Chart & Statistics Integrity Gate

A pre-publish quality gate that makes your agent a hostile reviewer of every chart, table, and statistic before a human sees it.

What this skill does

Most agents present data agreeably. They describe a chart, quote a number, and move on. A skeptical data reviewer does the opposite: they assume the visual is misleading and the statistic is overstated, then try to prove it. This skill installs that posture as a final pass. The agent stops being the analyst and becomes an adversarial reviewer of its own data presentation.

The output is not a redrawn chart or a recomputed number. It is a structured verdict: the weakest visual or claim, every misleading element, every missing piece of context, and a clear decision — PASS, REVISE, or FAIL.

When to use it

Run the gate as the last step before delivering anything that shows or cites data: dashboards, report figures, chart descriptions, data summaries, KPIs, or any statistic that will be quoted or acted on. It is most valuable for confident, polished-looking visuals, because that is exactly where distortion hides.

The five review passes

1. Axis & scale check. Flag truncated or non-zero baselines, dual axes, inconsistent intervals, and log scales presented as linear.
2. Cherry-pick scan. Catch suspicious date ranges, dropped outliers, missing denominators, and selective comparisons that flatter one side.
3. Chart-type fit. Check that the chosen visual matches the data (e.g., pie charts used for non-parts-of-whole, line charts implying continuity across categories).
4. Statistical-soundness audit. Surface correlation stated as causation, small-sample or no-error claims, misleading averages, and percentages without base rates.
5. Hostile-question rehearsal. Generate the three toughest questions a skeptical analyst would ask, and check whether the presentation already answers them.

The verdict format

A compact, consistent block: the weakest visual/claim, each issue found (with a suggested fix), missing context, the three toughest analyst questions, and a final decision — PASS, REVISE, or FAIL — with a one-line justification.

Why it works

It separates the analyst role from the reviewer role. The same model is far more critical when explicitly told to argue against its own figures and to score them on adversarial criteria rather than on whether the chart looks clean. The structured passes stop the review from collapsing back into approval.

What it is not

This is a reasoning-and-prompting skill, not a charting tool or a statistics engine. It cannot run calculations, render charts, or access data. It surfaces misleading presentation and unsound inference in the data the agent already holds — it does not certify that the underlying numbers are correct. Pair it with an evidence or claim-grounding skill when factual verification is also required.