The Entity Operating Layer

For forty-nine years, your divergence from the model was recorded as a deficit. This work inverts the sign.

Andreas Ehstand · Independent Researcher, AUGMANITAI · ORCID 0009-0006-3773-7796 · 23 September 2026

The assumption nobody inverted

In February 1977, Brian P. Carr and Ira P. Goldstein published Overlays: A Theory of Modelling for Computer Aided Instruction at MIT. Their definition, verbatim: “Overlay modelling is a technique for describing a student's problem solving skills in terms of [a] modular program designed to be an expert for the given domain. The model is an overlay on the expert program in that it consists of a set of hypotheses regarding the student's familiarity with the skills employed by the expert.”

The expert program is the standard. The individual is modelled as a subset of it. Whatever the individual holds and the reference does not is a gap — a misconception, a deficit, something the system exists to close.

That assumption did not stay in 1977. It is the operating assumption of contemporary practice: models are aligned to aggregated human preference, and the individual who diverges from the aggregate is treated as an outlier to be smoothed, corrected or persuaded.

The reference layer is the baseline, not the standard. The divergence of an entity from it is not a deficit to be closed. It is the entity's holding, and it is what the record exists to preserve.

What follows from one reversal

• If divergence is a deficit, you overwrite it. If it is a holding, you version it.
• If divergence is a deficit, fidelity means convergence on the expert. If it is a holding, fidelity means faithfulness to the entity — and becomes something you can measure.
• If divergence is a deficit, nobody asks who owns it. If it is a holding, provenance, custody and succession become necessary — and that is what the prior art does not address at all.

What is recorded

The meanings an entity attaches to its own words; where its boundaries begin and end; what takes precedence when two valid commitments collide; on what grounds it decides; and the affective states that explain those decisions — recorded because they explain the decisions, not as psychological assessment. All of it versioned, so the record has a history rather than a single state.

The boundary is language, not subject matter: whatever an entity can put into language belongs to its operating layer. It applies to any entity describable in language — a person, a team, a company, a discipline, a machine, a software agent, a swarm. The human case is the first and most valuable application, not the object of the method.

It is not a copy of a brain, not voice or likeness cloning, not a psychological profile, not a personality score, and not an authentic new decision by the person. A later system working inside the record produces a model interpretation of a documented frame. Promising less than others is the point — it is what makes the remainder provable.

Why this was not possible earlier

Three conditions had to hold at once. A machine capable of producing the general layer — Boose's 1984 system explicitly required none and had none available; Carr and Goldstein hand-built an expert program for a single game. A reason to preserve the divergence — once systems act rather than answer, the frame becomes the standard against which the action is judged. And a way to date authorship without disclosing content.

A falsifiable consequence

Alignment to an aggregate cannot be verified, because there is no party to verify it against. Alignment to a named entity can: the entity reads the output and says whether it is faithful, and the disagreement is measurable.

The only form of alignment that admits verification is alignment to a specified entity with a recorded operating layer. Personalisation is not a convenience feature on top of alignment. It is the only version of alignment that can be audited.

This is falsifiable. If a published method verifies alignment to an aggregate without reference to any identified party's own judgement, it is wrong.

The dated record

The sealed register publishes each entry as a cryptographic fingerprint only — number, function, SHA-256, no content. Any single entry can be opened later and re-computed by anyone, with no key, no service and no permission required. Root fingerprint over all 803: 41f9f070f92476b668c54ecd350459b2d094c7365cda4e338f4ad80d0a134f5f

Statement of position

I claim no world first, and none is needed. The elicitation of individual distinctions is documented from 1984 (Boose, AAAI). The comparison of an individual against a reference model is documented from 1977 (Carr & Goldstein, MIT). I have found no published work that treats the divergence as the holding rather than the deficit — versioning it, measuring fidelity to the entity with a published instrument, and anchoring its provenance to a date by third parties. For the individual components there is adjacent work, cited above. For that combination I have not found it, and for provenance, custody and succession of a layer of meaning I have found nothing at all.

This is the result of a bounded search, not a proof of universal novelty. Where such work exists, its authors are invited to date it.

Scope This page sets out foundations. It does not set out the working procedure — the questions that locate divergence, the conditions under which a definition becomes valid, the individual fidelity scoring rules, or the corpus.