Provenance: Making Every Recommendation Traceable to Its Method

Six months after a major decision, someone will ask how it was made. Sometimes it is an auditor. Sometimes it is a new executive who inherited the consequences. Sometimes it is the board, after the assumption underneath the decision turned out to be wrong.

In most organisations, the honest answer is: we cannot fully reconstruct it. There is a deck. There are minutes. There is a person who remembers roughly. The reasoning that connected the evidence to the conclusion existed, briefly, in a room, and then dispersed.

This is normal, and it is a much larger problem than it appears — not because organisations are careless, but because no part of the standard toolchain is designed to record derivation.

What provenance means here

Provenance is a term from data management and archives: the documented chain of custody from source to artefact. Applied to decision-making, it means being able to answer, for any recommendation or conclusion:

• What is being recommended?
• Which method produced it?
• What inputs went in, in which version?
• Which validated source does that method come from?
• Who decided, and when?

Notice that this is a chain, not a citation. A citation says “see appendix B.” A provenance chain says: this output is the result of applying this specification to these inputs, and here is the record.

Why documents cannot carry it

The reason provenance is rare is not indifference. It is that documents are the wrong substrate.

A slide holds a conclusion. It can hold a source note in eight-point type. It cannot hold the relationship between the conclusion and the input data, because a slide has no concept of an input. When the input changes, the slide does not know. It simply becomes wrong, silently, and stays wrong until someone notices.

Spreadsheets are better — a formula is a derivation — but the derivation is trapped in one file, expressed in cell references, and unintelligible to anyone who did not build it. And the method behind the spreadsheet, the reason those particular calculations constitute the right analysis, is nowhere in the file at all.

This is why “just document your decisions better” fails as an instruction. The organisations that try it end up with a decision log: a list of what was decided, which is the least interesting part of the record.

The chain, concretely

In a system where methods are executable specifications rather than documents, provenance is not an extra feature. It falls out of the architecture.

Link	What it records
Recommendation	What the engine proposes
Method	The UNITE model that produced it
Inputs	Your data, versioned
Source	The validated corpus entry
Decision	Approved by a person — recorded

Each link is a real reference, not a label. Which means the chain can be traversed in both directions: from a recommendation back to its source, and from a source forward to every recommendation that depends on it.

That second direction is the one people underestimate. When a foundational input changes — a market assumption, a regulatory interpretation, a cost baseline — the question is not “what did we decide.” It is “what does this invalidate.” Only a traversable chain answers that.

Provenance is not the same as automation

An important distinction, and one we insist on: recording the chain does not mean the machine decides.

The final link in the chain above is a human decision, explicitly recorded as such. The system produces the recommendation, shows its derivation, and hands it to a person who approves, modifies or rejects it. That approval is part of the record.

This is not a compliance gesture. It is the correct division of labour. Methods can be specified; judgement in a specific organisational context cannot. What can be systematised is making sure the person exercising judgement can see exactly what they are judging and where it came from.

A system of record for how the decision was made — including the part where a person made it.

Where this becomes non-negotiable

For some organisations provenance is a nice-to-have. For others it is the precondition for using any system at all.

Regulated industries have the clearest case: if a decision affecting capital allocation, risk posture or customer treatment cannot be reconstructed, it cannot be defended to a supervisor. But the requirement shows up well outside regulated contexts — in post-merger integration, where two organisations’ decisions must be reconciled; in any multi-year programme that will outlive its original leadership; and in any organisation where the same class of decision is made repeatedly and inconsistently.

There is also a quieter benefit. When derivation is visible, disagreement becomes productive. Two people arguing about a conclusion are usually arguing about an input or a method without knowing it. Make the chain visible and the argument moves to where it belongs.

What to test

If you want to know whether your organisation has provenance today, pick a significant decision made twelve months ago and try to reconstruct it from records alone — no interviews.

Most organisations cannot get past step two. That is not a failure of diligence. It is a missing capability, and it is the kind of capability that has to be built into how the work is done rather than added afterwards.