A regulatory return was submitted three months ago. A reference-data correction arrives showing that a counterparty classification was wrong for 14 months, which changes figures in that return and in the five before it. What happens next, and what must the pipeline have supported for this to be manageable?

Hour by hour, what happens

The first question asked is not "what is the right number" but "what number did we submit, and from what?" If the pipeline cannot reproduce the exact output of a run from three months ago, the conversation with the regulator starts from an admission rather than from evidence.

The second question is the size of the difference, which requires running the corrected data through the same logic as it stood at each submission date. Two dimensions of time are now in play at once: when a fact was true in the world, and when the system knew it. A pipeline that stores only current values cannot answer either question, because the correction has overwritten the state that produced the original figure.

The third is the decision about whether to restate. That is not an engineering call, but engineering determines whether it can be made on evidence — materiality is assessed against the difference, and the difference cannot be computed without the previous two answers.

Where it amplifies

• Six returns, not one. A classification wrong for 14 months touches every submission in the period, and each restatement needs its own reproduction, difference and explanation.
• Downstream consumers. Internal capital calculations, management reporting and any disclosure that used the same figures now differ from the restated ones unless they are corrected too.
• The explanation is the deliverable. Regulators generally expect not just a corrected figure but an account of the difference — which rows changed, why, and what has been done so it does not recur. That account is assembled from lineage, and lineage that was not captured cannot be reconstructed afterwards.
• Scrutiny compounds. A restatement invites examination of the control environment, and the finding is frequently about the absence of reproducibility rather than the original error.

What the user sees

Nothing externally, if it is handled. Internally, the cost is measured in weeks of senior attention, and in a supervisory relationship that is now more expensive for a year.

What stops it being a crisis

Specific properties, all of which must exist before the correction arrives:

• Bitemporal storage on reference and transaction data: valid time and system time, so "what did we believe on 31 March" is a query.
• Versioned, deployed logic. The calculation code and its configuration tagged per submission, so the same inputs can be run through the same rules rather than through today's.
• Frozen, retained inputs. The exact input snapshot stored alongside the output, because reconstructing inputs from upstream systems reproduces today's view of them, which is the thing that changed.
• Row-level lineage from every reported figure back to its contributing records, so a difference can be explained rather than merely asserted.
• A rehearsed restatement path, exercised at least once deliberately. An amendment process first executed under supervisory attention is the worst possible time to discover it.

What would have to be true for this to self-heal

Only if corrections propagate as events rather than as overwrites. In a pipeline where reference data changes are appended with an effective date and every dependent figure is recomputed and compared automatically, the correction produces a report of affected submissions within hours rather than a project. That design costs more storage and more compute on every ordinary day — typically a 2-4x increase in stored volume for the retained snapshots — and it converts a rare, expensive, manual scramble into a routine, cheap and automatic one, which is the trade worth making precisely because corrections are guaranteed to arrive and impossible to schedule.

Common weak answers

• "We can rebuild the figure from source systems." Rebuilding reproduces today's view of the inputs, which is exactly the thing that changed. It fails silently and confidently.
• "The report is in the warehouse, we can rerun it." Rerunning applies today's logic. A restatement needs the logic as deployed at the submission date, which is why the code version is part of the record.
• "Materiality means we can ignore small differences." Materiality decides whether to restate, and the assessment still requires computing the difference — which is the capability being tested here.