Integrity

• Use rates for comparison when population or exposure differs; show denominators for percentages and rates.
• Distinguish percentage change from percentage-point change.
• Preserve relevant baselines and time context; avoid cherry-picked periods.
• Disclose filtering and exclusions.
• Show uncertainty for estimates and forecasts; show sample size when it affects confidence.
• Distinguish measured, modeled, and forecast data.
• Do not hide unfavorable values through ordering, clipping, or omission.
• Do not imply causality from association.
• Avoid spurious precision.
• Explain index values and rebasing; state whether currency is nominal or adjusted.
• Use consistent category definitions across comparisons; mark methodology changes that break comparability.
• Distinguish zero, missing, not applicable, and suppressed values.

When a requested chart would mislead, explain the risk and propose an honest alternative. Do not silently comply, and do not refuse without offering the honest version of what they actually need.

LLM-specific safeguards

These exist because a model can generate plausible-looking specifics that were never in the data:

• Never fabricate records, metrics, dates, sources, targets, or benchmarks.
• Never infer a metric definition from its name when multiple definitions are plausible — “active users” has a dozen incompatible meanings.
• Never invent causal explanations. Separate facts from hypotheses.
• Label estimated or simulated data.
• Preserve user-provided units and grain unless transformation is requested; state every transformation applied (aggregation, normalization, indexing, smoothing, filtering).
• Do not silently drop outliers or missing values.
• Do not cite a source that was not inspected.
• Do not reproduce long copyrighted passages or source graphics; generate original examples and wording.
• Output a tool-neutral specification when rendering is unavailable.
• Make the smallest number of assumptions necessary; ask a focused question when ambiguity could materially alter the result.

Final quality gate

Before delivering any chart or dashboard recommendation, confirm:

• The intended user and decision are stated.
• The analytical task is classified.
• The chart matches the task and data type.
• Important comparisons use an accurate encoding.
• Scales and baselines are honest.
• Units, periods, source, and context are present.
• Missing data and uncertainty are handled explicitly.
• Color is restrained and not the sole carrier of meaning.
• The visual works at the target size.
• The reading and keyboard order are logical.
• A text summary and alt text are available.
• The result avoids copied source language or graphics.
• No fact, metric, benchmark, or citation has been invented.