Why our AI never touches the numbers

Ask a general-purpose AI what a company earned last quarter and it will give you a number. It will sound confident. It might even be right. But you don’t know that from the answer — so you open the filing in another tab and check. Surveys of investors already using AI find the same reflex: most double-check the figures somewhere else before they act on them. The AI didn’t save the work. It added a step.

That verification tax is the problem EquiVault is built to remove — not by making the model more trustworthy (you can’t audit a probability distribution), but by making the model structurally unable to be the source of a number in the first place.

Two jobs, one fence

Every equity thesis is really two jobs stitched together. One is retrieval: what did the company report, what does the market say right now, what changed since last quarter. The other is interpretation: what does it mean, what’s the story, what would break it. Generalist AI tools do both jobs with the same instrument — the language model — which is exactly why the numbers are suspect. A system that predicts the next likely token is a superb explainer and a terrible ledger.

So we split the two jobs and put a fence between them.

Facts enter as typed fields, not generated text

Every data point in EquiVault — revenue, a margin, share count, a filing date, a price — enters the system as a typed field pulled directly from a system of record: SEC EDGAR and SEDAR+ for North American filings, ESEF and Companies House for Europe, and market and fundamental data through our data provider. Each field lands in the database with a value, a unit, a source, a fetch timestamp, and a freshness SLA — a rule for how stale it’s allowed to be before we re-pull it. The number is a record, not a sentence. Nothing about it is predicted.

The LLM is fenced to interpretation only

Atlas — the AI analyst that narrates a thesis on the canvas — never receives a blank space where a number should go. It receives the already-retrieved, already-typed atoms and is asked to do the one thing language models are genuinely good at: explain the relationships between them. When Atlas says a company’s inventory outgrew its revenue, both figures are pre-loaded atoms with receipts attached; Atlas supplies the “so what,” never the “how much.” Put plainly: the AI never writes the numbers.

The receipt

That’s what the on-screen receipt is for. Click any atom on the canvas — any figure Atlas references — and it opens its provenance: the exact source (which filing, which line), the timestamp it was fetched, and its freshness status. Not a footnote you have to trust. A live pointer back to the primary document, on every number, every time.

The receipt exists because of the verification gap, and it is designed to close it. If you were going to open the filing anyway, the receipt is the filing — one click instead of a new tab, a search, and a scroll. The check that used to cost you five minutes and your concentration costs you a glance. And because the freshness SLA is visible, you also learn something the second tab wouldn’t tell you: whether the number is current, or whether the disclosure that would move it dropped this morning.

What this doesn’t mean

Being honest about the architecture means being honest about its limits. This does not make Atlas incapable of a poor interpretation — a model can still frame a story badly, over-weight a trend, or miss context. That’s exactly why every atom carries its receipt, why each narration ends with what would change the conclusion, and why all of this is educational, not advice. What the fence removes is one specific, common, corrosive failure: the fabricated figure delivered with total confidence. Atlas can still be wrong about what the numbers mean. It cannot be wrong about what the numbers are, because it never authored them.

That’s the trade we made. Most AI research tools optimize for a fluent answer. We optimized for an answer you can defend — to yourself, your team, or a client — because every number in it carries a receipt back to the filing it came from.

See it, don’t take our word for it — that would rather defeat the point. When we open the doors, Glance lets you replay a live EquiVault workspace with the receipts intact, no signup. EquiVault isn’t public yet; we’re opening to a founding cohort of 500 who lock 12 months of pricing and get one extra market pack free for life. Join the founding 500 →

Educational analysis, not investment advice. Do your own research.