A conviction number for
every position.

Sheaf doesn't predict the market. When you've got several reads on a trade — your own, a quant model, the analyst notes, the macro print — it measures how much they actually cohere, turns it into a conviction number, and flags the part of your thesis that's standing on disagreement rather than consensus.

It doesn't tell you what to trade. It tells you how much to trust the view.

See it live on the markets desk → How it stays in your walls →

A five-discipline desk that scores exactly where your analysts agree — and where they split.

See the whole argument — not just the verdict.

you ask Trim my NVDA position before earnings?

FundamentalPriced for perfection — the multiple leaves no room for an in-line print.

TechnicalDisagree on timing — options skew says the downside is already hedged.

RiskAsymmetric either way: trim half, keep the upside, cap the tail.

·measuring consistency …

sheafThe desk splits on timing but agrees on the asymmetry — conviction is in the shape of the trade, not the entry. Trim 50% before the print, ride the rest with a stop.

What it gives a desk

∑

A conviction meter, not a crystal ball

When your inputs genuinely cohere, that's conviction — size up. When they irreducibly disagree, that's ambiguity — size down or pass. Put a number on the thing you've only ever been able to feel.

⚠

Disagreement is an early warning

When diverse, independent inputs suddenly stop cohering, that's often a regime change before the price shows it. Sheaf surfaces the fracture in your view while it's still cheap to act on.

⊞

The whole argument, on the record

Every discipline's case, objection, and revision is visible — and the output is a timestamped record of where your inputs agreed and split. The audit trail you'd otherwise reconstruct after the fact.

A five-discipline desk, in one call

Sheaf runs your question past a panel that argues like a real desk — each discipline pressing its own case — then measures exactly where they converge and where they fracture. Proven on the hardest version of the problem first.

FundamentalMacroQuantTechnicalRisk

Your data stays on your desk

Built to run inside your walls

Sheaf is designed to run in your environment, calling your model endpoints (Azure OpenAI, Bedrock, or your self-hosted models) with your keys. The coherence math runs locally on the outputs — so your positions and research never leave your perimeter, and we're never in the data path.

Your environmentRuns in your VPC or on-prem. Nothing phones home; it can run fully air-gapped.

Your models, your keysProvider-agnostic. Point it at the endpoints you already trust; keys stay in your vault.

Not in the data pathThe measurement is math on outputs you already generated. We never see your data.

For desks with strict requirements, we scope a private pilot to fit your environment. Start the conversation →

Why this isn't a similarity score

The hard part isn't asking models to agree — it's measuring it correctly. A naive "do they agree" score looks right and then fails silently on cyclic disagreement: three views where every pair agrees but the trio is inconsistent, invisible to any pairwise statistic. Sheaf computes the actual obstruction — real sheaf cohomology, with a whitepaper behind it.

See the consistency check → Read the whitepaper →

Run your own question through the desk

No signup. Watch the panel argue, and read the conviction the math returns.

See it live on the markets desk → Talk to us

Building on it? Put the desk behind one API call — api.sheaf.one

Many reads. One number you can size.

A conviction number forevery position.