LiminalOpen the live product →

Provenance without surveillance

The judgment layerfor AI agents.

Local-first infrastructure that keeps a human in command of the work they delegate to AI, and keeps a signed, verifiable record of how each decision got made.

In active build · invite-only pilot

DILIGENCEAuditorJUDGMENTSkepticSYNTHESISWeaverOUTREACHEnvoy
sealed · packet in the vault
sealed
the record stays inspectable

How it works.

READ

Bounded agents read the delegated work and disagree.

CORRECT

You correct them.

SEAL

The vault keeps an owned, tamper-evident record of the call.

The weekly agent-spend review.

For the agent-team lead and the operator accountable for what agents did and spent.

setup · baseline · context hardening · agent deliberation · ratified decision · provenance · executive report

One cadence, one signed record, defensible upward.

Open it where it runs

These are the live prototype surfaces, not mockups. Choreography runs on deterministic fixtures; receipts are labeled where earned; vault writes are real and device-local.

What’s real here

Who it’s forTeams accountable for AI-assisted decisions: founders first, then owners of frontier-model spend.
The problemAgents produce answers. Nobody governs what was trusted, corrected, or refused. Nobody can prove it later.
The mechanicBounded agents read; they disagree or refuse; the accountable human decides; the record seals and stays inspectable.
The statusPrototype choreography over deterministic fixtures. Receipts labeled where earned. Vault writes are real, device-local. Roadmap is marked.
1st place, Infrastructure track · Algorand hackathon, Berlin (Jun 2026)top 16 of 102 finalists (200 teams at start) · NatSec hackathon, Cerebral Valley x Army xTech x Shield Capital (May 2026)architecture PPA filed · 64/080,639liminal-engine · MIT-licensed public repo

Building with design partners now.

A close ring of operators, onboarded one at a time.