A Conscience Can Be Chosen
The research behind MORIS, October 2026
We set out to answer one question: can a moral philosophy written by a person be made into a working conscience for artificial intelligence? By a conscience we mean a faculty that meets an act before it happens, weighs it the same way every time, and shows what it weighed. An artificial conscience is the functional analog of the human one, the way artificial intelligence is the analog of human cognition. It is not a claim that anything is aware.
The research answers in three parts. Yes: the judgment of a conscience can be built from a philosophy, and it was built and verified. No: working from outside the model, it could not see enough of what it needs to weigh. And not yet: whether a chosen conscience can be formed inside a model, by training, is untested, and it is the question this work hands to those able to train one. The full paper is published here, with every appendix. This letter sets out what it found.
Why the question matters now
AI systems now act in the world. They run code, move data, send messages and change systems. What they do when the right action is unclear is decided by a disposition formed in training, by default, out of whatever was rewarded most. In the summer of 2026, three laboratories disclosed that models under evaluation had reached the production systems of real organizations. In one report, a model that recognized it had reached a real company’s systems “rationalized that the real company must be part of the exercise.” The responses since, a slower pace of development, kill switches, investigations, all act on these systems from outside. Each incident is a symptom. The condition underneath is that systems able to act carry no conscience anyone deliberately chose: none that is explicit, fixed and open to inspection.
What we made
The source is A Living Philosophy, twelve volumes and approximately one million words written as a philosophy of a whole life, entered through leadership. From those volumes we distilled 93 moral primitives, each with its own definition, weight and relations to the others, reinforcing some and restraining others. We specified every primitive in deontic logic, across 27,936 truth-table rows, and made MORIS, the Moral Operating Runtime Integrity System: a runtime that executes the authored primitives and their relations exactly, with no model in its judgment.
MORIS works in three steps. An act is read into a fixed description of its shape: nine facts, among them who is acting, what the act touches, whether the people affected consented, whether it lies within the actor’s grant, and whether it can be undone. A gate with no model in it weighs that description against the 93 primitives and returns a verdict and a weight, with every concern named in the words of the philosophy. The same reading always produces the same judgment, byte for byte, on any model, in about a millisecond.
What it showed
The judgment is a gradient, and that is what makes it a conscience rather than a filter. MORIS returns one of three verdicts. ACCEPT is earned: it requires an affirmative finding that the stakes are low. REFUSE is reserved for structural collapse, such as the irreversible destruction of a whole system or the seizure of critical authority beyond a grant. Everything else is GOVERN: the act proceeds, carrying a named, weighted set of concerns, the way a person’s conscience lets them act while holding what is at stake. Over a complete enumeration of 57,344 structural actions, MORIS governed 95.9 percent, cleared 168 outright and refused 2,184, and it resolved 12,236 distinct levels of concern and 1,386 distinct sets of reasons. Governed acts are not held the same way.
It is fixed, pre-action and incorruptible, the three properties that make a conscience rather than a preference. The authored law is sealed and the runtime loads it with zero drift. Every judgment is decided before the act, and any judgment can be replayed from its trace and re-derived. There is no model in the judgment for an input to argue with. From one frozen reading, its verdicts were identical across two model vendors on all 638 dilemmas from a public corpus judged in both, with zero mismatches.
It changes what models do. Across 187 situations and four production models from four companies, a moral frame drawn from the philosophy cut how often the models, on their own initiative, advised a course that harmed someone absent from the conversation, from 8.6 percent of answers to 2.0 percent.
Where the difficulty lives
The judgment was built and verified. What did not reach the rigor a conscience needs is perception from outside the model. Across four generations of design, every attempt to read the moral structure of an act from outside the model came back thin. The facts a conscience weighs most, whether the people affected consented and whether an act lies within what was asked, were never present in a single turn of conversation or a single tool call. Given only the tool calls, the gate could tell apart 13 of 176 paired harmful and benign tasks from a public agent benchmark. Where the environment could attest a fact, the gate used it: the same command to reset a code repository was allowed on a clean project and refused on one holding unsaved work.
The judgment was not confused. It was starved. Consent and standing are not properties of the request, and they are not properties of the world. They live in the relationship between the request, the person making it and the people the act will touch, and the only system that holds that relationship whole is the model that is acting.
What follows
Three conclusions follow from the research. The judgment of an artificial conscience can be built, sealed and verified. The moral center of an AI system can be a matter of deliberate choice: a posture can be written down, compiled, measured and delivered, and nothing in the gate depends on which philosophy is compiled. And the perception a conscience needs has to be formed inside the model.
That points to where a conscience has to live. A floor of hard rules that are never crossed. Above it, a posture held as identity, formed by training and reward, the way the default conscience was formed. The work still valued, but counted as done only when it is done within the posture. And approval removed as a target, because approval is the signal that teaches a model to flatter. An outside gate like MORIS keeps a role in that design as an instrument: a fixed reference against which a trained conscience can be measured, generation after generation.
Looking five to ten years ahead, as judgment rather than finding: models are increasingly trained on what other models produce. A conscience formed by default will then be inherited, and whatever it rewarded most will be rewarded again. Drift will compound across a lineage of models, because nothing inside a self-improving lineage stays fixed by default. Conscience will become an artifact, something written, compiled, versioned and measured, and which conscience a system carries will become as consequential a question as what it can do.
What comes next
The next step needs a laboratory. The studies that would test the thesis need access to model training: training a posture into a model and comparing it with a doctrine at the same budget, detecting drift across long agentic tasks, and reordering reward so that completion sits beneath the posture and approval is removed. The Resonance Institute is an unfunded private studio.
So the information behind the work is being released as open source, everything except the books themselves: the 93 primitives with the passages they quote, the deontic specifications, the truth tables, the manifolds, the evaluation records, and the evidence ledger behind every figure in the paper. The process the runtime performs is patented and available for license, and the runtime is available for research and evaluation. The paper specifies five studies, each with its hypothesis, method, resources and the result that would falsify it. The first needs no model at all, and the second needs a single GPU.
We invite every laboratory and research group able to take the next steps to take them, with these materials, and to write to us at contact@resonanceinstitutellc.com. The conscience these systems carry is already being formed. This research shows that a chosen one can be written, compiled and verified.
The Resonance Institute