NEW CAPABILITY 05 / 06
Document ingestion with provenance
Turn a stack of documents into facts, each one carrying where it came from.
A worked example on fictional documents, produced to show the shape of the output.
A fact without a source is a rumour. This shows an ingestion run over five documents: what was extracted, from which page, by what method, who approved it, what is still held back, and what to do when two sources disagree.
TRY IT / RUNS IN YOUR BROWSER
Use it here.
Ingestion with provenance
Every fact knows where it came from, and who signed it.
Five documents from the fictional Apex Instruments were read in one pass. Open any fact to see the file, the page, the section, how it was read, how sure the reader was, and the role that approved it. This runs on this page, on the data files beside it.
Loading the run
| ID | Claim | Value | Method | Conf | Status |
|---|
No fact matches that search.
Open a fact to see its provenance
Pick any row. The panel shows the source document, the page and section it sat on, the extraction method, the confidence, and the full status history.
What this proves: a fact with no approved source never reaches a document. Three of these facts were never signed and two are contested, so a sheet built from this ledger today would carry no battery figure, no partner discount, and no priority uplift, and it would say so rather than guess.
WHAT IT PRODUCES
The artifacts.
THE MECHANISM / SHOW THE SHAPE, NOT THE SOURCE
How it works.
- 01Receive
Each source is logged with its supplier, date, and a content fingerprint.
- 02Extract
Facts are pulled with a recorded method and a locator down to the page and section.
- 03Score
Each fact gets a confidence, and anything uncertain stays pending.
- 04Resolve
When two sources disagree the fact is marked contested, both values are kept, and a written rule plus a person decides.
- 05Approve
Only approved facts become available to the rest of the system.
WHAT IT READS
- Source documents with dates and suppliers
- Extraction rules for tables and headings
- An approver and a review queue
THE RULES IT HOLDS ITSELF TO
- Every fact names its document, page, and section
- Nothing pending is available for use
- A contested value is never silently chosen
- A superseded fact keeps a link to what replaced it
This is a new capability, built for this portfolio. It is not part of a client engagement, and nothing here is connected to a company’s internal system. The sample company, Apex Instruments, does not exist.