Trust, made formal. Every AI output gets co-signed.

AI drafts the work; humans verify it. Lornian gates every agent action by the risk of the tool it wants to use (declared in code, enforced the same way whichever surface asked), so “almost right” never ships.

Trust, before and after Co-Sign

Without Co-Sign

AI output ships on faith, not verification

"Almost right" bugs surface in production

No record of who approved what, or why

Trust in AI output stalls at 29%

With Co-Sign

Every draft arrives with its confidence score and rationale

High-risk tools pause the run and wait for a signature

Medium-risk writes record their own undo before executing

Every co-sign, score, and checkpoint is auditable

Three things the Agent will never do alone

Risk is a property of the tool, declared where it is registered. It is not a number the model reports about its own work.

Co-Sign HITL Governance · Risk-Tiered Verification
Awaiting Signer
agent_action: publish_knowledge_doc
Target artifact ADR-042 Token Encryption Protocol is locked against mutations via SetCoSignChecker while review #891 is pending.
Alex Rivera (Principal Eng)
✓ Verified ED25519
Sam Chen (Staff Architect)
⏳ Pending Approval
  1. Publish a document

    The doc locks the moment review opens. It cannot be published or rejected out from under the reviewer — not by another agent, not by another person.

  2. Ship a deployment

    The Agent can build the case for a deploy: the diff, the failing checks, the last change-failure rate. It cannot press the button.

  3. Message your client

    Nothing reaches a client portal without a name attached to it. The draft is the Agent’s; the decision is yours.

  4. Everything else runs

    Reads execute immediately. Writes execute with an undo already recorded. Gating only where it earns its cost — a review queue nobody reads protects nothing.

A paused run, in the four ways it can go

  1. You approve it Tuesday

    The parked run lives in Postgres, not in memory. Restart the server, approve three days later, and it resumes at the exact step it stopped on.

  2. 48h

    You never look

    The request times out and the run closes cleanly. No stalled agent sitting on a lock on your knowledge base.

  3. You reject it

    The step records that it was skipped in review and the run carries on without it. A rejection is data, not a crash.

  4. Two of you approve at once

    The decision is claimed compare-and-swap. Two people hitting approve in the same second produce one execution, not two.

Bring verified trust to your team

Every AI output, checkpoint-gated and auditable. Join the waitlist, or book a walkthrough of the ritual.