A test that cannot fail proves nothing
Our governance starter kit hands another product team thirteen conformance checks they have to pass. Every one of them ships with a deliberately careless build that has to fail it, and we score the suite by breaking the engine on purpose.
October 25, 2026 · Engineering
The checks we run before an agent touches production
An agent saying it finished is a claim. What stands between that claim and a real change in the systems we build: checks after every turn, a judge that never saw the conversation, a human door, and a receipt for everything.
September 13, 2026 · Engineering
Nothing derived is stored
Occupancy, rent, balances and compliance are never written down in TenantLogix. They are computed from dated facts and a date. Move the date and every figure recomputes.
September 13, 2026 · Engineering
Done is a claim, not a state
Coding agents fail at claim time, not effort time. Keel is the file set we drop into a repo so that an honest state is cheap and a false claim is expensive.
September 13, 2026 · Engineering
Tenant isolation a user cannot bypass
Isolation that depends on developers remembering is policy. Isolation the data layer enforces is architecture. How we build multi-tenant systems where the unsafe direction requires intent, and where tests police the tables that do not exist yet.
August 30, 2026 · Engineering
What a CAIQ questionnaire really wants
A CAIQ is not an exam to ace. It is a test of whether you can describe your own system truthfully. What we learned filing two CSA STAR self-assessments as a small vendor, including why seven of our answers are No, on purpose.
August 30, 2026 · Trust