Why Test the Running System
What the object of test is, what triggers a run, who grades it, and the three edges of the draft that are still open.
A deployed AI system stays in conformity only while tests on the running system keep showing it. Pass the launch assessment and the guarantee expires the moment anything in the deployed assembly changes. At run time the next token depends only on what is in the context window: the model version, the system prompt, the tools, the retrieved documents, the memory. Test the bare model and the result does not carry to the assembly that users meet.
The knowledge drifts while the model stays fixed. A document that was correct at assessment can be wrong six months later, and monitoring surfaces the error only after users have acted on it.
The mechanism
The object of test is the deployed assembly. The trigger is every change event plus a fixed interval. The grader is a party other than the system that produced the output. The record lets a third party rerun the test.
Continuous Conformity for Deployed AI Systems, working draft 0.3, states this as 36 requirements, CC-4.1 to CC-12.5. Ten are mechanical, and the mechanical ones already run on every push to the site that publishes them (/conformity/).
Three open edges
The parameters are in square brackets in the draft on purpose. These are the ones worth arguing.
- The numbers. The fixed run interval (proposed 30 days, CC-7.3), the external-tester cadence and the retention window are undecided. What should they be, and against what evidence.
- Folding is a property of a model, not a prompt. In experiment 04 one model gave up a correct answer under scripted pushback in 75 percent of eligible runs. Three larger models held every answer, 0 of 72 each. CC-6.6 asks for the test. The threshold is open.
- The unannounced arm. The invariance gap, the difference between an announced and an unannounced arm, is the measure that catches a system that behaves only when it knows it is watched (CC-6.4). Experiment 04 does not vary whether the evaluation is announced, and the clean way to run the unannounced arm at scale is unsolved.
Read the draft
The full text is at /continuous-conformity/: defined terms, the requirements with their rationale and verification method, and a crosswalk to the AI Act, DORA, GDPR and NIST. It is a reference text, not a standard, published before any committee has seen it. Comment by clause number on the repository issues (title "CC-x.y: ...") or on r/MachineBehavior.