Matrix

A Matrix is a comparison grid you build yourself. Companies go across the top, questions go down the side, and every cell is answered for that company from that company's own documents. A row therefore reads as a like-for-like comparison rather than a set of unrelated summaries.

Building one

1. AI > Matrix > New Matrix. Name it after the comparison you are running — “Payments peers, Q2” beats “Matrix 3”.
2. + Symbol opens the global symbol search. Pick from the results rather than typing a ticker: a ticker can belong to more than one listing, and picking binds the column to the exact instrument you meant.
3. + Documents chooses what every column reads. See Source documents below — this is the part worth understanding.
4. + Analysis writes a question. It is asked once per company.
5. Run. Nothing is sent to a model until you press it.

The toolbar runs in the order you use it: Symbol and Documents set up what there is to read, Analysis and Data say what to pull out of it, Run does the work, Chat interrogates the result.

Source documents

You choose documents once, for the whole matrix — but every column is a different company, so “the same document” cannot mean the same file. It also cannot mean the same fiscal period: NVIDIA's “Q2 2027” earnings call was held in August 2026, so lining companies up by fiscal label would quietly compare unrelated quarters.

So a selection is a recency rule: “each company's most recent earnings call”, “each company's two most recent”. It is then resolved per company, and what each one actually resolved to is shown in the Source Documents row and previewed in the picker before you save.

MVP1 offers earnings call transcripts. Filings (10-K, 10-Q, 8-K, DEF 14A) are built and will appear here once they can be read at full length rather than from the first few pages.

Analysis rows

An analysis row is a field definition, not a chat message. Write what you want extracted and in what units; the model is already told to answer only from the documents, to state figures exactly as reported, and to say Not found in documents rather than guess.

Improve with AI rewrites a rough line into a proper field definition and strips out anything that would break the row — company names, tickers, formatting instructions. Output length (Quick, Bullet points, Detailed) controls the shape of the answer.

Data and financial rows

Not every row needs a model. + Data adds a row read straight from our own database: it fills instantly, costs nothing, and can never go stale.

Running, and what it costs

Every analysis cell is a separate AI call against that company's documents, so running is always something you ask for and never something that happens because you edited a prompt.

• Run does the cells that need it — empty, failed, or stale.
• The caret beside it offers Run whole matrix, which re-asks everything including cells that are already up to date.
• The play icon on a row header runs that row across every company; the one on a column header runs that company down every row.
• Hovering a cell gives you Run / Re-run for that cell alone.

Every one of these tells you how many cells it is about to run before it runs them, and cells run a few at a time so a slow document never stalls the rest of the grid. Stop halts the queue.

Stale cells

A cell remembers the prompt, the output length, the document set and the model that produced it. Change any of those and the cell is stale: the answer stays on screen, and the run bar tells you how many are affected. Nothing is re-run and nothing is charged until you decide it is worth it.

Reading a cell

Click any cell to open it in the side panel: the full answer, the question it came from, when it was answered, and the documents it was read from — each of which opens in the same panel.

Chat

Chat answers questions across the grid — “which of these grew fastest, and by how much” — using the answers already in it. It reads what is there; it never triggers a run, so one question costs one call however large the matrix is.

Every claim carries a citation chip naming the company and the row it came from. Click one and the cell it cites opens beneath the conversation, so the claim and its evidence are on screen together. Each message can be copied or deleted, and long threads load earlier messages on demand.

Sharing and working together

Share in the matrix menu grants access to a team or to named people, at view or edit level. Sharing is what decides who can open a matrix — publishing does not: Published versus Draft is a label you set to tell collaborators whether the contents are settled.

While others have the same matrix open you will see them: an avatar for each person in the toolbar, and a coloured outline with their name on whichever cell they are looking at. Cells appear as they are produced, wherever they were run from, so two people do not pay twice for the same column.

Limits

Analysis cells draw on your AI credit allowance; data and financial rows do not. Running out stops the queue rather than failing every remaining cell one at a time.

Known limits

• Documents are truncated to fit the model. Very long documents are read from the front. Where that matters the cell says so.
• The fundamentals feed can lag the transcript feed. A financial row may show an older period than the earnings call in the same column. The period label under every financial value is how you spot it.
• There is no live market capitalisation. The financial one is as of the period end it names.
• Answers are extraction, not judgement. A cell reports what a document says. It does not check whether the company is right.