Why we reconcile disagreeing values without a model

Any system that reads more than one document will eventually hold two values for the same thing. A management pack says ARR is €12.4m; the audited accounts say €11.9m. A facility agreement sets a covenant at 3.5x; a side letter amends the definition it is measured on. This is not an edge case — it is the normal condition of a firm's paperwork, and how a system handles it is most of what distinguishes one from another.

The obvious move is to ask a model. It is also the wrong one, for three separate reasons.

Why not just ask a model which value is right?

First, cost. Conflicts are common. Paying for a model call every time two documents disagree turns an ordinary property of a corpus into a recurring bill that scales with how messy your records are — which is precisely backwards, because the messier the records the more the firm needs this.

Second, repeatability. Ask the same question twice and you may get two answers. For a system whose entire pitch is that the record can be examined, an arbitration step that cannot be reproduced is a hole straight through the middle of it.

Third, and most importantly: the answer is usually already determined by facts nobody needs a model to weigh. Audited accounts outrank a management pack on a financial figure. An amendment outranks the agreement it amends. A later filing outranks an earlier draft. These are not judgments — they are rules your profession already has, and they can be written down.

What are the three rules?

In order: tolerance, authority, recency.

Tolerance comes first because most disagreements are not disagreements. Two documents reporting €12.43m and €12.4m are reporting the same number at different precision. Applying a tolerance before anything else stops the system generating conflicts out of rounding, which is the single biggest source of false noise in this kind of work.

Authority comes next, and it is per module rather than global. This is the part people find surprising and it is the part that makes it work: a source's authority depends on what you are asking it about. A CRM record beats a pitch deck on what stage a deal reached, and loses to audited accounts on a revenue figure. One document is not simply “more reliable” than another; it is more reliable about some things.

Recency settles what authority leaves tied. Two documents of equal standing, one from March and one from July, and the July one wins — which is how a restated figure supersedes an original without anyone rekeying anything.

What happens to the value that lost?

It stays. This is the part that matters most and it is the easiest to get wrong.

The record is append-only, so nothing is overwritten. The figure the committee approved on remains readable after the figure is restated, attached to the claim that superseded it. That is not sentimentality about data — it is the difference between a system that answers “what is true now” and one that can also answer “what did we believe when we decided, and on what”. The second question is the one that gets asked in a review, an exit, or a dispute.

Contradiction becomes judgment rather than overwrite. Both values stay attached, and the person reading can see the disagreement that produced the answer.

When is a model used at all?

Only when the rules genuinely run out: two sources at the same authority tier, disagreeing by more than tolerance, with no recency separating them. That is rare, and it is the only case worth spending on — it is a real conflict rather than an artefact of rounding or hierarchy.

The economic shape of this matters for what the product costs. Deterministic reconciliation is free, runs on every value, and is where most of the work happens. Model spend is reserved for the exceptions. A design that arbitrated everything with a model would have a bill proportional to the disorder of the corpus, which is a bill nobody can forecast.

Why is authority per module rather than per source?

Because “which document do you trust” is not a question with one answer, and treating it as one is how systems in this space get their most embarrassing results.

A CRM record is authoritative about what stage a deal reached and close to worthless about a company's audited revenue. A pitch deck is authoritative about what the company claims and not about what it delivered. An information memorandum is authoritative on deal structure and self-interested on financials. Rank sources globally and you must pick one of those to be wrong about.

So authority is declared per module — per kind of thing being asserted — and it arrives with the pack for your discipline rather than being hard-coded. A firm that disagrees with a default can say so, which matters because these rankings are genuinely contested between houses: some credit funds treat a maintained tracker as authoritative over the executed agreement, and they are not wrong, they are running a different process.

What does this buy a firm?

An answer you can interrogate. When the memory reports a covenant level, it can name the clause that sets the test, the certificate that reported against it, the amendment that moved the definition, and why one won. None of that requires trusting a summariser — the reasoning is a rule, and the rule is the same one your own people would have applied.

It also makes disagreement visible rather than resolved out of sight. A credit team that can see headroom is 3.1x on the amended definition and 3.4x on the original is in a different position from one handed a single confident number. The number is the same work; the difference is whether the system told you which one it picked and why.