From a self-description to the organization it describes

A self-description is a record within an organization. It can omit the distinction needed to predict a mechanism, preserve information that cannot currently be accessed, or describe a disposition accurately while leaving its authority untouched. The source–readout results make the first distinction exact. The later mechanism supplies the acquisition and write routes needed to connect an adequate account to conduct.

7 From a self-description to the organization it describes

7.1 Five uses of source

The unknown fundamental ground, a nominated mathematical source, an accessible readout, a self-description and an agent as a causal source of action are different objects. A source model $X$ is a declared domain of states or preparations. An observation $m:X\to M$ specifies a readout. A self-description is a retained representation within an organization, possibly derived from that readout. An agent is a causal source when its realized activity makes the relevant contribution to action. Neither the word “source” nor an incomplete $m$ proves that these objects coincide.

The canonical source-readout papers establish exact structural restrictions on nominated targets, not a theory that every finite model is the ground (Rodgers 2026a, 2026b, 2026c). Consider a surjective readout $m:X\to M$ and an independently defined target $c:X\to A$. Here $c(x)$ might be a specified interventional signature of a recommendation-producing mechanism: its responses under nominated input, model-read and authority-write tests. An independently fixed signature avoids defining recognition as whatever a controller happens to report.

Proposition 7.1 (Inherited descent and target completion). A function $\bar c:M\to A$ satisfying $c=\bar c\circ m$ exists exactly when $c$ is constant on every $m$-fibre. Adjoining any auxiliary $w=\omega\circ m$ cannot repair a failure. The joint image

$$E_c=\{(m(x),c(x)):x\in X\}$$

is a coarsest target-sufficient extension: if $r:X\to E$ is a surjective extension with $m=\pi\circ r$ and $c=v\circ r$, there is a unique map $h:E\to E_c$ with $h(r(x))=(m(x),c(x))$.

Proof. Descent implies fibre constancy immediately. Conversely define $\bar c(m(x))=c(x)$; constancy makes it well defined. Since $(m,w)=(m,\omega(m))$, every function of that pair is still a function of $m$. Finally set $h(e)=(\pi(e),v(e))$. Surjectivity of $r$ gives its image in $E_c$ and uniqueness. It also shows that $h$ is onto $E_c$. ◻

This is the canonical completion result, restated for inspection; the self-mechanism target is its application (Rodgers 2026b, Theorem 3.1, Proposition 3.4, Theorem 8.3). It says what a sufficient extension must distinguish. It supplies neither a diagnostic procedure nor its physical availability. A description can be insufficient for this target while already sufficient for a task. Appendix A preserves the blind-control and powerless-informed-controller counterexamples.

There is a quantitative counterpart, using the total-variation distance defined in Section 2. Suppose two organizations have the same self-description but different laws $P_0,P_1$ for an admitted authority diagnostic. A description-only forecast must use the same law $Q$ for both. The triangle inequality gives

$$\max\{\operatorname{TV}(P_0,Q),\operatorname{TV}(P_1,Q)\}\geq\tfrac12\operatorname{TV}(P_0,P_1).$$

The midpoint $(P_0+P_1)/2$ attains the bound for the pair. This is the two-source radius result of Sealed or Leaky, now applied to self-mechanism forecasting (Rodgers 2026p, sec. 5, Theorem 5.2). It diagnoses an inadequate nominated account, not a loss of agency or the unique ontology behind an adequate one.

7.2 Information, access and authority

Three repair problems differ. Relevant distinctions may be absent from the current input. They may already be retained but inaccessible under the current operations or deadline. Or they may be accessible while the organization assigns them an inappropriate role in assessment. Passive postprocessing cannot solve the first problem. A new interpreter, dispatch rule or control route can solve the second or third without increasing Shannon information in a retained record. Thus not every recognition process needs new information, and not every useful instruction adds a previously absent fact.

Nor does the passive theorem show that words do nothing. An utterance can be a physical input that changes state, supplies a procedure or opens an existing route. That changes the unchanged-readout premise. In the construction below, paid diagnostics add a previously unavailable distinction, while a separate installation operation changes its practical use. Keeping those operations separate explains why a correct verbal self-description can coexist with an unchanged policy.

The canonical internal-observer ceiling likewise requires actual factorization of the admitted probes through the nominated readout (Rodgers 2026e, Assumption 7.1 and Theorem 7.2; §8). Being physically internal does not prove that factorization. A new paid probe can refine the accessible distinction. Exact projected closure is also a separate question: a many-to-one description can evolve autonomously if its fibres are dynamically preserved (Rodgers 2026d, Theorem 5.2). Incompleteness alone establishes neither nonclosure nor nonzero memory.

Sealed or Leaky distinguishes preserved information from accessible observables. Transport through a bimeasurable bijective evolution preserves full measurable distinguishability while a restricted readout can cease to expose it. A retained earlier record prevents that information from vanishing from the cumulative record, but a mathematical inverse is not automatically an implemented operation (Rodgers 2026p, sec. 7, Theorem 7.1). Its operational-surrogate result concerns equality of complete experiment laws, including adaptive records and interventions, rather than a unique physical ontology (Rodgers 2026p, sec. 5, Theorem 5.1). Its Bell results have additional preparation, outcome-determinism, measurement-independence and conditional no-signalling premises; they do not supply a universal no-agency theorem.

The extended treatment Bounded Perspective, Real Control distinguishes bounded observational competence from complete source reconstruction (Rodgers n.d.a). The separation of retained history from present authority is explicit in A Chooser with a History (Rodgers n.d.b). In the latter’s reversible archive example, an old rule remains recoverable while future action uses its replacement. The new question is how the organization acquires a warranted account of that rule and uses it in selective revision. Preservation, accessibility and authority are therefore not three names for one quantity. Information can survive without governing present action; a model can be accessible without being used; an accurate self-account can acquire a causal role through a retained assessment-and-write procedure. No external observer is required. The observer and updater are themselves parts of the organization being described.