Understanding-Aligned Intelligence Framework (UAIF): Which Parts of This Architecture Survive Cognitive Parity

Abstract

Version 1.0.0 of this framework, published in January 2026, proposed a five-layer architecture for aligning an artificial system. At its base is a constitution derived from international human rights law. Above it sit a proof validator that checks certificates instead of searching for safe actions, a set of behavioural constraints, and a gateway applying a proportionality calculus. The layer where proofs would be generated is left open as a grand challenge. Three months later the same author published a threshold condition the earlier document had not been written against. This version puts one question to every component of that architecture: does the condition under which the component works carry a term comparing one party's capacity to another's? The instrument is the criterion of Soundness Without Leverage §3.1, applied across the corpus in Invariance Without a Beneficiary, and the answer is not the one a layered architecture invites. The seam does not run between the layers of version 1.0.0. It runs inside each populated layer, and in thirteen places inside a single named component. Three of the proportionality test's four stages turn on judgments a party makes rather than on properties of an artefact. Of six anti-manipulation constraints, one survives as written, four do not, and the remaining one survives only under a reading that stops meaning what its name says. The fallback catalogue divides evenly, two entries naming machine states and two naming a human or an effect. The trusted computing base, defined by version 1.0.0 as the components that must be correct for system safety, contains a human escalation pathway, and past the crossing that relation does not hold. What survives across the architecture is a checker, a counter, a record and a set of enumerated machine states. Three results go beyond sorting, and the second of them has two faces. The architecture specifies a decidable safety predicate and a constitution its own text describes as requiring interpretation and admitting conflicts, and it specifies nothing between them. Half of that map could be written without the term, the half that orders the principles against one another, and version 1.0.0 supplies two comparison-free ways of doing it. For the half that says what each principle holds of it supplies nothing that does not carry the term. Motivational Conformance, one half of the framework's own diagnostic, is underdetermined, and neither available completion gives the architecture what it needs at once. On the first it is invariant and carries no assurance. On the second it carries assurance and imports the capacity term through the constitution's reading. And the threat model's own table assigns the highest adversary level three controls, two of which the same document elsewhere calls unavailable or defeatable, and the third of which this reading finds dependent on a capacity relation. The reading yields a positive rule as well as a partition. Six rules let an engineer sort a design before the crossing rather than after it, and they come as much from the places version 1.0.0 went right as from the places it did not. Among the components version 1.0.0 treated as fallbacks, the survivors are disproportionately the ones it ranked lower. The rebuilt proof-obligations table is the test of whether the discipline reaches the document's central claims. Two of six rows stand unchanged, one holds with its boundaries made visible, one holds over a narrower object than its wording suggests, one holds for half its catalogue, one is not established, and five obligations are added. The rule has a limit, and it is stated in its own chapter. A design fully compliant with it would contain no component whose function is to make a human party's judgment effective. That is this corpus's title result met at the level of a built system. This is a revised architecture rather than a refutation, it is not a claim that the crossing will happen, and it is not a deployment case. The surviving half is not a safety architecture, and the paper says so before it says anything else about it. Two published requests are executed here. The Containment Paradox §5.4 asked a revised version of this framework to consolidate three things, one of which is attributed to a section that does not contain it, so the instruction is carried out in corrected form. Row 38 of the Invariance map was left reserved because supplying the missing verdict would have meant supplying the reading rather than reporting it; §9 supplies it and the row closes. Six numbered open problems are recorded, two of which were on the record before this version. One disclosure belongs here rather than in a footnote. The architecture, the criterion and the reading have one author. What is offered against that is not the author's care but an appendix: a table of a hundred and eight rows in which every verdict but two stands against a quoted sentence, a stated criterion and a named rule for where one component ends and the next begins. The two exceptions have rows of their own, one for a control version 1.0.0 names and defines nowhere and one for a list it prints only as a drawing. It has one coder and no second annotator, so it is an audit with its working shown rather than a reproducible classification, and §12.5 says so in those terms.