What did this decision rest on?
The question arrives eighteen months later, from an auditor, about a decision whose author has moved on. The answer has to be a document rather than a reconstruction — and the reconstruction is what normally happens.
The problem
Assurance functions are asked to justify a decision after the fact, from artifacts that were never designed to be read together. The spreadsheet is a version behind, the assumptions are in a thread, the approval is in a ticketing system, and the person who made the call has a different job now.
What comes out is a narrative assembled to fit the conclusion, and everyone in the room knows it. The alternative is not a better narrative; it is a decision that was recorded at the time in a form that can be replayed.
This is also where most governance tooling stops short: it records that a model existed, not what it said, which assumptions it rested on, and whether any of that has changed since.
What the product does about it
Four capabilities, ordered by how an audit actually proceeds — from the pack the reviewer is handed, backwards to the individual numbers.
- 01
One decision, self-contained and fingerprinted
The evidence pack for an intervention returns the decision, the analysis it was approved against, the exposures, and the reconciliation — assembled, self-contained, and fingerprinted so the same pack can be shown to have not changed since. A reviewer does not need access to the system that produced it to read it.
What this does not establish: A fingerprint proves the pack has not been altered; it does not prove the contents were right. It is a statement about integrity, and presenting it as a statement about correctness is the most available misreading of that field.
- 02
The decisions taken, in order, with their analyses
An intervention carries its approvals in sequence: which option, approved against which stored analysis, by whom, when. Because the analysis is stored rather than referenced by name, an approval remains checkable after the model has moved on.
What this does not establish: The record holds what was approved and against what. It does not record why the approver chose that option, and no schema will — that remains a question for a person.
- 03
Who changed which number, and when
Every probability row that was overridden by hand is recorded separately from the rows that were fitted or elicited, and the network keeps a revision history. So a question about a specific number resolves to an event rather than to a diff between two files somebody still has.
What this does not establish: Revision history covers the model. Changes to the surrounding decision — the value model, the capacity assumption, the target definition — are recorded as part of the intervention, not as revisions to it, and that boundary is a design decision worth knowing about.
- 04
Replay: does the answer still hold?
Re-ask a stored question and the answer is classified: identical, drifted, or reidentified — meaning the same question now identifies differently, which is not the same failure as a number moving. Each classification carries what actually moved between the two runs.
What this does not establish: Replay covers the effect. Refutations, decisions, segments and targets are stored and re-runnable but not yet all classified the same way, so a replay of those returns the current answer rather than a statement about what changed.
When not to buy this
If the requirement is audit over a model that was fitted from data by another team, this records the decision layer and not the training pipeline — you would be adding a second record rather than replacing one. And if the decision was made without any explicit model, the honest first step is to decide whether it is worth giving one, which is a different conversation.
What the record keeps
The complete chain from the question asked to the outcome measured, in a form that survives the departure of everyone involved.
- The question as asked, not just a fingerprint of it — added deliberately, because a digest cannot be replayed
- The structure as it stood, with each probability's provenance and any manual override
- Every approval in sequence, each against the stored analysis it was decided on
- The replay classification when the question is asked again, with what moved
Bring the decision you have to defend
The most useful session shows the reconstruction you would have to produce today, and how much of it is already recorded rather than remembered.