Close

Every point is one searchable passage of the engine's evidence, placed by its vector.

How the field was made

1. The engine's index was built on from the acquired sources: public web pages and PDFs cut into passages, dataset descriptions, catalogue observations and the analyst cases. Every passage keeps its source link, file and character range.
2. Each passage was read into 384 numbers by . Passages about similar things end up pointing in similar directions.
3. To draw them, the 384 directions were flattened to two so that passages with small angles between them sit close. Distance on this screen is an approximation of that angle; nothing else.
4. When a question is asked, it is read into the same 384 numbers on the private hosted engine. The nearest passages by angle light up. The number beside each is the cosine of that angle: 1 would be identical wording, 0 unrelated.

What this is not

It is not a map. Where a point sits says nothing about where the thing is. The map stays the place view.

Similarity is not confidence, source authority, demand, commitment or a forecast. A lit passage is a lead to inspect at its source, not an answer.

Numeric dataset cells and raster values are not in the field, only their descriptions. Scanned and dynamic pages are gaps.

On this page

Signed in to the workspace, a question opens an investigation, a hold saves the record's evidence snapshot, set aside and next step are saved as notes, and the investigation comes back when you do. The workspace keeps who did what and when. "Catalogue since asked" compares the catalogue revision now with the one at the time of each question; it does not say what changed, only that review is due.