed audience — someone with no background in the subject and the engineer who has to reproduce it read the same document, so it is written in three layers. The two failure modes it exists to prevent: - **The confounded result** — a real difference measured, attributed to the wrong cause. Condition B looks 20% better because it ran on twice the sample, not because it is better. - **The trusted label** — a column called `duration_ms` that holds seconds, a file named `run-30fps-high` recorded at a different setting. The name is a hint; only the content is evidence. --- ## Invariants - **Work read-only.** Never modify the source data or the runtime code of the project under study. Temporary scripts and intermediates go to the scratchpad directory. - **Verify the input; do not trust the name (MUST).** Open a field, column, flag or file and confirm what it actually carries. A field called `roll` may not be r