Dossiers

A dossier is a formal PDF report on part of a board. Select the nodes it should be about and DeepState gathers them, every node one link away, and the links among them, and typesets the lot as a document you can hand to someone who has never seen the board.

Where a presentation tells a story in slides, a dossier is the record: every entry in full, every link with its reasoning, every source with its capture date.

Building one

1. Select one or more nodes on the board.
2. Click Dossier in the selection bar above the ask bar, or in the status bar next to Delete.
3. Set the options and click Build PDF, then choose where to save it.

Option	What it does
Title	Leave it blank and the assessment names the dossier; without AI, the first subject's title is used.
Brief	What the reader should come away knowing. It steers the summary and findings.
Prepared by	Shown on the cover. Remembered for next time.
Paper	A4 or US Letter.
Handling marking	Internal, Confidential or Restricted puts a coloured band on the cover and the marking on every page.
Include directly connected nodes	Adds every node one link away from the selection. Regions and RSS feeds are left out unless selected.
AI summary and assessment	Uses your active AI provider. Off, the dossier contains only what is on the board.
Full texts in an appendix	Long notes, articles and extracted document text reproduced unedited at the back.

Building takes seconds without AI and usually well under a minute with it. Stop cancels.

You don't have to wait: Run in background (or closing the dialog) lets the build carry on while you work. A pill in the status bar shows its progress (click it to reopen the dialog), and a card with Open PDF and Show in folder appears when it is done. One dossier builds at a time; it is built from the board as it was when you clicked Build PDF.

What is in it

• Cover with the case, a reference number (DS-yyyymmdd-xxxx), the date, the scope and the handling marking.
• Contents with page numbers and links.
• Summary (AI): an executive summary answering the brief, key findings rated high, medium or low confidence, how the material connects, open questions, and a note on the sources.
• Subjects: each selected node in full: its fields, pictures, text, an AI assessment of why it matters, its verification notes, every connection (with the reasoning recorded on the link), tags, narratives and sources.
• Chronology of every dated entry, subjects highlighted.
• Network: a figure of all the entries and links, and a table of every link with its basis.
• Connected entries, in brief.
• Sources: every distinct URL, with publisher, author, publication and capture dates and the archived copy where one exists, and which entries cite it.
• Appendices: full texts; plates, every picture in the dossier on its own page; and a method note recording when and how the dossier was compiled and which parts were written by which model.

Pictures are embedded at their original resolution. JPEG and PNG files go in unaltered; other formats (WebP, GIF, HEIC…) are converted without scaling, and SVGs stay vector. Only pictures larger than 10,000 pixels on a side are reduced. Wherever a picture appears (a subject's figure, a connected node's thumbnail), clicking it jumps to its plate, which gives the pixel size, format, file size and where it came from; zoom in there for full detail. A dossier with many large photos is correspondingly large.

Entries are numbered as exhibits (E-01, E-02, …). Citations in the AI text such as [E-03] are links to the entry.

How the AI reads the material

Nodes can be very large, a scraped article, a long note, the text of a 300-page PDF, so the model never reads the board wholesale:

1. Short entries (up to about 1,600 characters) are used as they are.
2. Entries that already carry an AI summary (notes and articles the assistant has summarised) reuse it.
3. Long entries are split at paragraph boundaries into chunks of about 12,000 characters. Your summariser model condenses each chunk to its facts, and the notes are merged into one digest of about 200 words. Very long texts are read in samples (start, evenly through, end) rather than whole.
4. The main model then writes the assessment from the digests, the fields and the list of links, a request kept under a fixed size however many entries there are.

Digests are cached by content, so rebuilding a dossier, or building another that shares entries, does not pay for the same text twice.

Every machine-written passage is labelled machine-generated in the PDF, and the method appendix names the models used. Treat them as analysis to check against the exhibits, not as evidence. If the AI step fails, or no provider is set up, the dossier is still built, without the summary and assessments, and the dialog says why.

Privacy

The PDF is rendered on your machine by a sandboxed, offline browser window: pictures are read from the board's media folder (or fetched once if only a URL is stored) and embedded, and the page cannot load anything from the network. Only the digests and the assessment request go to your AI provider, and only when the AI option is on.