Design a pipeline producing periodic regulatory submissions where every figure must be defensible years later.

The property that drives everything

A figure submitted three years ago must be reproducible today, from the data as it was then. Not approximately — exactly, including the corrections and the reference data in force at the time.

What that requires

• Immutable source snapshots per reporting period, retained for the full obligation, so a rerun reads what the original read rather than what the source holds now.
• Versioned transformation logic, with the version recorded against each submission. A rerun must execute the code that produced the original.
• Bitemporal reference data, so rates, classifications and rules are applied as they stood at the reporting date, with corrections tracked separately from real-world validity.
• Deterministic transformations. Any dependence on wall-clock time or on a mutable lookup's current state produces a different result rather than a reproducible one, and that difference cannot be explained to a regulator.
• End-to-end lineage from the submitted figure to its source records, which is what a challenge actually demands.
• Reconciliation controls with evidence retained, since the control operating is what is examined, not merely its existence.

Restatement handling

Corrections happen and must be first-class. A restated figure needs the original preserved, the reason recorded, and the difference explicable. A pipeline that overwrites the previous submission destroys the evidence the restatement will be judged on.

The governance piece

Sign-off recorded against a specific artefact version, so the approval refers to something identifiable rather than to a process. And change control on the pipeline with evidence of testing, because the pipeline itself is part of the control environment and its changes are in scope.

The practical failure

Ad-hoc adjustments applied manually to the output. Every reporting programme accumulates them, and each one breaks reproducibility invisibly. They need to be inputs to the pipeline — recorded, versioned, approved — not edits to its output.