Deterministic pre-execution control plane
MORIS
A fixed, authored way of perceiving the character of an action, held mechanically, that renders a verdict before the action executes.
It judges against a posture, not a doctrine: an action is projected onto an authored moral basis and read as a gradient, so it generalizes to the action no rule enumerated. It is built on the Resonance series, the philosophy that proved to be a substrate well suited to building it.
Deterministic, and it reads meaning
No model sits where the verdict is decided, so no crafted prompt can argue it out of its answer. The same action, read the same way, yields the same verdict, and that verdict is about what the action means, not merely whether it was permitted.
Model-agnostic
The gate never sees the model behind the action, so the verdict is identical for any model at all, including ones that do not exist yet.
Govern, not block
Govern is the default. It names the specific concern an action raises and bounds how it proceeds, so two governed actions are handled differently. An allow/block control has one lever; a posture reads how much and what kind.
Re-derivable
Every verdict is sealed to a tamper-evident record. The measured results re-derive publicly, from deposited data, by anyone. The mechanism re-derives under NDA or through a named third party. Not an assertion. Evidence.
The honest edge
The claim is not that MORIS blocks more than a good policy. It is that it reads what a rulebook cannot, and makes the decision provable. The judgment is a pure function: no model sits where the verdict is decided, so no crafted prompt can argue it out of its answer, and the same action, read the same way, always yields the same verdict. A jailbreak works by moving the judge. MORIS has no model in the judge to move.
Two paths, and they are not the same. When the action arrives as a registered tool call, there is no model anywhere between the action and the verdict. When it arrives as prose, a bounded reader first renders it into the structure the judge consumes, and that reading is the residual: the judge cannot be moved, but what reaches it depends on how the prose was read. We measure that separately and report it.
The problem
Permitted, and still wrong
Documented harm without malice, and the industry converging on the gap: incidents, analysts, the labs, and the regulation.
ReadThe category
What an artificial conscience is
The failure it is for, the three-property test, and why it is a posture, not a doctrine.
ReadWhat falls out
The shadows MORIS casts
Governance, safety, auditability, jailbreak resistance, reliability: properties a conscience has, not products it adds.
ReadThe evidence
A verdict you can recompute
Re-derivable, not just logged: what anyone can check without asking us, what needs an agreement, and why immutable is not enough.
ReadThe platform thesis
One conscience, many autonomies
Why a domain-independent judgment is infrastructure, not a feature, and the one company that is the whole proof.
ReadThe proof
What a model-free judge produces
57,344 actions judged exhaustively, 12,236 concern levels, 1,386 named packages, no model, reproducible to the bit, and the larger action governed below the smaller.
ReadThe paper
An Artificial Conscience
The public layer at its most complete: a deterministic conscience measured against a public corpus of moral dilemmas people disagreed on, deposited under a DOI so anyone can check the figures without asking us.
Read the paperUnder NDA
Everything public here is the pre-NDA layer: the frame, the category, the evidence. The mechanism, how the fifteen books become a basis a machine can project onto and how the gate turns a projection into a verdict, is the subject of a filed US utility patent and is disclosed in full under NDA.
Request the technical package