Support Boundaries

This page defines the current public support boundaries for Epsilon. It is not a release certification page. Epsilon is still on the 0.1.0-dev line, so the supported surface is intentionally narrower than the full public export list.

Supported

Area	Current Support
Fitting	Turing/NUTS MCMC for supported `TimeSeriesMMM` and bounded `PanelMMM` models
Data	Config-driven CSV workflows with explicit date, target, media, holiday, control, and panel columns
Pipeline	Stage-based local runs through `run_pipeline`, `pipeline_main`, and `runme.jl`
Time-series validation	Blocked holdout validation through Stage `35`
Panel modelling	One or more declared panel dimensions on a deterministic flattened panel-cell axis
Post-model analysis	Contributions, decomposition, response curves, saturation curves, adstock curves, and marketing metrics from grouped MCMC results
Optimisation	Fixed-budget total-response optimisation; panel optimisation uses historical within-channel panel shares
Plotting	Static CairoMakie-backed plots and stage-local PNG artifacts
Scenario planning	Non-UI comparison tables over current, manual time-series, and solved optimisation scenarios
Calibration	Time-series MCMC lift-test and cost-per-target calibration on the bounded centered-logistic path

Explicitly Unsupported Or Deferred

Variational inference. Epsilon supports MCMC/Turing fitting only.
Dashboard, hosted UI, AI advisor, or background scenario-store workflows.
Panel holdout validation.
Panel calibration.
Free channel-by-panel optimisation.
Panel seasonality, trend, events, and richer controls beyond the currently supported panel surface.
Automatic refitting of every prior-sensitivity scenario.
Arbitrary future spend-path simulation.
Portable binary interchange for Julia .jls artifacts.
Loading serialized artifacts from untrusted sources.

Unsupported paths should fail explicitly rather than silently approximating a different model.

Pipeline Stages

Pipeline runs use fixed stage directories:

Stage	Directory	Notes
`00`	`00_run_metadata/`	Source config, resolved config, data and model metadata
`05`	`05_prior_sensitivity/`	Optional scenario planning, not automatic refitting
`10`	`10_pre_diagnostics/`	Prior predictive artifacts where supported
`20`	`20_model_fit/`	Fitted model and grouped inference artifacts
`30`	`30_model_assessment/`	Observed/fitted, posterior predictive, residual summaries, plots
`35`	`35_holdout_validation/`	Time-series holdout validation only
`40`	`40_decomposition/`	Contributions and decomposition artifacts
`50`	`50_diagnostics/`	Sampler, convergence, predictive, residual, and design diagnostics
`60`	`60_response_curves/`	Response, saturation, adstock, and metric artifacts
`70`	`70_optimisation/`	Optional budget optimisation

When a stage is skipped, Epsilon writes SKIPPED.json in that stage directory and records the marker in run_manifest.json.

Config Guidelines

Use fit.backend: mcmc or omit the backend for the maintained MCMC path.
Keep media spend nonnegative.
Use explicit panel dimensions under dimensions.panel for panel configs.
Keep demo optimisation disabled unless a deliberate budget block is supplied.
Prefer runme.jl ... --quick for local demo checks and increase sampler settings only when inspecting modelling behaviour.

Local Checks

Use focused checks for routine development:

make format-check
make smoke
make smoke-demo-configs

Use the full suite only when a broad local gate is needed:

make test
make docs