What we claim, and what we do not
Two sentences are worth more than a page of assurance, so here they are, and the specification says the same thing in Section 7.
The fate of a finding does not depend on the order the lenses ran in.
The text is frozen, so every lens reads the same words. Lenses report and never edit, so none can contaminate another. Conflicts are resolved by published rules with a declared tie break. Release is a property of the final text, not of the path taken to it. Break any one of those three conditions and the guarantee goes with it.
Two careful readers will not always raise the same findings. Neither will two runs.
Human judgment varies between people and days. Language models vary between runs. We do not pretend otherwise. The variance is contained rather than denied: written interrogation sets anchor human readers to the same questions, and agentic runs use pinned models, deterministic settings and an optional majority of three, with disagreement on a Blocking finding escalated to a person. The policy used is recorded on the Version Record, in daylight.
Seven standards can be run in 5,040 different sequences. The frozen text is the only input each one sees, findings are a set rather than a list, and the Canon has a declared tie break, so the union and its resolution cannot depend on sequence. This is the whole of the central claim, and it is the reason it can be tested rather than asserted: the build compares ledgers across randomised orderings on every commit, and a difference fails the build.