What It Is

The product is a research workspace built around a canvas of cards. A card can be a person, an event, a source, a claim, a task, a note, a group, or any content type the user invents. Cards can be dragged, nested, searched, assigned, discussed, and connected with typed relationships like supports, contradicts, cites, depends on, or flows to.

The important decision is that a card is not just a visual shape with a label. Every card is a real content record with fields, relationships, permissions, and history. That one choice is what makes the rest of the system work.

• Drop in a URL and the system fetches, evaluates, and stores it. If the page is a paywall, bot wall, empty shell, or failed fetch, the card says so instead of quietly storing garbage. Blocked pages can fall back through the Wayback Machine.
• Drop in a Word document, PDF, or spreadsheet and it becomes structure. Documents split along their own headings into a linked tree. Spreadsheets become row-level cards, grouped and searchable. Scanned PDFs can be OCR'd by a vision model and clearly marked as machine-read.
• Ask the assistant about the board and it sees both the cards and the connections between them. It has explicit verbs for the actions it can take: create, link, unlink, remove, trace, break apart, summarize, and compare.
• Work together live with presence, selections, per-card assignments, locks, team chat, typing indicators, reactions, team invites, and read-only guest links.

The Rule: Absence Is an Answer

Nearly every hard bug we fixed had the same shape: a missing capability produced a confident invention instead of an honest gap.

When asked what was inside a PDF the system had not actually read, the assistant answered with an invoice number lifted from the filename. When asked how two people were related, it created a family link, then read its own fabrication back as fact. The fix was not a better prompt. The fix was to give the model the missing read operation, then make absence loud.

Now the system can trace relationships through the stored graph and answer, "No connection is recorded," instead of filling the blank. Files carry explicit verdicts such as ok, scanned, and service_down, so a card that was never read looks unread.

The same rule shapes the analysis layer. Claims cluster by meaning so users can see where sources agree. Sources that share a wire service collapse into one origin, because ten outlets running the same copy are not ten independent confirmations. Story outlines refuse to assert what the reporting has not established.

The Stack

The stack is boring where it can be and specific where the product needs it.

Two Decisions That Did the Heavy Lifting

Cards Are Content, Not Shapes

Because every card is an Entry of a ContentType, and content types are user-definable, the canvas was general before we deliberately made it general. Adding databases, services, queues, and other node types required an icon, a color, and a shape. The graph, AI, search, permissions, and sharing model already knew what to do with them.

The Canvas Sends Operations, Not Documents

The canvas used to save the entire graph behind a version guard. That made the unit of concurrency the whole board: when two people dragged different cards, one person's work could silently overwrite the other's.

Now the client sends what the user actually did: move this card, draw this line, update this field, remove this relationship. Those operations merge cleanly, and they are exactly the messages that need to be broadcast to everyone else. Live collaboration came from the same architecture rather than being bolted onto it later.

What It Is Not

It will never be draw.io. There are no freeform arrows, no arbitrary swimlanes, and no shapes that mean nothing. Every node is a record. That constraint is what lets users trace it, search it, assign it, discuss it, and ask an AI about it without asking the AI to guess what it is seeing.

That trade, structure over freedom, is the product. The goal is not to make a prettier whiteboard. The goal is to build a canvas where evidence has weight, uncertainty has a place, and the assistant is allowed to say, clearly and usefully, "we do not know that yet."