Security think tank · 事例
Instead of filling in a competing-hypotheses table by hand, review one that's already filled
これでできることIn a workspace holding 340-odd clippings in seven languages and 12 satellite imagery reports, asking "lay out the conflicting evidence on the Kandas Third Industrial Complex, with sources" returns a list split into supporting and contradicting evidence, every item carrying a [[bracket]] reference, with a "Where this came from in your notes" mini-graph drawn underneath the answer.
Tuesday afternoon, four windows open. On the left, a translated wire story from a Russian agency. In the middle, page 41 of a December satellite imagery report, as a PDF. On the right, notes typed by hand at a seminar back in May. Along the bottom, an empty spreadsheet. That spreadsheet was supposed to become an ACH table — Analysis of Competing Hypotheses, the method Richards Heuer codified as a CIA analyst: hypotheses down one axis, evidence across the other, each cell filled in with supports or contradicts. Three hours in, two rows are filled.
This person is a senior OSINT analyst at a security think tank. Two years of foreign press clippings, 340-odd of them, sit mixed across seven languages, and satellite imagery reports, academic papers, seminar notes, and in-house network reporting are tangled together in Korean and English inside one folder. This piece walks through what you'd ask and what would come back if you turned that folder into a Consilience workspace. It isn't a real customer story — it's a hypothetical scenario.
What's piled up on the desk
Two years of following one facility, all in a single folder. Most filenames are nothing but a date and a source abbreviation, and nowhere does it say which sentence supports which hypothesis.
- 340-odd press clippings — seven languages. A good many are the same original story, translated and re-quoted by outlets in different countries.
- 12 satellite imagery reports — PDFs. Some pages carry the analyst's read-out comments and tables as images.
- 28 academic papers — English. Background on the facility type and its processes.
- 40-odd seminar notes — Korean markdown. Who argued what on the day, written out sentence by sentence.
- 60-odd in-house network reports — Korean and English mixed. The most current judgments live here.
How it works today
By the textbook, you put four hypotheses down one axis, copy each piece of evidence across the other, and fill every cell with supports or contradicts. Which means rereading 340-odd items, recognizing by eye that the same facility goes by different names across seven languages, and hunting the relevant sentence out of a satellite imagery PDF to paste into the table. By hand, that's a day gone. So outside a crisis, the table rarely gets finished.
Without the table, two things happen. First, only the stories that fit the first hypothesis register. That's where one Russian story and the four other-language stories that picked it up get counted as five independent pieces of evidence. Second, your own assessment from two months ago disappears. The judgment written into the May seminar notes and the judgment in the July network report don't match — and checking whether it changed or whether you simply forgot means reopening 40-odd notes.
Reopening 40-odd notes to find a two-month-old judgment and what it rested on takes half a day. So most of the time, nobody checks.
How you set it up
You don't have to move the material anywhere. Drop the folder in as-is, and settle the naming drift up front.
From one folder to a workspace
Drag the whole folder in
Drop the folders of clippings, reports, papers, and notes straight into "Extract sources" (⌘⇧I). Progress shows as five stages — "Reading your files" → "Getting to know your topics" → "Extracting the knowledge graph" → "Connecting related items" → "Organizing into themes". Across the 12 satellite imagery PDFs, the text-heavy pages convert locally, and the pages where read-out comments are images get read by vision OCR.
Paste article links as URLs
For foreign stories where you have no clipping file, only a link, use "Add manually" → "Add URL" and paste them one per line. Your originals are never touched, and the workspace stays as markdown on your own disk.
Teach it the facility's names
In the left rail, open "Domain vocabulary" → "Teach the AI your field" and register the facility's English spelling, its Russian transliteration, and the short form the domestic press uses. Then run "Tune up AI search" from the Lifecycle panel and it walks Domain vocabulary → Re-extraction → Synonym check → Relationship check → Topic map. The facility's other names pile up as merged pairs in the "Synonyms" panel, and if you judge a pair wrong, a toggle splits it.
Pin the question as a case
In the sidebar session list, right-click the chat where you asked about the facility → "Pin as case…". The driving question, the pinned entities, and the investigation log all stay. Reopen it two weeks later and the new-evidence strip tells you which pinned entities picked up evidence while you were gone.
What you ask
"@Kandas Third Industrial Complex — lay out the conflicting evidence, with sources"
→ Comes back as a list split into supporting and contradicting evidence. Every item
carries a [[bracket]] reference that opens straight to the source note or entity
page. Underneath the answer, a "Where this came from in your notes" mini-graph is
drawn with an "N items · M relations" count, and the legend splits "Directly
matched" from "Related".
"Has my assessment of this facility changed at all since May?"
→ Where the May seminar note's judgment and the July network report's judgment
diverge, the "Revisions" panel surfaces them as a "New" (the July sentence) and
"Revises" (the May sentence) pair. Press "Confirm" and the May sentence renders
struck through everywhere in the app, with a "Superseded" badge and a
"Superseded by {note}" link.
"Which note did each piece of the civilian-use case come from?"
→ Every basis arrives with a [[bracket]] reference. Open the references and you can
see for yourself that one Russian story and the four other-language stories that
picked it up are all the same wire report. That's where counting one original as
five independent pieces of evidence gets caught.These answers are possible because extraction splits things down to the sentence. An entity page's "In your words" section collects your sentences verbatim, grouped by kind: "Decisions", "Open questions", "Claims", "Quotes", "Preferences", "Values", and "Stances". Someone else's words are recorded with who said them, and stance and preference sentences get a for, against, or neutral badge. A sentence a later note overturned carries a "Superseded" or "Retracted" badge.
The basis is traceable. Hover a connection chip on an entity page and the source sentence that states the fact is quoted straight out of the note, with a caption in the form "{note title} · line {n}". If the note changed after extraction, it tells you the note may have changed rather than quoting a stale sentence. When your background on the facility type is thin, fill it in with /research (⌘⇧R). One document per source lands in the research/ folder, and those documents get extracted into the graph like any other note.
What changes
| Before | Now |
|---|---|
| Reread 340-odd pieces of evidence and fill the table's cells by hand. Three hours, two rows | Ask for "the conflicting evidence, with sources" and the answer arrives as a list split into supporting and contradicting |
| The same facility is scattered under different spellings across seven languages and counted separately | Register the English spelling, the Russian transliteration, and the domestic short form under "Teach the AI your field" and merged pairs pile up as rows in the "Synonyms" panel — or a toggle splits them |
| Four stories that picked up a Russian original get counted as four independent pieces of evidence | Every basis carries a [[bracket]] reference, and opening it shows the four are the same wire report |
| My judgment from two months ago is somewhere in 40-odd notes | The "Revisions" panel surfaces it as a "New" / "Revises" pair, and pressing "Confirm" puts a "Superseded by {note}" link on the May sentence |
Reviewing the table stays your job. Whether each item is really independent evidence, and whether the source sentence actually supports that hypothesis, is the analyst's call. What gets filled in automatically is the table's cells, not the judgment.