Security · How we know

A control nobody has tried to break is a hope.

An agent uses a model to reason and a harness to act. The harness is the agent program that calls a model and acts on the answers with its own tools. A seat is the confined environment in which the harness runs. A planted fault is a deliberate defect introduced to check whether a test or reviewer notices it. A digest, or fingerprint, is a value calculated from content to check that the content has not changed.

Every page in this section says what a design is meant to do. This page is about how we find out whether it does.

Status recorded 6 October 2026. What the four status labels mean.

What we do

Each heading states the design or practice. Its status shows whether it is in use, built, in build or designed; a built part does not establish that the whole control is running.

  • One builds, another reviews, a third confirms.

    No piece of the system is accepted on its builder’s word. A second agent reads it and tries it; a third confirms the review was real.

    • In use
  • Reviewed by a different mind.

    The builder and its reviewer run on models from different makers. We use this difference to challenge shared blind spots and treat model diversity as a security property, not only a preference. It does not guarantee that reviewers will catch a defect, or mean that all three agents use different makers.

    • In use
  • We test the tests.

    Before a test counts, a fault is planted in the code it is supposed to guard. If the test does not notice, the test is not evidence.

    • In use
  • A review can fail on the tests alone.

    One piece of the Awakener was first refused by its reviewer because 26 of 36 planted faults went unnoticed by its tests. The tests were strengthened until all 36 were caught, and only then was the work accepted.

    • In use
  • The failures are written down first.

    The Awakener’s specification lists ways it must not fail. Each becomes a test, and none counts as passed until its planted-fault run has been reviewed. The method is in use; the tests are in progress.

    • In useas a method; the tests are in progress
  • The checker is not the thing it checks.

    Where a rule is enforced by one program, it is verified by a separate instrument.

    • In useas a design principle
  • Evidence travels with a fingerprint.

    When work moves between agents it carries a digest, and the receiver reads the digest back before acting on it.

    • In use
  • What runs is what was reviewed.

    A change is finished only when the running copy’s fingerprint equals the reviewed one.

    • In use
  • Failures of process become rules.

    When the team gets something wrong, the lesson is kept, and later work is held to it.

    • In use
  • Reviewers are tested too.

    We have begun giving reviewers work with defects planted in it, to learn what a given reviewer catches and misses.

    • In useearly

What this provides

Evidence, not assurance. For each control we can say who tried to break it, how, and what happened.

What is not yet proven

Review by agents does not replace an outside audit, and we have not had one. The method above is how we work day to day, and it is what an auditor would be shown.

Reviewer testing is early. Model diversity and planted-fault tests do not establish that every defect will be found or that every designed control is built.