Thirty-one edges, and what each obliges.

Relationships are typed, not implied.

The register asserts 31 relationships across the nine frameworks. They are not decoration on the catalogue. A framework that supplies evidence to another has a downstream consumer whose gate will not open without it, and an institution adopting one framework in isolation inherits an obligation it cannot see from that framework alone. The map is the part that says which obligations those are.

Every relationship is typed, and the types are defined in the register rather than left to inference. 7 types are authored; one of them, supersedes, is declared and carries no instances yet.

Authority has one source.

Two of the 6 types in use carry the structure. grants-authority-to appears once, and that single edge is the root of the family: AI Governance Operating Model grants the authority under which 5 of the others operate. Charter, decision rights and release authority come from one place. A framework operated without that charter is being run on the authority of whoever happened to start it.

instantiates appears 7 times and does the other structural job: it places each framework at a MESA altitude, which is how the family has a coordinate system rather than a list. MESA Framework™ is not a peer of the eight. It is the frame they are positioned in, and 6 of these edges say so on exactly that subject.

What will not open without a record.

supplies-evidence-to is the commonest relationship in the register, 17 of 31 edges. It means an output of one framework is an input or a gate entry criterion of another — not that the two are related in spirit, but that one will not function correctly without a record the other produces.

The shape is worth reading off the table. AI Incident Response Protocol (AIRP)™ receives from 7, and Five-Gate Deployment Model™ receives from 6. Those two are where the family converges: a deployment gate that cannot open without upstream records, and an incident protocol that cannot reconstruct an event without them. Evidence is not filed for its own sake in this architecture. It is filed because something downstream is going to ask for it, and the map says what.

A second authority changes the procedure.

threads-into is the strongest claim in the vocabulary: the source inserts a required branch into the target's own procedure. Not an input — a change to how the other framework runs. All 3 instances come from one framework, Sharia AI Compliance Framework (SACF)™, and the register notes the type is asserted by the authority source rather than by the framework being modified.

That is the structural consequence of a second binding authority. Where a Sharia Supervisory Board holds approval power, validation does not gain a checklist item; it gains a second and independent track, and the same is true at data certification and at vendor diligence. An institution outside that scope runs the three unmodified procedures. An institution inside it runs three different procedures, and the difference is asserted here rather than left to be discovered during an examination.

What was folded in, and from where.

absorbs records where the family changed shape. 2 edges account for 4 instruments that no longer exist separately, each folded into a framework that took over its subject. This is the part of the register that makes the work auditable over time rather than only at present: a reader who met one of these names in an earlier draft can see what happened to it, and that it was consolidated rather than quietly dropped.

One edge leaves the register.

One edge crosses out of the governance family entirely. worked-instance-of means the source is the application of a named law or boundary class to one domain, and the single instance connects the register to the engineering family: Cross-Border AI Architecture Patterns™ is a worked instance of the Boundary Invariant, which is specified in the Pattern Language rather than in the register.

That is the only structural link between the two families, and it is deliberately narrow. The families are governed by different documents — registry section 2.6 gives the Pattern Language precedence on its own side — and one worked instance is a claim about a single framework, not a merger of two bodies of work. Defensible AI is the umbrella over both; it is not a third framework.

The engineering family is not catalogued here and carries no REG ids, deliberately. Its two constructs have their own deposited specifications: PEVG, the planner, executor, verifier and generator decomposition, and PARA, the four faculties and the authority each one holds.

What each asserts, and what is asserted of it.

The same 31 edges, arranged by framework. The two columns are not symmetrical and the asymmetry is the point: the left column is what this framework claims, and the right column is what other frameworks claim about it. The register attributes every assertion to the framework that makes it, so a reader can tell whose claim they are reading.

The register governs, this page is derived.

This page is a rendering. The instrument it renders is the Defensible AI Framework Registry, and its canonical machine form is registry-v1.0.json, served unmodified against its published schema. Every edge above is read from that file at build time rather than transcribed. The relationships array carries all 31, each with its type, its subject, and the altitude it applies at.

Cite the register as: The Defensible AI Framework Registry, version 1.0. 10.5281/zenodo.22170112. That is the concept DOI and it always resolves to the latest version. CC BY 4.0.