analysis-authoring

Follow a fixed section sequence: analysis question, data assets summary, configuration, parameterized loading, structure and quality checks, univariate distributions, then conditional sections for multivariate, temporal, and outlier analysis, ending with insights and next steps. Conditional sections appear only when the data supports them.

Select the figure from the analytical goal: histograms for numeric distributions, bar charts for categorical counts, scatter plots for two-variable relationships, heatmaps for correlations, and line charts for temporal trends. Apply scale thresholds such as sampling above roughly fifty thousand points and grouping high-cardinality categories.

Cache serializable results like loaded frames and aggregates with a data-caching decorator, since they are copied per session. Cache non-serializable resources like database connections with a resource-caching decorator and treat them as read-only. Hold user selections in explicit session state so they persist across pages.

Encode every meaningful distinction in a channel besides colour, such as marker shape, line dash, position, or direct labels. Use colourblind-safe qualitative palettes for categories and perceptually uniform scales for continuous values, and avoid red-to-green ramps.

Default budgets are under roughly three seconds for initial page load, under roughly one second for interaction response, and stable memory over extended sessions. These are defaults that a project-stated budget supersedes, and every budget should be recorded alongside its observed measurement.

Route column semantics and profile contracts to data-catalog, persistence and versioning to dataops, accessibility conformance to the accessibility skill, and trained-model evaluation to ml-experimentation. The conventions cover authoring judgment only, not tool mechanics or pipeline ownership.