Cognilateral

A trust control plane for autonomous software delivery.

Coding agents already plan, edit, test, open pull requests, and run on a schedule. Cognilateral admits bounded work, hands execution to a replaceable runtime, verifies the result independently, records evidence and cost, and permits merges only through a fail-closed gate.

What is proven, and what is not

proven One runtime (Claude Code) executes bounded work under the plane, with an evidence chain and a nightly proof artifact.

unproven A second, replaceable runtime (a local-model agent) has not yet run under the same contract.

unproven The fail-closed gate has not yet accumulated 30 consecutive nights of unattended operation.

Try it

The public edge package is cognilateral-trust on PyPI, with source at vorthruna/cognilateral-trust.