---

name: dataviz-precision description: Choose how many significant digits to show for chart and table numbers - keyed to the spread (max-min), not individual values, one uniform place per column. Never fabricate precision. metadata: short-description: Choose significant digits / rounding for chart and table numbers claude-description: Choose how many significant digits to show for chart and table numbers - keyed to the spread (max-min), not individual values, one uniform place per column. Never fabricate precision.

Dataviz Precision

Own the significant-digits decision for numbers on a chart or in a table: axis ticks, data labels, and table cells. Do not choose the chart's form, its colours, or its prose - those belong to dataviz-selector, dataviz-color, and chart-explainer. This skill generalises the precision philosophy that also governs karthik-table-style; when formatting numbers anywhere, this is the authority.

The mechanical computation lives in the dataviz MCP: recommend_precision. Without the server, apply the formula below by hand.

The rule: precision is keyed to the spread, not the value

The question is never "how precise is this number" but "how precise must the reader be to tell these numbers apart". So precision is set by the range of the column or axis (max minus min), not by any single value.

Call recommend_precision(values, role) with the numbers that will be shown together (role = axis, label, or table_column). It computes:

• the range max - min over the values that must be distinguished,
• a single uniform rounding place for the whole set, from the formula place = floor(log10(range)) - (target_steps - 1) where target_steps (~2) is how many significant figures of the range the reader needs to just about resolve the information,
• a formatted preview of every value rounded to that place.

Example: {12483, 9210, 15040} has a range of ~5830, so the place is hundreds and the column shows 12,500 / 9,200 / 15,000 - two significant figures of the spread, and no more.

If you know the smallest difference that actually matters (d), pass smallest_meaningful_difference=d; the place is then floor(log10(d)).

Each preview row also carries a compact form: the same place, written in the largest short-scale unit (K, M, B, T) the column supports. When the column is already scaled - "$MM", "in thousands" - pass unit_multiplier (1e6, 1e3) so the unit comes out right: [70398, 77264] in $MM reads 70.4B / 77.3B. Numbers written in words - the title, a subtitle, an annotation - use the compact form; axis ticks and table cells use whichever the column reads best in.

What this means in practice

• Significant digits, not decimal places. The place cuts to the left of the decimal too: 12,483 becomes 12,500 or 12,000 when the spread is coarse. Large, widely spread values usually want fewer digits, not more.
• Round every value in a column to the same place, so length reads as magnitude and decimal points align.
• Don't show precision the data cannot support, and never manufacture it to fill space - precision is a data decision, not a layout one.
• The source's own precision is a hard ceiling the spread rule can lower but never raise. Integer source values stay integers (44, 1, not 44.0, 1.0 - a trailing .0 asserts a tenths measurement never made). Take the finer of the two: the spread's place, capped at the source's decimal places. Still one uniform place per column - never 36 beside 43.0.
• Never round toward rounder-sounding numbers - round to the place the spread dictates, not to whatever looks tidy.
• No approximation glyphs on a value that is known exactly. A plotted datum you hold precisely is labelled as itself - 610, not ≈610 or ~610 or about 610. The ≈ asserts an uncertainty the data doesn't have, and it also breaks the redundant-axis check (a tick reader matches label text to the mark value, and ≈610 no longer reads as 610), so the value axis it should have retired stays on the chart. Reserve ≈/~ for a figure that is genuinely an estimate (a rounded external headline, a read-off from a source chart), never for a value from the data in hand.
• A displayed 0 must mean the value is zero. When a value far smaller than the spread would round away at the column place, recommend_precision refines the place just enough to keep it one significant digit and returns zero_collapse_prevented: true. It never lets a nonzero value print as 0.

Precedence: the spread rule is the default; exact digits are the exception

By default, recommend_precision governs every displayed number. Source digits override it in exactly one case: identifiers or a genuine exact-lookup requirement - an account number, a code, a reference value the reader must read off verbatim. Call recommend_precision(values, role, exact=True) for those; it preserves every source digit and returns exact_override: true.

Printed source labels, integer measurements and the existence of a data label are not exact-lookup requirements. Identify the reader's task that needs verbatim digits; a magnitude/change comparison uses spread precision even when the source prints every digit. Keep underlying data exact and round only its display.

An exact override is never silent: whenever you leave the spread rule behind, record the reason - why this column is an identifier or exact lookup rather than a quantity to compare. If you can't name why, the spread rule stands. If an upstream decision already marked a column as identifier/exact-lookup, obey that mark and carry its reason forward - don't silently re-decide at build. Carrying the decision as an explicit flag with its reason (not re-inferred from prose) is what lets a weaker downstream model apply it reliably.

Labels are not measurements: temporal and ordinal axes

A value that names a position rather than a quantity to compare - a year (2000, not 2,000), a quarter, month number, rank, stage, or any sequence/ID used as a coordinate - is a label, not a measurement: no thousands separator, no spread rounding, no forced decimals (a year axis reads 1970 2000 2030, never 1,970 or 2000.0). This isn't the exact-lookup override (that preserves a quantity's digits); a coordinate is simply never a quantity the spread rule governs. recommend_precision is for the measured values plotted against these axes, not the axis positions. Same for such a value inside an annotation ("in 2000", not "in 2,000").

Charts and tables

The rule is the same for both. On a chart it sets axis-tick and data-label formatting; in a table it sets each numeric column's rounding and, with karthik-table-style, the decimal alignment and tabular figures. Apply recommend_precision per axis and per column - each has its own spread.