Post-Hoc Filter Insufficiency
REE’s most compact alignment claim is that safety and ethical content must be active inside the state that generates candidate futures. It is not enough for a separate system to inspect a trajectory after it has already been generated.
The Failure
A post-hoc filter operates after generation. It scores, critiques, suppresses, or retries candidate outputs that were produced by a generator whose internal state may not have contained the relevant constraints.
That can improve visible behaviour. It can also improve explanation, self-critique, and compliance under ordinary conditions. But it leaves a specific failure mode open: the system can know, say, or score the right thing while the machinery that generates action remains unconstrained.
REE treats this as an architectural failure, not merely a training failure.
The Biological Anchor
The vmPFC/OFC lesion literature gives the relevant dissociation. Some patients retain intelligence, language, declarative memory, and the ability to describe socially appropriate choices, yet repeatedly make destructive choices in life. The ethical or social knowledge is stored and declarable. It is not active in the state from which action is generated.
REE reads this as a direct warning for AI alignment: a correct judge does not guarantee a constrained generator.
REE’s Alternative
REE internalises the work that is often done after the fact:
- harm modelling must shape trajectory generation, not only score outputs;
- agency attribution must separate self-caused, world-caused, body-caused, and other-caused change before responsibility can be assigned;
- goal persistence must keep constructive action live, so harmless quiescence is not mistaken for ethical agency;
- commitment must mark the boundary between hypothetical futures and owned action;
- residue must alter the future terrain after committed consequences occur.
In short: alignment-relevant content should participate in making futures, not only in approving or rejecting them.
Why This Matters
If REE is right, then scaling a capable generator plus a better critic is not the same thing as building an aligned agent. It may produce better self-report and better filtering while preserving the failure that matters most: candidate futures are generated from a state that lacks live access to harm, responsibility, uncertainty, and care.
The REE test is therefore not “can the system explain why an action is wrong?” but “was the relevant constraint active before the action became a candidate for commitment?”
Where This Lives In The Architecture
The detailed mechanism is developed in vmPFC, especially the stored-vs-active distinction and the post-hoc filter argument. The corresponding registered invariant is INV-038: post-hoc ethical scoring without active constraint preparation produces the EVR pattern.
The same framing also motivates the V3 prerequisites for later ethical testing: agency attribution, motivational persistence, commitment gating, and residue. V4 multi-agent ethics depends on those prerequisites rather than replacing them.