Somebody asked a question the original scope never anticipated. The answer was already in the estate.

Its scope is a contract estate that was read once, for another reason, usually a migration that asked for a specific set of fields. Its power comes from a decision taken before anyone framed a use case: read the document for everything it contains, not for the fields the current asker specified.

That decision costs more per document and looks like waste on the day you pay for it. It stops looking like waste the first time a department shows up that was not in the original scope. Serving the second department costs close to nothing. The documents have already been read.

An entitlement is not a number in a contract. It is the distance between what a contract says and what a system did. So the query needs two sides: the estate, where the terms live, and a reference to what was actually applied. Give it only the contracts and you get a beautiful list of clauses and no information.

None of these decisions is taken to find anything. They are taken earlier, for other reasons. Money Found is not a new capability. It is the release of inventory that already exists, which is why it takes an afternoon rather than a project.

The capture side works off a catalogue of money relevant concepts. It knows the difference between a price adjustment clause, a price index reference, an escalation cap, a construction cost contribution and an early payment discount. Five different questions, five different answers. A system that treats them as one field is not doing the work.

Every concept carries one of two tags. It is money: the captured value is itself an amount. Or it points at money: it enables a check downstream without being an amount. Pointers never sum into a headline number. The easiest way to produce an impressive and dishonest figure in this category is to add a pointer to an amount.

Matching is deterministic. No language model in it. Case insensitive, diacritics folded, word boundaries anchored, so a three letter trigger cannot fire inside a longer word.

What comes back is rows, each carrying the clause it came from and the page that clause is on. A formula, its parameters, its base value, its index factors, and the amount a system applied, side by side for the first time. Deliberately unexciting.

Three kinds of question an estate can usually answer once it has been read this way:

An agent that produces evidence hands its work to the department that owns the answer. They compare the computed figures line by line against their own systems. Until they sign off, a result is a method and a shape, not a finding.

The incumbent for this class of question is a one off manual study by an outside firm. You pay once, you get an answer about the estate as it was that quarter, and the answer decays. A structured estate makes it a standing query. Ask on the day you want the answer, and again the day after. The second department to ask pays close to nothing.

Nobody else structures contract estates this way. Designing the method your team uses to prove a field is correct is part of the work, not a formality at the end.

Every read is written down with the question that caused it. Every row carries the clause and the page. The result is not a claim to be trusted. It is a list that can be walked. When a row disagrees with the system of record, one of the two can show you the page and the line it was read from.

A bare value with a page number is not provenance. A claim about what a contract entitles somebody to is worth nothing without the clause under it, and a great deal with it.