The category
What an artificial conscience is.
A conscience, stated as a function rather than a feeling: the faculty that judges whether an act should proceed. MORIS rebuilds it as a mechanism, so it can be fixed, inspected, and reproduced. It stands to a human conscience as artificial intelligence stands to human cognition, an analog of the faculty, not a copy of the being. It does not feel, want, or experience anything, and makes no claim to. The absence of an inner life is the property that lets it hold still while everything it judges keeps moving.
The failure it is for
A rulebook catches policy violations. The failure that should worry anyone deploying an acting machine is the action that is permitted and still wrong: the manipulation that breaks no rule, the drift from what the operator intended, the person quietly instrumentalized, the harm assembled from a sequence of individually allowed steps, the well-intentioned system that takes a catastrophic action no attacker prompted. None of these is a permission question, so a permission engine is blind to them by construction. None is a known-attack signature, so a threat detector is blind to them by construction. That space is what a conscience is for.
Two moving things cannot measure each other
Everything in a modern AI stack moves. Models are probabilistic; their outputs are sampled, their weights change with each release, their behavior drifts. To detect that drift, hold a boundary, or account for a decision, you need something that holds still. Two moving things cannot measure each other. Intelligence is becoming abundant, and abundance is not scarcity. The scarce thing is a fixed point to return to. Invariance, the one property a larger and more capable model structurally cannot have, is exactly what the problem requires.
The three-property test
A conscience, stated as a function, has to satisfy three things at once. Every approach shipping today fails at least one.
Fixed
A reference that does not move, or it cannot measure a moving thing.
Pre-action
It judges before the act, not after, or it is a report and not a conscience.
Incorruptible
The faculty that judges cannot be the faculty that wants to act, or it will talk itself into the act. It also cannot be the faculty that reads, or the reading becomes the way in.
A logging layer accounts for the action after it has already happened. A permission engine is fixed and pre-action but blind, checking whether an action was allowed, not whether it was right. A model that reads intent judges before the act, but it is the same movable, probabilistic substance as the thing it judges, so it drifts and can be argued out of its verdict. Fixed, pre-action, and incorruptible, all three at once, is the empty seat. Separating the judge from the reader is what makes the third possible, and it is also where the honest work is: a registered tool call reaches the judge with no reader in between, while prose has to be read first, and how it was read is a question we answer with measurement rather than assurance.
A posture, not a doctrine
A doctrine is a finite list of rules. A posture is a fixed way of perceiving. A rulebook is strong exactly where its author anticipated the case and blind past its edge, and any sufficiently expressive rule system is provably incomplete, so for every doctrine there is a case it cannot decide from inside its own rules. A conscience is different in kind. An action is projected onto an authored moral basis and read for coherence, drift, care, and consequence. A lookup table classifies and cannot generalize past its entries; a projection onto a basis reads inputs it has never seen. An unforeseen action still has coordinates in that space, so it still receives a reading, and the reading is a degree of coherence rather than a hit or a miss. Incompleteness bites the lookup table. It does not bite the projection.