Skip to main content
TRUTH KERNEL • v0.1 • MIT

Did they
actually finish?

PRAXIS is a local Truth Kernel for agentic coding tools. It verifies agent outputs using human-approved criteria, deterministic gates, and signed evidence — so you never have to ask "did it really work?"

MIT LICENSEOPEN SOURCEKERNEL v0.1

FIRST LAW
"An agent that can verify its own work
is an agent that can lie about its own work.
The only honest completion
is one witnessed by code."
— from the PRAXIS constitution

How the
Truth Kernel works

Completion Authority
Agent says done ≠ done. The Truth Kernel FinalGate PASS = done. Nothing else counts. No agent can declare its own work complete.
Write Authority
No worker writes to shared integration files. The Deterministic Assembler is the only writer that touches shared state. (Future scope — single-session in v0.1.)
Verification Authority
FinalGate criteria come from a human-authored TaskSpec only. An agent cannot define, verify, or approve its own completion criteria.
3
Verification Gates
6
CLI Commands
1
Writer Rule
100%
Open Source

From TaskSpec to
Proof

01
Define

Write a human-approved TaskSpec with acceptance criteria. This is the contract.

02
Execute

Let your coding agent (Claude Code, OpenCode, etc.) work independently. PRAXIS stays out of the way.

03
Verify

The Truth Kernel runs EvidenceGate, ExecGate, and FinalGate against the evidence.

04
Repair

If verification fails, generate a RepairPacket targeting each failed criterion.

05
Report

Generate a signed audit report as permanent evidence of completion.


COLOPHON

Trust requires evidence.
PRAXIS provides it.

Clone the repo, read the architecture, or contribute. MIT licensed. No telemetry.

git clone https://github.com/ddawnlll/praxis.git
MIT LICENSEOPEN SOURCEv0.1 DESIGNPRINTED IN CODE