Investigative journalism · Étude de cas
The relationship map that used to need a dedicated data team, built by two reporters on a laptop
Ce que ça vous donneWith 4,200 leaked scans sitting in a folder on the laptop, ask in one line for "companies registered at a building owned by the procurement director's spouse, and the contracts those companies appear in" — and get an answer that quotes the sentence from the corporate registry.
Four stacks of scanned printouts on the desk. 4,200 pages, handed over on a USB stick by a source. Corporate registry extracts, contracts, board materials, and someone's printed-out email, all mixed together. A reporter is drawing names and arrows on a sheet of A4 in ballpoint. "Hanbit Development — auditor Park — Semyeong E&C CEO Park?" The question mark is there because confirming whether the name she saw three pages back is the same person or just a namesake means digging through that stack again. The desk gave her six weeks. She's in week two.
Two reporters on a local newsroom's investigative desk. What they have: 4,200 scanned pages from a source, 260 public tender notices downloaded from the procuring agency's website, and two laptops. No data team, no server, and the leaked documents can't be uploaded to an outside service, because protecting the source comes first. This piece walks through putting that material into a Consilience workspace: what you ask it, and what shape the answers come back in. It isn't a real customer story — it's a hypothetical scenario.
What's on the desk
The decisive sentence is never in one document. The fact that company A's auditor and company B's CEO are the same person is split across two registry extracts, and the fact that B's registered head office is a building owned by the procurement director's spouse sits somewhere else entirely, in the contract stack. It takes three documents to produce one line of the story.
- 4,200 scanned pages from the source. 180 corporate registry extracts, 340 contracts, board materials, and printed emails, mixed together in no particular order.
- 260 public tender notices from the procuring agency. PDFs off the website, listing the winning company and the amount.
- The same company is written differently in every document. "Hanbit Development Co., Ltd.", "Hanbit Development", "Hanbit Dev." — and some contract signature blocks just say "Hanbit".
- About 40 people whose names recur in the registry's officer sections. Whether two are the same person or namesakes comes down to the birth-date digits alone.
- Two weeks of the reporters' handwritten notes: six A4 relationship diagrams and their own notebooks.
How they do it today
Right now the work happens in a spreadsheet. One reporter pages through the 180 registry extracts and copies out company names, head-office addresses, and officer names; the other pulls the winning companies out of the 260 tender notices into a second sheet. Then they try to match the two sheets on company name. Rows where the spelling drifts never match, so a miss tells you nothing: no connection, or just a different way of writing the name?
The second problem is finding things again. "I've seen this address somewhere" doesn't tell you which of the 4,200 pages it was on — that means digging through the stack again. Everything confirmed by week two lives on six sheets of A4, and every item with a question mark next to it needs the same dig repeated to settle.
There's a precedent worth studying. For the Panama Papers, the International Consortium of Investigative Journalists (ICIJ) loaded 11.5 million documents into a graph database and let reporters in some 80 countries follow names and addresses across it. But that's what you do when you have a dedicated data team and servers. Here there are two reporters and two laptops, and the leaked documents can't go up to an outside service.
How the material goes in
Setup takes half a day. Drop the folder in whole, wait for extraction to finish, then open a case around one driving question.
From 4,200 scans to a case
Move the scan folder into the workspace
Move the folder holding the 4,200 pages into your workspace folder. Don't sort it first. Once untouched documents pile up, the Activity center offers the bulk route with an "Build this workspace's knowledge graph?" card, and extraction mints the people, companies, and addresses in each document as entities. The original scan files stay exactly where they are.
Keep the public material in its own folder
Put the 260 tender notices in a different folder from the leaked documents. Splitting "material we can name in print" from "material we can't" by file location makes the citation calls faster when you're drafting. Public background on the agency and the companies can be gathered separately with
/research(⌘⇧R), and those documents get extracted just like any other note.Ask only after the build finishes
Extracting 4,200 pages takes a while. Send a chat message mid-build and a "Knowledge graph still building" dialog appears with the progress. Pick "Send when extraction finishes". An answer based on half the registry is especially dangerous in this kind of story, because a company with no connections and a company nobody has read yet look identical.
Open a case around one driving question
Ask broad questions in a "New chat" (⌘⇧C), @-mentioning the key names, then right-click that chat in the sidebar session list → "Pin as case…". The toast tells you how many entities got pinned. Into "What is this case about?", type the reporting question as-is: "Did companies registered at a building owned by the procurement director's spouse win this agency's tenders?"
Check the registry to confirm it's the same company
Whether "Hanbit Development Co., Ltd." and "Hanbit Development" landed as one entity is something you check on the entity page. Hovering a connection chip shows which note and which sentence the fact came from, so you read the head-office line in the registry yourself and make the call. A wrong edge dies with "This connection is wrong", and re-extraction can't resurrect it. Neighbors that check out go into the case via "Add to case" on their page.
What you actually ask
"Any company registered at a building held in the name of
@Director Park's spouse? If so, show the contracts
those companies appear in too"
→ Companies and contracts come back as a table. Each row carries a
reference like [[Hanbit Development]], quoted together with the note
and the sentence that states the fact. Below it, the mini graph
"Where this came from in your notes" is drawn.
"Is there a path connecting @Director Park to @Hanbit Development?"
→ The path comes back hop by hop (spouse-owned building →
registered head office → Hanbit Development). Each hop carries the
source sentence that states the connection, so the reporter hovers
the chip, reads the registry sentence, and verifies the hops one by one.
"Put the notices these companies won in chronological order,
and add each company's incorporation date"
→ The notices come back as a table sorted by date, with an
incorporation-date column. Whether the gap between notice date and
incorporation date is suspicious is the reporter's call, from the table.These questions work because extraction mints the people, companies, and addresses across the 4,200 pages as entities and wires the relations between them into a knowledge graph (a map that connects concepts as points and lines). "Companies that share an address" just means companies pointing at the same address entity, so one question catches it without paging through the documents in order. The conclusions and the sentences in the story are the reporter's to write.
What changes
| Before | Now |
|---|---|
| Page through 180 registry extracts, hand-copying company names, head-office addresses, and officers into a spreadsheet | Drop the folder in the workspace, click the "Build this workspace's knowledge graph?" card, and extraction mints the people, companies, and addresses as entities |
| "Hanbit Development Co., Ltd." and "Hanbit Development" stay separate rows and you can't tell if it's one company | Judge from the connections and source sentences on the entity page, and kill wrong edges with "This connection is wrong" |
| Dig through 4,200 pages again to settle "where did I see this address?" | Hover the connection chip and the note and sentence the fact came from are right there |
| Give up on tooling entirely, because the leaked documents can't go to an outside service | The workspace is a folder on the laptop. The 4,200 scans stay in that folder, and nothing gets uploaded to a shared service |
| Reopen it two weeks later and how far you got lives only in each reporter's A4 notes | "Pin as case…" the chat, and on reopen the "New evidence on N pinned entities since your last visit:" strip points at what to catch up on |
What disappears from the two reporters' manual work is the copying-out and the digging-again. Judging which company looks suspicious, reading the registry text to confirm it, and writing the story stay exactly where they were: with the reporters.