datavizlisted

# dataviz Turn a concrete claim and its source data into a publication-quality chart that reads like an editorial illustration, not a generic dashboard widget. ## Start with the claim Inspect the real data before choosing a chart. State the sentence the figure must prove, the comparison that matters, and any uncertainty or missing measurements. Never invent values, smooth away inconvenient results, or imply causality the experiment does not establish. Choose the smallest form that carries the claim: | Question | Form | |---|---| | How does a metric change over time or training? | Line chart; add uncertainty bands or error bars when available | | Where is the quality/cost/latency frontier? | Scatter plot; label points directly and explain the favorable corner | | What contributes to a total? | Stacked bars; use a shared scale for comparisons | | How does behavior differ by mode or corpus size? | Small multiples with identical axes | | Is only one comparison important? | Annotated slope, dot, or bar chart instead of a full dashboard | ## Editorial visual language - Put a conclusion-oriented title inside the figure. Add a short orientation such as `top left is best` when direction is not obvious. - Prefer a single dark canvas (`#0f172b`) with slate grid lines and labels. Reserve saturated colors for data. - Use a compact technical or monospace face when it matches the publication; keep numerals tabular. - Use thin grid lines, clear axes, restrained legends, and generous i