New · Sumex · proof of computation

A runtime that proves what it computed.

Most systems give you an answer. Sumex gives you an answer plus a checkable record of how it was produced, whether it reproduces, and who vouches for it — for the moment someone has to trust a number they didn't compute themselves.

Content-addressed, reproducible, cryptographically signed. The record travels with the result.

What it does

The record travels with the result

Give Sumex the shape of a computation. It compiles the intent, executes it under the constraints you set, and returns a result that can defend itself.

Content-addressed

The graph and its inputs have a fingerprint. The same computation always has the same one; a changed computation never does. Tamper-evident by construction.

Honest trust tiers

A result is Computed, Reproduced, and/or Signed — independent and composable. Signed is not reproduced; reproduced is not proven. The label never overstates.

Enforced limits

You declare time, memory, and size limits. The runtime refuses to exceed them and fails clearly — oversized work is rejected before it allocates, not after it crashes.

It can find the computation

Give it inputs, a target, and constraints, and it searches for a computation that satisfies them — then proves each candidate by running it. It proposes; the engine verifies.

Who it's for

For results that cross a boundary

Sumex earns its place when a computed result is handed to someone who has to trust it without redoing the work. When you just need a fast number, an ordinary script wins.

Model validation & internal audit starting here

A valuation or risk figure produced by one team, relied on by risk, compliance, and eventually an auditor. Reproducing last quarter's number becomes automatic instead of a manual reconstruction.

Computational research starting here

Results that feed papers, collaborators, and review. A reviewer verifies the artifact instead of trusting the author — and you can reproduce the figure from months ago.

Trustworthy compute for AI agents available

An agent decides what to compute and calls Sumex over MCP; the result comes back with its trust status attached. Sumex is the hands that compute verifiably — not a competitor to AI, the tool it reaches for when the number has to be right.

Cross-party & regulated handoffs the horizon

Lab to regulator, subcontractor to liable prime, model to auditor. “Trust me” becomes “check it yourself”: a signed artifact a third party confirms independently.

Where it is, stated plainly

Sumex is an early, working v0.1 runtime. The trust foundation — identity, reproducibility, resource safety, cryptographic attestation, bounded synthesis — is built and continuously verified on a clean machine.

Sumex is early and evolving in the open. Agents can call it over MCP today; long-running jobs and published pricing are still ahead, shaped with the first pilot users. This is an invitation to build with us, not a finished product to buy.

Pricing will be usage-based — a credit-based pilot tier for early partners, set with you rather than handed down.

Pilot it with us.

Early access is for people who feel the reproducibility problem directly and want to run real work against it. Tell us what you compute and who has to trust it.