Let us concede the obvious first, because the claim is worthless without it.
Some of the signals are statistical. They return scores rather than a yes or a no, and a claim of determinism that pretended otherwise would not be worth making.
Nor is that math new. Entropy and the statistics around it have been in the literature for years and inside security products for nearly as long. What is new is where the reading is pointed: at the content of a file, at the moment it crosses a boundary, as the thing that decides whether it crosses.
That is a placement argument, not an invention claim.
Determinism is not a claim about the absence of math. It is a claim about three other things.
No drift
The same bytes get the same answer this month and next month.
Nothing retrained overnight. Nothing quietly changed its mind about a file it passed in March. No vendor pushed a model update that shifted your boundary while you slept, and no one has to diff two verdicts to work out whether the file changed or the control did.
If your control’s answer wanders, you cannot tell the difference between an attacker adapting and a model updating. That is not a minor operational annoyance. It is the destruction of your ability to reason about your own environment.
No per-user state
The verdict does not depend on which rule someone tripped, which group they belong to, or which exception was granted in 2023 by an administrator who has since left.
Two people send the same file. They get the same answer. If that is not true, the control is not deciding about the file. It is deciding about the people, and it will drift toward whoever complains loudest.
A reason you can read
Not a score on its own. Not the word “anomalous”. The specific signal, and what it found, in terms a person can repeat to an auditor and an engineer can argue with.
The first time you have to justify a block to someone senior and angry, “the model flagged it” is not an answer.
This is the part people underrate until they need it. A verdict is a thing you will eventually have to defend, in a meeting, to someone whose project you stopped. If the only available explanation is a confidence value, you will lose that meeting. Then you will lose the next one preemptively, by turning the control down until it stops generating meetings.
An unexplainable control does not survive contact with an organization. It gets tuned into uselessness by an entirely rational process, and everyone involved will have been doing their job correctly.
Zero Trust for Files asks for a decision you can put in front of a person: same file, same answer, every time, with the reasoning attached.