Master Auditor, a check that takes nothing on trust

What it is for

Master Auditor is the tool we use to check our own websites, web apps and Android apps before they are released. It looks for what is broken, tries to break what seems to work, and gives a verdict with the proof for each part of it written down.

Nothing here takes its word for anything

An AI assistant can sound confident about a piece of software without having checked anything at all, and we built Master Auditor so that confidence alone counts for nothing. Results are recorded by programs that measure, and a check cannot be marked as working unless the right kind of proof stands behind it.

How an audit runs

An audit starts with a checklist made for the project in front of it, because a plain website and an app with accounts need different checks, and any check that is skipped is written down with the reason. Plain programs with no AI then run the code checks, the browser checks and the Android checks.

A final step, which we call the gate, looks at what the checks found, gives the verdict and writes the reports.

It cannot change what it checks

An audit physically cannot change the project it is checking, so it can never tidy up a problem in the course of reporting on it. When we want something fixed, a separate skill does that work on its own branch, and only when it is asked to.

How we test the auditor

On 18 September 2026 we ran it on a project we had broken on purpose, and 32 of 32 checks reached a real result, finding 14 genuine problems, including a password left in a shipped file. Every security hole found while reviewing the auditor itself was fixed and turned into a test written as the attack.

As of 18 October 2026, 99 of 99 tests pass, and in our evals it finds 9 of 9 planted faults with no false alarms. Its first run on one of our own apps found 3 real problems and 5 false alarms, and we fixed each false alarm with a test that pairs the harmless case with the real one.

There is one limit we know of, which is that when the auditor is pointed at its own code, it flags its own lists of warning patterns.

Where it is going

Master Auditor is still being built, and it has so far been run on one real project of ours. It is there to help with the checking before a release, rather than to replace a person reviewing the work.