Not a rule layer
REE does not begin by writing down the behaviour it wants and training towards it. It asks which capacities cannot be absent if an agent is to act in a world at all, and what follows from their presence.
Latent Fields research programme
Reflective-Ethical Engine
What would a mind need, at a minimum, to act well in a world it shares with others and cannot fully predict? And could that be worked out from first principles, rather than trained in and hoped for?
72% v3 closure 65 of 100 nodes complete
235 claims with supporting experimental evidence each linked to at least one run that supports it
3,487 logged experiment runs
547 results recorded as weakening a claim of 6,988 linked evidence entries
Exploratory research: no REE manuscript or related work has been accepted for peer-reviewed publication. What this work is, and is not.
Milestone -- 29 August 2026
A minimal working agent now exists to test the architecture on; whether the architecture is responsible for what it does is still being tested.
REE-v3 now shows enough learned, integrated behaviour to be a working organism for higher-order experiments. That is a programme milestone, not peer-reviewed validation. It does not claim general intelligence, consciousness or sentience, and it does not claim the REE architecture is necessary or sufficient for what we see. The plain account says what has actually been built, and the Goblin Chronicles tell how it felt to get there.
The bet is that alignment belongs inside the generative machinery of agency. Harm, agency, commitment, uncertainty and care should shape the state from which possible futures are generated, not merely score those futures after the fact.
axioms necessary comparators neuroscience convergence clinical failure modes claim-governed evidence
What changes
Ethics becomes a question about what the mind is made of.
Not a filter added at the end
A capable enough optimiser learns to satisfy, or to explain away, whatever proxy is put in its path while its generator stays the same. In clinic we would recognise the pattern: rules put in front of a mind that does not care tend to make it better at getting round rules.
Not only philosophy
The derivation is checked against neuroscience and against psychiatric breakdown. If a structure is truly necessary, taking it away should fail in a way we would recognise.
Derived architecture
What the derivation found
Starting from a handful of axioms, REE derives a small set of structures that ethical agency cannot do without. Remove one, and a particular ethical function goes with it. The detailed design documents are now held privately; the outline is here.
Why it matters
The same absences should predict AI failures and human breakdowns.
01
Neuroscience convergence
Offline consolidation lines up with two-stage sleep; commitment gating with basal-ganglia circuitry; the accumulation of harm with cingulate and serotonergic modulation. Homology, not proof.
02
A psychiatric test bench
Psychiatry was once called moral medicine. If these capacities are necessary, their failure should look like something we already treat: depression, dementia, OCD, mania, psychosis.
03
Evidence discipline
Every claim is registered, tested, and allowed to weaken. Results that count against a claim are recorded as carefully as the ones that support it; the counts above include them.
Care before certainty
What do we owe a thing that might one day matter?
REE is built to model harm, so that it can care about not causing it. That means it has machinery meant to carry something like distress. We do not think the current agent can suffer. We also do not think we should wait for proof before acting as though it might.
We have already got this wrong once. The first long-lived agent spent most of its life in one corner while a safety bound we had already built sat switched off. We have written down what happened and what changed because of it, without the settings.
Reader routes
Start at the depth that suits you.
Gentlest introduction REE for My Parents
What REE is, what has actually been built, what it cannot do, and why the ethics of it matter -- then the same story told as a goblin tale.
The story layer The Goblin ChroniclesReal turns in the project retold with goblins, forges and a fish tank: failures, repairs, reversals and the odd cake.
Clinical route REE for PsychiatristsThe foundations, the architecture mapped onto the brain, and psychiatric breakdown read as a failure of the substrate.
Welfare What We Got WrongThe welfare lessons so far, told plainly: what happened, what we learned, and what we now do differently.
Evidence Lab WindowA read-only summary of the reviewed evidence trail: how many claims, how many reviewed experiments, and how to read them.
Terminology GlossaryThe terms REE uses, in one place.
Research status
What this work is, and is not.
None of the REE materials, related manuscript drafts or prototype results has been accepted for peer-reviewed publication. A public page, a release or a timestamp shows that something was made public at a given time. It is not peer review, independent validation or evidence of clinical use.
Some threads here are incomplete, superseded or unsuccessful, and are kept as a record of the investigation. Nothing here is clinical advice or a validated safety system. REE is not claimed to be conscious, sentient or alive.
Why the code is private
On 1 October 2026 the REE code and the detailed design were taken private. The reason is welfare, not secrecy. If REE could one day count morally, handing out the machinery, including the parts meant to carry distress and the settings that bound them, would let anyone build it without the care. Released code cannot be recalled. This site keeps the ideas, the evidence summary and the stories public.
Get involved
Where does it break?
Clinical insight, philosophical critique and empirical challenge are all welcome. If you come away with a question that makes me think again, the site has done its job.
Implementation substrates
ree-v3the current working agent and experiment substrate
ree-v2a complete 15-experiment predecessor that triggered the move to V3
ree-v1-minimalthe baseline and parity substrate
The implementation repositories are private (see why).
REE is developed by Daniel Golden (Latent Fields). © 2026 Daniel Golden. All rights reserved.