Show your work: extracted values and their sources

Ask a general assistant what the renewal notice period is and you get a number, immediately, in a confident sentence. Some of the time it is right. The trouble is that you cannot tell which time this is without opening the contract — which was the work you were trying to avoid. An answer you have to re-derive is not an answer. It is a draft you now have to check, and checking a draft against a hundred-page agreement costs roughly what reading the agreement cost in the first place.

The receipt is the product

The extracted value is the cheap part. Any competent model will pull a date off a page. What decides whether that date is usable is whether you can get back to where it came from without leaving your seat: which document, which page, which line. Where a reliable source match exists, reviewers can inspect the available source evidence beside the extracted value — the field on one side, the page on the other, one click apart.

That changes what a review is. A discrepancy stops being a search and becomes a line you open. When the rent roll says the expansion commenced July 1 and the first amendment says August 1, you are not hunting through two documents to work out who is wrong. Both values are in front of you, each attached to the page that produced it, and the only thing left is the judgment about which one controls — which was always the part that needed you.

It also changes what happens six months later. A record whose values carry their provenance survives the person who prepared it. The successor does not have to trust the spreadsheet; they can open it. That is the difference between a data set and an audit trail, and it costs nothing extra at the moment of extraction but is impossible to reconstruct afterwards.

What happens when there is no reliable match

Not every value can be traced, and pretending otherwise is where this goes wrong. Source evidence is not available for every value or every file. A scan goes soft in the gutter. A figure appears in three places on the page and none of them is the one that governs. A term is established across two clauses by implication rather than stated in one. Some documents arrive as photographs of paper taken at an angle in a stairwell.

In those cases the value stays uncited. It does not get paired with a nearby page that looks close enough. Alembic will not manufacture a source relationship to make a record look clean. This is the single most tempting shortcut in the category, because a screen with a citation on every row looks finished and a screen with three uncited fields looks like the software struggled. The second screen is the honest one, and it is the one that saves you.

An uncited field is information, not an error. It is the record telling you exactly where to spend the attention you have. Three uncited fields out of forty is a five-minute job with a known shape. Forty fields that all look equally confirmed is an afternoon of spot-checking that you will not actually do, which means you will trust all forty because you checked six.

Why there is no number on it

We do not print a confidence score, and that is a considered choice rather than an omission. A percentage invites you to build a policy on it — anything above this threshold goes through, anything below gets a look — and the threshold is a fiction. The number is not calibrated against your documents, it does not mean the same thing on an invoice as on a lease, and nobody can tell you what happened on the wrong side of it last quarter.

Categorical states do the job the number was pretending to do. A value is prepared, open, or decided by someone with a name. A conflict is a conflict. A missing document is missing. Those are states a person can act on, argue with, and explain to an auditor. A decimal is not.

What it looks like on the page

The three packets below are the configured examples published on the industry pages. They are built from synthetic documents rather than customer data, and each one is walked through end to end, including the part where something does not add up.

In accounts payable, the line extensions total 7,440 while the invoice states 8,060, and the receiving report records five units against the six billed. Both the subtotal line and the quantity-received line open beside the fields that disagree with them. The controller reads two pages, not two documents, and then makes the accounting and payment call — which never left them. That workflow is described in full on the accounting and AP page.

In a vendor packet, the MSA requires sixty days of renewal notice and the order form says thirty. Both clauses sit side by side with the extracted terms above them. Nothing gets reconciled behind your back, and nothing is chosen for you; legal decides how the documents relate. The four-lane version of that review is on the SaaS contracts and vendor operations page.

In a pay-application packet, the conditional waiver covers a different amount than the payment due, and the required inspection record never arrived at all. The waiver amount opens to its through-date line. The inspection record opens to nothing, because it does not exist, and the record says so instead of inferring that the requirement was satisfied. That is on the construction page.

The standard this sets

A tool that shows its work is agreeing to be checked. That is uncomfortable to build and it is the entire point. It means every value we hand you is a claim we are prepared to have opened, and the ones we cannot support are labeled as such rather than dressed up to match the ones we can.

The alternative — confident prose with no page behind it — is faster to produce and quietly expensive to use. You can only find out which of those you bought by opening a value and seeing whether anything is there.

Each industry workflow walks through a labeled example built from synthetic documents, including the part where something does not add up and a person has to decide. The longer argument for why this matters is on Why Alembic.