Reproducibility Artifacts¶

This page defines the minimal, authoritative artifact set required to reproduce any evaluation, selection, or policy decision produced under the Electric Barometer framework.

Reproducibility in Electric Barometer is not aspirational. It is a hard requirement enforced through explicit artifacts, fixed assumptions, and deterministic workflows.

Why reproducibility is treated as an artifact¶

Operational readiness decisions often outlive: - the people who made them, - the models that generated forecasts, - and the systems that executed them.

As a result, Electric Barometer treats reproducibility not as a logging concern, but as a deliberate output of every workflow stage. If a result cannot be reproduced exactly from declared artifacts, it is considered incomplete.

Reproducibility principle¶

A result is reproducible if—and only if—the following question can be answered unambiguously:

Given the same declared inputs, parameters, and versions, would this workflow produce the same diagnostics, selections, and policy decisions?

If the answer is “not guaranteed,” required artifacts are missing.

Required artifact categories¶

Reproducibility requires artifacts from each stage of the workflow pipeline.

1. Evaluation context artifact¶

Defines what was evaluated.

Must include: - entity definitions and identifiers, - evaluation resolution and horizon, - aggregation semantics, - inclusion / exclusion rules, - and the evaluation window.

This artifact establishes the coordinate system for all downstream reasoning.

2. Input data snapshot¶

Defines what was observed.

Must include: - realized demand values aligned to the evaluation frame, - explicit representation of missing or zero-demand intervals, - and a versioned reference to the data source.

Downstream artifacts may reference this snapshot but may not alter it.

3. Forecast candidate manifest¶

Defines what was evaluated.

Must include: - an explicit enumeration of forecast candidates, - identifiers or hashes sufficient to retrieve each forecast, - confirmation of structural alignment to the evaluation frame.

Forecasts not listed in this manifest are considered out of scope.

4. Evaluation parameter declaration¶

Defines how behavior was interpreted.

Must include: - asymmetric cost ratios, - tolerance values or grids, - snapping or representation requirements, - and any fixed thresholds used during evaluation.

All parameters are treated as governed assumptions.

5. Diagnostic output bundle¶

Defines what was observed diagnostically.

Must include: - readiness diagnostics for each forecast candidate, - evaluated under authoritative unit semantics, - with references to diagnostic definitions and versions.

Diagnostics without parameter and unit context are invalid.

6. Selection rationale (if applicable)¶

Defines why a forecast was chosen.

Must include: - the selection objective, - any secondary criteria or tie-breakers, - and the dominance logic applied.

This artifact explains selection without requiring re-evaluation.

7. Governance decision artifact (if applicable)¶

Defines what actions are allowed.

Must include: - the authoritative governance decision, - declared unit system and tolerance semantics, - readiness adjustment allowability, - and explicit reasoning statements.

This artifact is binding and terminates interpretation.

Determinism and versioning¶

All reproducibility artifacts must satisfy:

Determinism Identical artifacts must produce identical outcomes.
Explicit versioning Diagnostic definitions, thresholds, and policies must be versioned or hashable.
No hidden state No workflow may rely on implicit defaults, learned state, or external context.

What is intentionally excluded¶

Reproducibility artifacts do not require: - model training code, - feature definitions, - hyperparameters, - or optimization internals.

Electric Barometer reproduces decisions, not model training processes.

Governance notes¶

Reproducibility is enforced by governance.

If required artifacts are missing, incomplete, or inconsistent: - evaluation results are non-authoritative, - selection outcomes are invalid, - and policy decisions must fail explicitly.

Silent degradation is not permitted.

Summary¶

Reproducibility artifacts ensure that Electric Barometer workflows are: - reviewable, - auditable, - transferable across teams and systems, - and defensible over time.

They are not optional documentation—they are part of the system’s output.