HIFDocs
HIF / Documentation / Evidence representation

Data Visualisation

Status: normative foundation and engineering reference

Version: 0.1

This document defines how HIF turns data into inspectable representations for analysis, explanation, monitoring and decision-making. A visualisation is an interface to a data-generating process, not decoration applied to numbers.

1. Objectives

A conforming visualisation MUST enable its intended audience to:

  • identify what is represented and what is not;
  • recover the relevant values, relations, uncertainty and provenance;
  • distinguish observations, estimates, forecasts and targets;
  • perform the intended comparisons without avoidable perceptual distortion;
  • access equivalent meaning without vision or colour discrimination;
  • inspect transformations, filters and definitions;
  • detect stale, missing, suppressed and invalid data;
  • avoid acting on precision or causality that the evidence does not support.

Beauty, novelty, density, engagement and animation are secondary to truthfulness, task fitness and access.

2. Representation contract

Before chart selection, record:

decision or question
audience and data literacy
tasks: locate, identify, compare, rank, estimate, trace, explain, predict
entities and observational unit
measures, units and definitions
data types and admissible operations
sampling and data-generating process
uncertainty and missingness
time, geography and population coverage
transformation and aggregation
required comparisons
consequences of error
delivery media and accessibility modes

The visualisation MUST NOT imply a stronger measurement scale than the data possess. Identifiers are not quantities; ordered categories do not necessarily have equal intervals; a rate without its denominator is incomplete.

3. Data, task and encoding

3.1 Data semantics

Classify each field:

  • nominal category;
  • ordered category;
  • quantitative interval or ratio;
  • temporal instant, duration or sequence;
  • spatial position, region or topology;
  • network relation or hierarchy;
  • text or unstructured content;
  • uncertain distribution, interval or ensemble.

Also classify whether it is observed, derived, imputed, modelled, forecast, target or benchmark. The same numeric storage type can require a different representation because its semantics differ.

3.2 Task first

Common task classes include:

  • retrieve a known value;
  • find an extremum or anomaly;
  • compare two or more values;
  • assess distribution, association or change;
  • trace a path, hierarchy or provenance;
  • monitor a threshold or service state;
  • explain a conclusion;
  • explore hypotheses;
  • decide and act.

One display need not serve every task. A precise table, overview chart and interactive analytical view MAY be coordinated instead of compressed into one universal chart.

3.3 Expressiveness and effectiveness

Mackinlay distinguishes expressiveness—whether a graphical language represents all and only the intended facts—from effectiveness—whether it uses the medium and perceptual capabilities well. HIF applies these as gates:

  1. reject encodings that imply facts absent from the data;
  2. among expressive candidates, select the one that supports the task most accurately and efficiently;
  3. test the selection in the target context.

3.4 Perceptual encoding

Cleveland and McGill's original graphical-perception experiments demonstrated systematic differences in elementary perceptual judgements. Position on a common scale generally supported more accurate quantitative judgement than area, volume or colour saturation in their tested tasks.

Their findings MUST NOT be converted into an invariant chart ranking. Accuracy depends on task, marks, scale, density, expertise, device and interaction. Use the work to form testable preferences:

  • use aligned position for precise comparison where practical;
  • use length from a common baseline for magnitude comparison;
  • use angle, area and volume only when their lower precision is acceptable;
  • reserve hue primarily for categories rather than precise quantity;
  • use lightness or another ordered channel only where its perceived order survives the context;
  • add labels or a table when exact values matter.

Three-dimensional perspective, pictorial area and volume MUST NOT encode ordinary values when they introduce avoidable occlusion or nonlinear apparent size.

4. Chart selection

Analytical needStrong starting representationRequired challenge
Exact lookupTableCan scanning and comparison also succeed?
Category comparisonDot plot or bars on common baselineIs zero meaningful; are labels readable?
Trend over ordered timeLine or aligned small multiplesAre intervals, gaps and uncertainty honest?
DistributionDot/strip, histogram, ECDF, box/violin with contextAre sample size and distributional assumptions visible?
RelationshipScatterplot with appropriate model/intervalAre overlap, confounding and scale addressed?
Part-to-wholeStacked position/length, often table or barsIs the denominator stable and comparison possible?
GeographyMap plus non-spatial comparisonDoes area dominate the measure; are rates normalised?
Network or hierarchyNode-link, adjacency matrix or tree by taskIs topology more important than value comparison?
Status monitoringValues, trend, thresholds and exceptionsIs the threshold justified and freshness visible?

This table supplies hypotheses, not automatic prescriptions. Chart form MUST follow the comparison and data-generating process rather than a gallery name.

5. Scales, axes and coordinates

5.1 Scale contract

Every scale MUST define:

  • domain and range;
  • transformation;
  • units and reference period;
  • zero, baseline or centre;
  • inclusivity of boundaries;
  • treatment outside the domain;
  • ticks, rounding and label precision;
  • missing, infinite and invalid values.

5.2 Linear, logarithmic and transformed scales

A logarithmic scale MAY support multiplicative comparison but:

  • MUST be visibly labelled;
  • MUST NOT contain zero or unsigned negative values without a defined transformation;
  • SHOULD provide interpretable ticks;
  • MUST be explained for audiences who may infer equal additive intervals.

Any normalisation, indexing, smoothing, cumulative transformation, per-capita conversion or inflation adjustment MUST be disclosed near the representation or in directly available methodology.

5.3 Baselines

Bar length ordinarily encodes magnitude from a common zero baseline; truncating that baseline can greatly exaggerate differences and requires a different encoding or conspicuous justification. A line chart may use a non-zero range when the task concerns change within that range, provided the axis and context make the magnitude honest.

Dual axes SHOULD be avoided where independent scaling creates a visual relationship not present in the data. If used, their association, units, transformations and sensitivity MUST be explicit and tested.

5.4 Aspect and projection

Aspect ratio can change apparent slopes and correlation. Map projection changes area, shape, direction or distance. The chosen geometry MUST match the analytical task, and consequential interpretations SHOULD be tested under reasonable alternative aspects or projections.

6. Statistical integrity

6.1 Aggregation

Aggregation can conceal variation, composition change and Simpson's paradox. Record and expose:

  • unit of analysis;
  • aggregation function;
  • denominator and weighting;
  • group definitions;
  • time window and timezone;
  • suppression and privacy rules;
  • sample size and coverage;
  • relevant disaggregation.

Mean, median, rate, percentile and total are not interchangeable. A dashboard metric MUST link to its operational definition.

6.2 Missing and censored data

Missing data MUST NOT silently become zero. Distinguish:

  • not collected;
  • not applicable;
  • suppressed;
  • below detection;
  • delayed;
  • invalid;
  • estimated or imputed.

Visual gaps, symbols, labels and accessible descriptions SHOULD retain these differences. The method MUST explain any imputation and its uncertainty.

6.3 Association and causality

Correlation, temporal sequence and visual alignment do not establish causality. If a causal conclusion is claimed, the representation MUST link to the design, assumptions and analysis supporting identification. Model fit MUST NOT be shown without residual uncertainty, relevant diagnostics and scope.

6.4 Precision

Displayed digits MUST reflect measurement and model precision. Rounding MUST be consistent while retaining values used for computation. Apparent precision created by tooltips, smooth curves or pixel resolution MUST NOT exceed the evidence.

7. Uncertainty

7.1 What to represent

Relevant uncertainty can arise from:

  • sampling;
  • measurement;
  • missingness and imputation;
  • model parameters and specification;
  • forecast scenarios;
  • revision and data latency;
  • classification and linkage;
  • natural variability.

A point estimate MUST NOT be presented as a certain outcome when uncertainty could change interpretation or action.

7.2 Representation

Depending on the task, use:

  • intervals or bands with stated coverage and method;
  • distributions, densities or quantiles;
  • ensembles or hypothetical outcome plots;
  • ranges and scenario sets;
  • sensitivity analysis;
  • calibrated verbal statements linked to numbers;
  • explicit unknown or unquantified uncertainty.

An error bar MUST name what it represents: standard deviation, standard error, confidence interval, credible interval, prediction interval or another quantity. These are not interchangeable.

7.3 Comprehension

Uncertainty displays can be misunderstood. Test whether the audience can:

  • identify the central claim and range;
  • distinguish individual-outcome from parameter uncertainty;
  • compare uncertainty across groups;
  • avoid interpreting an interval as a binary significance test;
  • make the intended decision at the declared threshold.

Do not remove uncertainty merely because the first design was difficult to explain; revise the representation and decision support.

8. Colour

Use the policies in COLOR_AND_THEMING.md. In particular:

  • qualitative, sequential, diverging and cyclic schemes MUST match data type;
  • ordered palettes MUST have an appropriate perceptual order;
  • a diverging centre MUST be analytically meaningful;
  • colour MUST be redundant for consequential distinctions;
  • marks MUST be tested at rendered size and background;
  • missing and selected values need semantic roles;
  • legends MUST preserve order and symbols;
  • external palettes require licence and provenance records.

A corporate palette is not automatically a data palette. Brand colours MAY identify the publisher or highlight a focal series, but MUST NOT override accuracy, discrimination or accessibility.

9. Annotation, language and narrative

Titles SHOULD state the subject, population, measure and period. Subtitles, annotations and captions MAY state an evidence-supported takeaway, but MUST distinguish observation from interpretation.

Every visualisation SHOULD make available:

  • data source and retrieval/revision date;
  • measure definition and unit;
  • filters and coverage;
  • transformations;
  • uncertainty method;
  • author/owner and correction route;
  • accessible data download where lawful.

Narrative order MUST NOT hide material counter-evidence. Annotation must not crop a time range, select a baseline or emphasise one comparison without revealing the decision.

10. Interaction

Interaction MUST add an analytical operation, not conceal essential context. Supported operations MAY include:

  • filter and reset;
  • sort and group;
  • zoom and pan with overview;
  • select and compare;
  • inspect value and provenance;
  • change measure or normalisation;
  • annotate, save and share a reproducible state;
  • download data and methodology.

Requirements:

  • the default view MUST be meaningful;
  • current filters, time, units and selection MUST remain visible;
  • state SHOULD be addressable and shareable;
  • keyboard and non-pointer operation MUST be complete;
  • focus order and announcements MUST match the visual operation;
  • hover-only values are prohibited;
  • zoom MUST NOT silently change aggregation or semantics;
  • loading, partial, stale, empty and error states MUST be explicit;
  • animation MUST be interruptible and respect reduced motion;
  • an action that affects an external decision MUST preserve an audit record.

11. Accessibility and multimodal access

11.1 Layered access

An accessible visualisation SHOULD provide coordinated layers:

  1. title and purpose;
  2. concise summary of principal pattern and exceptions;
  3. structured description of axes, variables and encodings;
  4. keyboard-navigable access to meaningful groups and values;
  5. an accessible table or equivalent data representation;
  6. source data and methodology;
  7. optional sonification or tactile output where it supports the task.

Alternative access MUST communicate the same material facts and uncertainty, not merely the chart type.

11.2 Semantics

Prefer native HTML for headings, controls, tables and disclosures. SVG can carry structured text and semantics but requires tested labelling and reading order. Canvas MUST provide meaningful fallback content and operable alternatives; pixels are not an accessibility tree.

Do not add ARIA roles whose support is assumed but untested. Declare the supported browser/assistive-technology matrix and verify real output.

11.3 Tables

Data tables MUST:

  • identify caption, row/column headers and units;
  • preserve meaningful order;
  • expose missing and suppressed values;
  • support navigation without excessive repetition;
  • avoid visual-only merged structures that destroy header association;
  • match active filters and transformations;
  • remain available without requiring precision pointer use.

Large datasets MAY use pagination, grouping or download, but the summary and current subset MUST remain accessible.

11.4 Sonification

Sonification maps data to non-speech audio variables such as pitch, timing, location or timbre. It MAY reveal trend and pattern or provide an alternative mode, but MUST:

  • define the mapping and reference points;
  • provide controls for play, pause, replay, speed and volume;
  • avoid interfering with screen-reader speech;
  • include equivalent text/table access;
  • not use loudness alone for precise quantity;
  • be tested with blind and low-vision people for the intended task.

Sonification is not an automatic accessibility substitute.

12. Dashboards and monitoring

A dashboard is an operational interface, not a tiled chart collection. It MUST define:

  • decisions and owners;
  • update frequency, last successful refresh and latency;
  • service-level targets and threshold rationale;
  • normal variation and alert policy;
  • drill-through and recovery action;
  • data quality and outage states;
  • role-based access and sensitive-data handling.

Prioritise:

  1. current state and material exceptions;
  2. trend and comparison needed to interpret state;
  3. cause, provenance and affected population;
  4. available action and consequence;
  5. deeper exploration.

The colour red MUST NOT be the sole alert. A single KPI MUST NOT conceal distribution, denominator, uncertainty or an offsetting harm. Display targets separately from forecasts and actuals.

13. Misleading and prohibited practices

Unless a documented analytical need and clear explanation justify them, HIF prohibits:

  • truncated magnitude baselines for length/area comparison;
  • unequal intervals rendered as equal spacing;
  • hidden log or index transformations;
  • 3D perspective and volume for ordinary quantitative comparison;
  • dual-axis scaling chosen to manufacture correlation;
  • cherry-picked dates, groups or denominators;
  • cumulative totals presented as period rates;
  • missing values treated as zero;
  • smoothing that conceals reversals or outliers;
  • area maps of totals where population exposure is the relevant denominator;
  • rainbow order for magnitude without a reliable perceptual order;
  • animation that prevents comparison of prior state;
  • pictograms whose area grows in two dimensions for a one-dimensional value;
  • decorative precision or unlabelled extrapolation;
  • dashboards whose stale or failed data look current;
  • interaction defaults chosen to maximise a preferred conclusion.

Ethical review MUST consider who benefits from a framing, who can be harmed, which groups disappear through aggregation and whether the representation is usable to challenge its own conclusion.

14. Engineering architecture

14.1 Specification object

Each maintained visualisation SHOULD have a machine-readable specification:

VIS-ID and version
question and task
dataset/schema/version
transformations and filters
mark and encoding definitions
scales, axes and legends
annotations
interaction state machine
accessibility structure and summary
theme/token references
responsive transformations
export and print rules
tests and evidence
owner and review date

14.2 Reproducibility

Given the same data, specification, token set and renderer version, output SHOULD be deterministic. Store:

  • source-data snapshot or immutable identifier;
  • query and transformation code;
  • locale, timezone and rounding;
  • library and browser versions;
  • random seed where sampling is used;
  • generated accessible summary version;
  • checksum of export artefacts.

14.3 Responsive design

Responsive visualisation is task-preserving transformation. It MAY:

  • change from side-by-side to vertically aligned small multiples;
  • reduce tick density without hiding endpoints or material events;
  • move legends or use direct labels;
  • replace overview-plus-detail with staged disclosure;
  • offer a table as the primary narrow-screen view.

It MUST NOT merely shrink labels and targets or silently remove a material series.

15. Verification protocol

15.1 Data and statistical tests

  • Schema, units and categorical domains are validated.
  • Aggregates reconcile with a trusted independent computation.
  • Filters, timezone, denominators and missingness are tested.
  • Scale domain and transformation handle boundary values.
  • Uncertainty calculations reproduce the declared method.
  • Tooltips, labels, table and exports agree to declared rounding.
  • Stale, partial, failed and revised data states are exercised.

15.2 Perceptual and accessibility tests

  • Intended comparisons are completed accurately.
  • Chart interpretation and uncertainty comprehension are probed.
  • Keyboard, zoom, reflow, forced colours and reduced motion work.
  • Screen-reader structure, summary, controls and table are verified.
  • Colour-vision diagnostics and non-colour cues are checked.
  • Sonification, if present, is task-tested and controllable.
  • Small screen, print and representative display output are reviewed.

15.3 Adversarial review

Reviewers SHOULD attempt to:

  • reach a plausible but wrong conclusion;
  • find hidden transformations or excluded populations;
  • recreate the metric from source;
  • compare reasonable alternate baselines, aspects and aggregations;
  • identify inaccessible facts;
  • determine whether a stakeholder could tune defaults to favour itself.

Material findings MUST block release or be published with owner, risk and remediation date.

16. Sources

All sources are English-language original research, standards or official specifications. Retrieval date: 30 July 2026.