We build one thing, and we build it to be checked

01

An agent that cannot explain itself cannot be trusted with production code

so

Seven layers of reasoning trace, recorded per run and attached to the pull request. Not a confidence percentage, but the actual chain: which files were ranked and why, what the reviewer objected to, which rule the scanner tripped.

02

One model reviewing its own output is not a review

so

The pipeline is split across specialised agents, and the one that reviews the diff is not the one that wrote it. It arrives at the change with no memory of having produced it, which is the only arrangement where the review means anything.

03

Source code is a liability to hold, so we do not hold it

so

Runs execute in a workspace created for that run and destroyed after it. Nothing is written to a database, a cache or a backup. What persists is metadata and an audit trail, which is the minimum required to answer for a decision later.

04

A margin you cannot see is a margin you cannot trust

so

In a self-hosted deployment you hold the provider account, so inference is billed to you directly at whatever rate you have with them. We are not in the middle of it and take no cut. The product records what each run consumed so the bill is checkable against something.

05

Autonomy without a stop condition is just a faster way to be wrong

so

Confidence is scored before anything is applied, and two of the three outcomes end with a person rather than a commit. Nothing force pushes, nothing reaches your default branch, and a blocked change takes an explicit acknowledgement to override.