Auditable architecture · Unit 13 of 13
The zero nobody measured
A system measured three times whether a sub-agent wrote its plan before working, and all three times it reported zero. On that zero a comfortable conclusion was written: asking an agent for something doesn't change what it does. The zero was real as a number and false as a fact. The tool registry was a single one per process and discarded by NAME anything already registered; the planning tool isn't just any tool, it's a closure bound to one specific session. The parent registered its own first, the child's was discarded as a duplicate, and the child ended up writing in the parent's notebook. All three measurements consulted the child's notebook —a place where that fact could not appear— and read empty. None was badly written: they were badly grounded. Nobody asked whether the file being consulted is a place where the datum can exist.
By the end, you'll be able to
- Tell a measured zero apart from a zero produced by looking where the datum cannot be.
- Use the control as the only thing that separates those two, and audit first the data that confirm what you already believed.
Understand
A system measured three times whether a sub-agent wrote its plan before working, and all three times it reported zero. On that zero a comfortable conclusion was written: asking an agent for something doesn't change what it does. The zero was real as a number and false as a fact. The tool registry was a single one per process and discarded by NAME anything already registered; the planning tool isn't just any tool, it's a closure bound to one specific session. The parent registered its own first, the child's was discarded as a duplicate, and the child ended up writing in the parent's notebook. All three measurements consulted the child's notebook —a place where that fact could not appear— and read empty. None was badly written: they were badly grounded. Nobody asked whether the file being consulted is a place where the datum can exist.
What saved the series was the control. In the arm without the request, zero sub-agents planned —and there the zero really was zero. The difference between arms was real; what was wrong was its absolute level. A control doesn't only measure the effect: it's the only thing that distinguishes a measured zero from a blind one, because a blind instrument returns the same emptiness in every arm. And the corollary that makes the lesson expensive: the bad number survived two days because it CONFIRMED what the house already believed. The instrument was audited the day a new mechanism failed to fire, not the day the datum arrived. A datum that contradicts your hypothesis gets reviewed on its own; the one that confirms it is the least audited —and therefore the costliest.
See
Open the runtime x-ray
The x-ray shows each path of an operation leaves its own trace — and that asking the wrong trace returns an emptiness that reads exactly like a fact.
Do
Run the practice in your checkout and keep the output as evidence.
Verify
Show that you can apply the unit. Progress only advances once you pass the assessment.
Criteria assessed
- Take a metric in your system that reports zero today and say where the datum would appear if it weren't zero. If you can't point at it, the metric isn't measuring.
- Explain why a blind instrument is indistinguishable from a null effect without a control arm.
Primary sources
Content verified: 2026-08-04