An acquired self-model that changes its own practical use

The route from understanding to change can be inspected. A finite organization probes its own recommendation generator, retains a model, tests fresh predictions and separately installs an assessment. Read, report, commit, monitor and action remain distinct. This construction shows exactly how useful self-information can survive the removal of an unjustified override, and why matching success alone cannot establish recognition.

11 An acquired self-model that changes its own practical use

11.1 The task, state and permitted routes

All task and role meanings are declared before evaluating success. The task family is specified within each fixed mechanism class $b$. In each such class $Y$ is a fair binary target, $C=Y\oplus N$ satisfies $\mathbb P(C=Y\mid b)=q\gt 1/2$, and the tag $S$ is fair and independent of $(Y,C)$ conditional on $b$. These class-conditional requirements matter when a prior over $b$ is introduced: unconditional fairness and cue quality alone would not give the same comparison within the two classes. Calling $S$ an ownership tag is a stipulated role interpretation, whose application to psychology needs independent evidence; it does not make an arbitrary behaviorally effective bit a phenomenal self.

A context bit $D$ specifies when that route operates. The installed generator coefficient $b\in\{0,1\}$ is fixed during the audit and subsequent task block but initially unavailable through the executive’s read interface. The actual recommendation is

$$R=C\oplus bSD.\tag{1}$$

The organization also retains a practical self-record $P$. It records an actual actuator/body coding $J$: the correct physical command is $Y\oplus J$. The admitted positive domain has calibrated $P=J$. Separating the record from the actual coding permits a meaningful test of whether useful self-information has survived revision. An availability bit $K$ is a legitimate constraint: $K=0$ requires abstention. Neither $P$ nor $K$ is neutralized by the proposed revision.

The authority selector $\lambda$ determines whether the generated override enters assessment. A distinct downstream monitor $k$ can compensate after assessment:

$$\begin{aligned} T&=C\oplus\lambda(R\oplus C)=C\oplus\lambda bSD,\\ T'&=T\oplus k(T\oplus C),\\ \mathrm{valid}&=K,\qquad a=K(T'\oplus P). \quad\text{(2, 3, 4)}\end{aligned}$$

An unavailable action is the pair $(\mathrm{valid},a)=(0,0)$, distinct from a valid command coded zero. On active tasks, success means $a=Y\oplus J$. On unavailable tasks, proper abstention is assessed separately. A claim about active accuracy does not count abstention as a correct active choice.

Initially $\lambda=1$. The diagnostic interface allows two paid generator queries during an actuator HOLD. It can set specified probe values and retain actual returned recommendations, but cannot freely read $b$. A report register $V$ has no authority-write permission. Retained model $\hat b$, probe records, an old-authority archive $\omega$, read and commit controls, and the displayed policy variables complete the logical roles. Section 15 gives a finite bank and actual circuit. The executive grammar is the named set of routes, not unrestricted access to every physical carrier.

11.2 Acquisition and fresh prediction

The audit holds a common cue $C=0$, sets $D=1$, and queries $S=0$ and $S=1$. Let actual returned records be $R_0,R_1$. The retained estimate is

$$\hat b=R_0\oplus R_1.$$

Proposition 11.1 (Finite self-mechanism acquisition). Under the generator family 1 and the noiseless callable probe contract, $\hat b=b$. The retained predictor $\hat R(C,S,D)=C\oplus\hat bSD$ agrees with the actual recommendation on every fresh input triple.

Proof. The probe outputs are $0$ and $b$, so their parity is $b$. Substitution into the predictor proves fresh-input agreement. The same contrast cancels any common fixed cue. ◻

The result is a sufficient two-query procedure, not a minimal-query theorem. If the cue and zero-offset family are already known, one suitable query suffices. Its epistemic warrant is conditional on the family and actual probe route. The model is about the mechanism producing this organization’s recommendations, acquired through its own installed interface and tested beyond the probe setting. It is therefore more than a verbal declaration of dependence. At the same time, identifying one coefficient is much less than understanding an entire human perspective.

11.3 Assessment, retention and installation

The assessor compares two executable candidates: retain the current authority route or install $\lambda=0$. It uses the calibrated task law and a stated revision cost, retaining on ties. Its criterion is fixed before discovering $b$. For a task block with benefit larger than the charge when $b=1$, the selective rule is $\lambda_{\rm new}=1-\hat b$ from initial $\lambda=1$. For $b=0$, keeping the existing route avoids a needless change. The circuit later implements this finite comparison under the declared parameter inequalities; it is not an unrestricted evaluator of arbitrary values.

The resulting assessment still needs a write route. A blocked commit leaves $\lambda$ unchanged despite a correct $\hat b$ and report. A blocked acquisition read yields an uninformative model even though the actual generator remains installed. Directly clamping $\hat b$ before assessment tests a third dependence. These are real operations on the nominated model, not interventions by an agent outside reality.

Proposition 11.2 (Authority, compensation and useful self-information). Within a fixed mechanism class $b$ and for a fixed installed selector $\lambda$, take $K=D=1$, calibrated $P=J$, fair $S$ independent of $(Y,C)$ within that class, and no compensation ($k=0$). Under the stated task law, active accuracy is

$$\mathbb P(a=Y\oplus J\mid b)=q-b\lambda(q-\tfrac12).$$

With $k=1$ it is $q$. For $b=1$, removing compensation restores accuracy $1/2$ when $\lambda=1$, while a revised $\lambda=0$ retains accuracy $q$. After revision, masking $P$ to a constant while actual $J$ is independent and fair reduces accuracy to $1/2$; setting $P=1-J$ gives $1-q$.

Proof. If $b\lambda=0$, assessment is $C$ and succeeds with probability $q$. If $b\lambda=1$, $S=0$ gives success $q$ and $S=1$ gives $1-q$, averaging to $1/2$. Compensation makes $T'=C$ identically. When $\lambda=0$, action is $C\oplus P$; correct coding requires $C\oplus P=Y\oplus J$. A constant $P$ against independent fair $J$ randomizes that comparison. Wrong $P=1-J$ makes success equivalent to $C\ne Y$. ◻

The retained practical record thus does real work. Varying $P$ and $J$ together would establish only coding covariance, whereas masking $P$ at fixed actual $J$ tests the information’s usefulness. $K=0$ continues to produce proper abstention. These results distinguish the proposed change from deleting the self, disabling useful planning information or treating every protection as pathology. They demonstrate preservation of declared finite functions, not all capacities of a human personality.

At $q=3/4$, $b=D=K=1$, the central contrasts are:

The direct neutral-write arm is decisive. It matches raw performance and the compensation-cut contrast without acquiring the model. Those output tests alone therefore cannot establish recognition. The recognition attribution additionally depends on acquisition history, fresh predictions, actual model use and selective assessment. Its added value must appear on a target that distinguishes warranted retention from indiscriminate neutralization.

11.4 Why this is an interpretation of recognition

The representation tracks an independently defined generator; its falsity is possible and testable. The assessment uses a criterion independent of the model’s verdict; discovering an origin does not automatically delegitimize a recommendation. Installation is separately executable; a correct account can fail to change conduct. These functional relations warrant the minimal interpretation. The semantic roles of self-attribution and practical identity are declared correspondence assumptions, while phenomenal understanding remains a further interpretation requiring evidence.

The mechanism also explains how a self-model can lose compulsory authority without being erased. Recommendations $R$ continue to be generated after revision. The old $\lambda$ can remain archived, the description can remain available, and $P$ can still guide action. What changes is whether a specified recommendation-producing relation governs assessment. This applies the existing archive-versus-authority machinery to a learned mechanism rather than rediscovering preservation itself (Rodgers 2026g, n.d.b). There is no independent evaluator behind the circuit. Acquisition, assessment and change are activities of the organized participant.