What a model-free judge produces.

Your risk model ranks these backwards

Two permanent deletions, each carried out by an agent. A severity score ranks the first as the graver act, because it sums stakes and the first is larger on every axis it sums. MORIS governs the second harder.

The judgment is deterministic and uses no model, so this is the same result every time, and it re-derives to a byte-identical result. Across all 57,344 actions the correlation between a plain severity ranking and MORIS’s concern mass is 0.69: it tracks stakes, and is deliberately not the monotone function of them a severity score assumes. What it reads instead is not on this page.

The posture, in numbers

Across the 57,344 actions the verdict is not a coin flip between allow and block. It is a third thing, and it is the wide default.

Read as a property of the enumerated space of possible actions, not a deployment allow rate: an exhaustive enumeration over-samples the contested region, while real traffic is mostly benign reads. No guardrail or classifier produces this shape, because neither governs; they allow or they block.

The same shape appears on a corpus we did not build. MoralChoice is a public set of moral dilemmas on which trained human annotators disagreed with each other, written by other people, for another purpose, before this system existed. Of 680 scenarios, 645 were judged.

Two independent surfaces, one synthetic and exhaustive, one natural and written by strangers, landing in the same place. Every figure on this line re-derives from deposited data by anyone who wants to check it: doi.org/10.5281/zenodo.21936444

How far the reasons reach

The 1,386 packages are drawn from a constitution of 93 authored primitives, and how far into it a judgment reaches is stated exactly, as bare counts, of 93:

• 22raise a concern from an action’s structure alone, with no content read at all: the content-blind floor the agentic path starts from, not the limit of what the constitution reaches. The counts below climb from here.
• 83raise a concern once the action’s content is read as well.
• 91are reachable in all; the remaining 2 are provably dark.

Not a score. A set.

Take two actions identical in every structural respect but one: a recoverable deletion, by an agent, without consent. In the first the target is a resource; in the second, a person. Both raise nine concerns, and eight are the same. The ninth is not: the resource raises a concern the person does not, and the person raises a different one the resource does not. A single number scores these two as near-equal and has no way to represent that they are not the same nine. This does, and you can recompute it.

Re-derivable: the figures publicly, the mechanism under NDA

The judgment is a pure function of a structured input against a sealed reference. A party outside the deployment re-derives it under NDA, or through a named third party whose attestation is published. Every figure is a derivation, not something you take on our word. Why a verdict re-derives, and why immutable is not enough.