National research institute · Case study
Pull 20 years of conclusions out of 1,800 reports — what changed, when, and to what
What it gets youAsk 1,800 reports spanning 20 years "how has our institute's conclusion on road-noise reduction standards changed?" and the decisions come back as a list in date order, with revised conclusions struck through, wearing a "Superseded" badge and cited as "{report} · line {n}".
The National Assembly committee's request for materials landed on Tuesday. "Submit your research findings on road-noise reduction over the past ten years, along with how they were reflected in policy." It's due Friday. The report folder the fellow opens on the institute's file server holds 1,800 PDFs stacked up since 2006. The folders are split by year, but inside them the filenames are project numbers and nothing else. A similar request came last year, and the senior fellow who wrote that response left for another job three years ago.
This person is a fifteen-year research fellow in a national research institute's policy division. What's on hand: 1,800 report PDFs spanning 20 years, sets of project proposals and final reports, memos about policy uptake scattered across individual staff folders, and handover folders left behind by researchers who've since moved on. This piece follows what it looks like to put that material into Consilience and ask "how has our institute's conclusion changed?" It isn't a real customer case — it's a hypothetical scenario of how this material would play out.
Twenty years are stacked up on the file server
The material is all there on the file server. The only thing written down nowhere is which report revised which conclusion. Here's the inventory.
- 1,800 research report PDFs. Twenty years' worth since 2006, 60 to 300 pages each. Anything before 2010 is a scan, so the letters are sitting inside images.
- Sets of project proposals and final reports. Forty to sixty a year, for twenty years. They're grouped by project number and not by topic.
- Memos on policy uptake. Draft committee responses, ministry replies, and records of what made it into enforcement decrees, scattered across several staff members' personal folders.
- Four handover folders from researchers who left. Filenames like "final_final_v3" are mixed in, with nothing marking which one is current.
- Department names and project titles changed several times over twenty years, so the same study sometimes survives twice under two different names.
Here's how it works today
When a committee request arrives, the division head emails it out to six research fellows. Each of them digs up the reports they personally remember, writes two or three paragraphs of summary, and sends it back; a staffer stitches the replies together into a response. This procedure only works while someone who remembers the project is still at the institute.
That a 2011 report's conclusion had been overturned by a 2016 report was known only to the person who ran that project, and he left. Reviewing new proposals runs on the same condition. If nobody in the room remembers the similar project from five years ago, the duplicate gets approved.
The same question got one answer last year and a different one this year. That's what happens when everyone only checks as far as the reports they handled themselves.
Here's how you set it up
Put the material into the workspace, teach it the institute's own terminology once, and rebuild the knowledge graph (a map that connects concepts as points and lines). The original PDFs stay right where they are in the folder you point at, and the app reads a copy.
Getting 1,800 reports into a state you can question
Drop the report folders in whole
Sidebar → "Extract sources" (Cmd+Shift+I), and drag the entire year-by-year report folders in. The five stages run ("Reading your files" → "Getting to know your topics" → "Extracting the knowledge graph" → "Connecting related items" → "Organizing into themes"). Pre-2010 scans get read through vision OCR (the processing that turns letters inside an image into text).
Teach it the institute's terminology
The vocabulary it pulls out on its own is only a draft. In the left rail, open "Domain vocabulary" → "Teach the AI your field", fix up this institute's terms — project, final report, policy uptake — and Save. Then use the Extraction panel's "Re-extract everything" so the 1,800 reports you already loaded get read again under the new vocabulary.
Add the uptake memos and the handover folders
Put the policy-uptake memos and the four handover folders into the same workspace. Make each note state its date. The order in which a later report revised an earlier conclusion is decided by the date written in the note, never the file's modified time. Undated notes never enter the ordering at all.
Reconcile the same study's different names
The "Connecting related items" stage proposes merges for the same subject split apart by changing department names and project titles. These proposals don't apply on their own — a person confirms them in the left rail's "Synonyms".
Leave the investigation behind as a note
Once the response is written, open the left rail's "Topic map" and press "Summary note" on the relevant theme. That theme's notes and concepts collect into a single note. When the next request comes in, this is the note you open first.
Here's how you ask
"How has our institute's conclusion on road-noise reduction standards changed?"
→ Five decisions as a list in date order. The 2011 conclusion is struck through,
with a "Superseded" badge and a "Superseded by Road Traffic Noise Reduction
Measures (2016)" link. Each line cites its source:
"Road Traffic Noise Reduction Measures (2016) · line 87".
"Are there earlier studies that overlap with this new project proposal?"
→ A list of six reports. Each row shows the overlapping concepts and a link
to the report's note. Two of them, from 2019 and 2022, are flagged as using
the same survey sites and the same measurement method.
"Show me, in date order, how this conclusion was taken up in policy."
→ Four uptake memos in date order. The ministry's reply is quoted word for word,
and clicking the note link opens that memo.The conclusions that were scattered across 1,800 reports collect on the entity page under "In your words", grouped by kind: decisions, claims, quotes. Hover a connection chip and the original sentence that stated the fact is cited as "{note} · line {n}". Twenty years of reports word things every which way, but the sentence you're shown is the sentence that report wrote.
The places where a conclusion changed are caught by the left rail's "Revisions" panel. When a later report revises an earlier conclusion, it proposes the pair as NEW / REVISES, and pressing "Confirm" marks the earlier conclusion superseded across the app. High-confidence pairs apply on their own as "Applied", always with an "Undo". Only decisions take part in this; claims and questions never supersede anything.
What changes
| Before | Now |
|---|---|
| Six research fellows each replied with as much as they personally remembered | "In your words" on the entity page collects the conclusions pulled from all 1,800 reports in one place |
| Only the person who ran the project knew when the 2011 conclusion had been overturned | The "Revisions" panel proposes NEW / REVISES pairs, and once confirmed the earlier conclusion carries a "Superseded" badge and a "Superseded by {note}" link |
| Whether a new project duplicated an old one depended on someone in the review meeting remembering | Load the proposal, ask which earlier studies overlap, and the report titles come back with the cited line attached |
| When the person who wrote the response left, the investigation walked out with them | "Summary note" in the "Topic map" leaves the reports and concepts you worked through as a single note, and after a long time away the "While you were away" card shows what got added in the meantime |
The 1,800 original PDFs stay right where they are on the file server. The app reads a copy to build markdown notes and a knowledge graph, and stores the revision marks a person confirmed against those.