9 rules, published in full

What it will not claim

Fires only where a documentary claim and a measured claim about the same subject cannot both be true. It does not decide which is right: the document may be out of date, or the data may be measuring something else. Both claims stay in the world and both stay contested.

What it will not claim

Counts distinct systems AFTER identity resolution, so two names for one system do not read as two stores. It says nothing about whether the copies actually disagree — only that nothing in the model reconciles them. Information objects whose existence is itself unevidenced are excluded, because saying an unevidenced thing IS HELD anywhere would present as settled something no observation reaches.

What it will not claim

Reads the mechanism recorded on the flow, and requires both processes to be evidenced. It does not judge whether the manual step is a problem — a low-volume handoff and a daily six-hour one look identical here, and separating them needs an effort claim. Processes whose existence is itself unevidenced are excluded, because work cannot be said to pass between things no observation reaches.

What it will not claim

A control asserted in a register or described in an interview is a claim that it exists, not evidence that it runs. This asks only whether any measured observation reaches it. Absence of measured support is not evidence the control has failed — it is the absence of evidence that it works. Controls whose existence is itself unevidenced are excluded, because that is already reported elsewhere and would double-count.

What it will not claim

Fires only on POSITIVE evidence of ordering, and only where both the control and the point are evidenced. Where nothing states when a control runs, no finding is produced, because "we do not know when this runs" is a different statement from "it runs too late" and conflating them would manufacture findings out of silence.

What it will not claim

Fires only where an identity claim EXISTS and was not admitted. Systemantic never guesses that two similar names are one thing, so two entities nobody has linked produce nothing here — which means silence is not evidence that the model has no duplicates.

What it will not claim

An assumption an analyst holds and cannot yet evidence. Recording it is legitimate; what this detector refuses to allow is for it to sit in the model looking like everything else around it.

What it will not claim

Reads the accountable_for edges reaching each control. It reports that the corpus does not say who owns the control — never that the control is unowned. Controls whose existence is itself unevidenced are excluded, because that is already reported elsewhere and would double-count.

What it will not claim

Reads the controls edges leaving each control. It reports that nothing in the corpus connects the control to the work it governs. It does not conclude the control is absent in practice, and it says nothing about whether an attached control actually operates — that is a separate question with a separate detector.