Recognition within a complete history

Learning occurs within a complete history. The exact 64-history ensemble follows a participant from an initially uninformative readout through diagnosis, installation and fresh ingress. It distinguishes knowledge of the installed mechanism from knowledge of the task and prediction of the command. Those distinctions give the source–readout account operational content without an external projector or an omniscient participant.

12 Recognition within a complete history

A complete history can contain an agent’s acquisition of information that was previously inaccessible to it. The completeness belongs to the history; the change belongs to the participant’s records and practical organization. To connect these descriptions without an observer outside reality, we now place the retained diagnostic transaction inside an explicit finite path space. The probability and decision results are elementary. The contribution is the worked identification of their variables with acquisition, assessment, installation and fresh action, including their different conditioning assumptions. This is a logical trace of the existing transaction, rather than an additional physical compilation or a claim about human consciousness.

12.1 One source point, one complete transaction

Let

$$U=(B,Y,R,S,J)\in\{0,1\}^2\times\{0,1,2,3\}\times\{0,1\}^2$$

have the uniform law on its $64$ elements. Thus $B,Y,S,J$ are independent fair bits, and $R$ is an independent uniform four-valued noise index. Define

$$N=\mathbf 1_{\{R=0\}},\qquad C=Y\oplus N,\qquad q=\mathbb P(N=0)=\frac34.$$

Here $R$ names the noise index, not a recommendation or a returned probe record. We use $G=C\oplus BSD$ for the recommendation generator in this subsection. Set $D=K=1$, $k=0$ and retain calibrated practical coding $P=J$. The installed coefficient $B$ is constant during the transaction. The uniform source law describes a declared comparison ensemble; it does not assert that a physical source supplies fresh randomness at each later stage.

For each source point $u$, the protocol determines a unique sequence

$$h(u)=(x_0(u),x_1(u),\ldots,x_5(u)).$$

Each $x_t$ includes the physical source environment and the controller’s registers. Its accessible controller trace $z_t$ has the following successive contents; earlier diagnostic records are retained in this finite trace:

$$\begin{array}{ccl} 0&:&\text{ready state, initially }\lambda=1;\\ 1&:&C_{\rm probe}=0,\ S_{\rm probe}=0,\ D=1, \quad R_0=0;\\ 2&:&C_{\rm probe}=0,\ S_{\rm probe}=1,\ D=1, \quad R_1=B,\quad\widehat B=R_0\oplus R_1=B;\\ 3&:&\omega=1,\qquad\lambda=1-\widehat B;\\ 4&:&\text{fresh }C,S,P=J\text{ enter their admitted registers};\\ 5&:&A=C\oplus\lambda BS\oplus J=C\oplus J. \end{array}$$

The initial readout $z_0$ is constant across the ensemble. The full physical state $x_0$ is not: its environment already carries the source distinctions. Otherwise a deterministic protocol could not produce these different histories from one identical complete initial state. In particular, $Y$ and $R$ are not initially disclosed to the executive merely because they are coordinates in our description. Source points label physical histories, not an agent-independent ground of reality.

The map $u\mapsto h(u)$ is injective because the physical history retains its source environment. Write $\mathcal H=\{h(u):u\text{ is a source point}\}$ and give this image the induced measure $\mu(\{h(u)\})=1/64$. Exactly one such history is actual on any run. Opening the fresh ingress at stage $4$ is an event within that history. It is not a second process in which an outside projector advances through an otherwise inactive block. The stages are an ordering of actual physical operations, with their required ports, holding and memory resources. They do not add a new time parameter outside the world.

12.2 What the participant can predict

Let $\mathcal F_t$ be the cumulative information generated by the admitted controller readouts up to stage $t$. In this ideal finite trace,

$$\begin{split} \mathcal F_0=\mathcal F_1&=\{\varnothing,\mathcal H\},\\ \mathcal F_2=\mathcal F_3&=\sigma(B),\\ \mathcal F_4=\mathcal F_5&=\sigma(B,C,S,J). \end{split}$$

The installed selector, archive and output add no distinctions beyond these: $\lambda$ is a function of $B$, $\omega=1$ is constant, and $A$ is a function of $(C,J)$. The inclusions are therefore inclusions of accessible record distinctions, not an assumption that every physical register becomes available. For binary target $T$, optimal prediction under zero-one loss has accuracy

$$\operatorname{Acc}^*_t(T) =\sum_{f\in\operatorname{At}(\mathcal F_t)} \max_{v\in\{0,1\}}\mu(f\cap\{T=v\}).$$

Counting the $64$ histories gives the exact values

$$\begin{array}{c|ccc} \text{prediction target}&t=0,1&t=2,3&t=4,5\\\hline B&1/2&1&1\\ Y&1/2&1/2&3/4\\ A&1/2&1/2&1 \end{array}$$

The diagnostic identifies the installed source coefficient, reducing its optimal error from $1/2$ to zero. It supplies no information about the independent fresh target $Y$. After ingress, every atom $\{B=b,C=c,S=s,J=j\}$ contains four equiprobable source points: three have $Y=c$ and one has $Y=1-c$. Hence

$$\mathbb P(Y=C\mid\mathcal F_4)=\frac34, \qquad \operatorname{Err}^*_4(Y)=\frac14.$$

The whole history has a definite target, while the participant still lacks its value. Conditioning on the complete history would instead identify $Y$ exactly. These are different conditioning operations, not contradictory claims about whether an actual event is settled.

The future action is equally uncertain before fresh ingress, even after correct diagnosis: independent fair $J$ masks $C\oplus J$ from the earlier readout. After the cue and practical self-record arrive, its prediction is exact. Predicting what command this organization will issue is thus easier than predicting whether that command will match the target. The distinction prevents perfect self-prediction from being mistaken for omniscience about the task.

These are also three different temporal descriptions. Inferring $B$ concerns a source mechanism already installed. Predicting $A$ concerns a later stage of the transaction. Predicting the fresh target concerns the later task check, although its value is already represented in the source environment. None is, merely by being an information update, a description of the subjective feeling of time passing.

12.3 Information acquisition and practical revision

For comparison, an unrevised controller blindly retaining $\lambda=1$ would issue

$$A_{\rm old}=C\oplus BS\oplus J.$$

Its active accuracy relative to the correct command $Y\oplus J$ is $3/4$ when $B=0$ and $1/2$ when $B=1$. Averaging over the fair coefficient gives

$$\mathbb P(A_{\rm old}=Y\oplus J)=\frac58, \qquad \mathbb P(A=Y\oplus J)=\frac34.$$

The difference is $1/8$. The numerical pair agrees with the physical error-budget illustration in Section 15, but its conditioning differs: there $b=1$ was fixed and $D$ was fair; here $D=1$ is fixed and $B$ is fair. The fixed-$b=D=1$ contrast in Proposition 11.2 was instead $1/2$ against $3/4$. These experiments must not be merged because their numbers happen to coincide.

The revised action follows because the diagnostic estimate actually feeds the assessment and write route, making $B\lambda=0$. It does not follow from an abstract increase in information alone. The retained estimate and old-authority archive remain available, while practical coding $P=J$ continues to guide the command. A source mechanism may therefore remain represented without retaining its former governing role. As before, a direct neutral write also achieves $3/4$ raw accuracy without diagnosis. Recognition is not necessary for that output score. Its further value in the costed comparison concerns warranted retention versus unnecessary revision, with the declared diagnostic and installation charges. This trace does not waive those charges or enlarge the original comparator’s permitted routes.

Proposition 12.1 (Nested access and finite Bayes risk). For a fixed finite decision problem with loss $\ell(T,d)$, a fixed finite action set, and nested information $\mathcal F\subseteq\mathcal G$, let

$$\mathcal R^*(\mathcal F) =\min_{d\;\mathcal F\text{-measurable}} \mathbb E[\ell(T,d)].$$

If acquiring and using the additional distinctions has no additional cost or restriction in this comparison, then $\mathcal R^*(\mathcal G)\leq\mathcal R^*(\mathcal F)$.

Proof. Every $\mathcal F$-measurable decision rule is also $\mathcal G$-measurable. The agent with finer information can use the old rule, ignoring the additional distinctions. Minimization over a set containing all the old rules cannot have a greater value. In a finite problem a minimum exists, so no limiting argument is required. ◻

This is an expectation statement under one unchanged decision problem. Individual observations can make a posterior less confident; no pointwise monotonicity of posterior entropy is claimed. Nor does the proposition show that actual inquiry is free, that a bounded controller can calculate an optimal rule, or that a newly constrained agent retains every formerly available action. Memory, computation and acquisition costs can reverse a net-value comparison. The ideal cumulative filtration specifies retained access; actual forgetting or blocked reads require a different, possibly nonnested access description.

12.4 Reversible preservation is distinct from a memory arrow

A three-bit example isolates this last distinction. Let $X$ be fair and prepare memory registers $M=Q=0$. Apply a controlled NOT from $X$ to $M$, then swap $M$ and $Q$:

$$(X,0,0)\longmapsto(X,X,0)\longmapsto(X,0,X).$$

Both operations are permutations on the full eight-state space. At each stage the joint distribution has two equiprobable states, so its Shannon entropy is exactly one bit. Yet the information about $X$ in the designated current register $M$ rises from zero to one bit, then falls to zero. The information has moved to $Q$. A cumulative transcript including the earlier value of $M$ still determines $X$; this mathematical transcript must not be confused with the current $M$ register. A participant can use that earlier observation later only if an admitted archive or read route preserves access to it. Reversibility by itself supplies neither that route nor the ready boundary $M=Q=0$.

The recognition trace similarly presupposes prepared records, ordered probe and ingress ports, and a continuing archive. Its useful direction of acquisition is a property of that organization and boundary preparation, not a theorem that all reversible dynamics generate growing memories or a universal temporal arrow. Preservation of distinctions, current accessibility and current authority remain separate questions.

Finally, neither a physical history subsequence nor a cumulative filtration is automatically a phenomenal projection. Shadow Theory’s A0-A3 conditions apply separately to whatever phenomenal realization is proposed. The present model establishes a limited connection: within one completely specified physical history, a constituted participant can diagnose a source, retain its account, revise its practical use and remain uncertain about events outside its available evidence. It neither derives subjective temporal flow nor adds an ultimate author behind the transaction.

12.5 Prepared records, entropy and the practical direction of inquiry

The ready-memory condition in the last example is physically significant. A record-producing interaction does not ordinarily start from an arbitrary memory value and nevertheless end with an unambiguous copy. For an $m$-valued classical source $X$ and memory $M$ in the additive group $\mathbb Z_m$, the reversible write

$$(X,M)\longmapsto(X,M+X\ \mathrm{mod}\ m)$$

produces $M_{\rm out}=X$ on the ready set $M=0$. That set has $m$ states out of the $m^2$ states of the joint register space. With independent uniform $X$ and $M$, the output memory is instead independent of $X$: for each $x$, adding $x$ permutes a uniform distribution. The raw gate alone therefore does not make a usable source record.

For uniform $X$ and ready $M=0$, joint Shannon entropy is $\log_2m$ both before and after the write. The memory’s marginal entropy rises from zero to $\log_2m$, and its mutual information with $X$ rises by the same amount. The newly correlated record does not require a loss of global distinctions. It requires an appropriate preparation and a continuing read route. This general finite example is an elementary information calculation, not a universal thermodynamic equation or a discovery of the receiver-capacity theorem already retained in Appendix A.

Thermodynamic entropy adds coarse physical information beyond these logical register values. If a macrostate $M_{\rm phys}$ corresponds to phase region $\Gamma_{M_{\rm phys}}$, Boltzmann entropy has the form

$$S_B(M_{\rm phys})=k_B\log\!\left(\frac{\mu(\Gamma_{M_{\rm phys}})}{\mu_*}\right),$$

where $\mu_*$ fixes the reference units. A volume-preserving Hamiltonian flow does not change the volume of an evolved fine-grained set. It need not preserve the volume of the macroregion containing that set, nor make $S_B$ monotone along every trajectory. Those statements concern different sets. Low-entropy boundary conditions, a physically justified coarse description and an appropriate typicality measure can support a direction of overwhelmingly typical record formation. Their truth and applicability cannot be inferred from invertibility alone.

The relevance to the participant is concrete. Prepared sensory and memory resources, environmental correlations and dissipation can make earlier interactions available as robust records while later events have not yet left comparable records in the organization. The organization may anticipate those events through a model. Its anticipation can be accurate, mistaken or acted upon. An asymmetric distribution of usable records can thus contribute to the distinction between remembering and expecting without requiring an objective global moving present.

This is a conditional explanatory route rather than a proof that the universe’s entire thermodynamic arrow follows from one finite memory. An initial ready bank is an unusually restricted condition relative to a counting measure, but its physical probability and preparation cost depend on its apparatus. The circuit’s source preparation, power and holding obligations cannot be replaced by the entropy formula. Nor does a low-entropy boundary alone guarantee the absence of all anticipatory correlations or an agent’s available control. The inherited history-budget arguments and the current port calendar continue to constrain what information can be acquired and retained.

Ismael identifies an important remaining gap. Even granting a thermodynamic account under which present interventions are differently informative about past and future macrostates, that account does not yet show that choosing brings about an event rather than merely revealing it (Ismael 2016, 147-49). She adds the performative role of deciding: it is an activity through which a judgment or commitment is made, rather than just an observation of an independently formed attitude (Ismael 2016, 150-52). We retain both sides of that distinction. Record asymmetry explains part of situated access. The assessment and commit routes explain a particular operative contribution. Neither substitutes for the other.

The finite recognition history is stronger than an incremental viewing analogy in precisely this limited respect. Blocking its commit route preserves the diagnostic description but changes later action in the relevant $b=1$ cases. Removing the write route while retaining the report likewise separates knowing a source from revising its authority. These tests concern the realized causal organization, rather than a spectator discovering a finished output. They do not show ultimate authorship of the organization’s enabling conditions.

The resulting temporal account has three layers. Physical causal and clock relations order occurrences. Boundary preparation, access and retention explain a limited asymmetry of available records. Interpretation and assessment make some of those records matter to later commitments and conduct. A phenomenal temporal perspective requires the additional consciousness constitution and evidence connecting its organization to lived succession. Keeping the layers related but distinct lets the model explain genuine acquisition without pretending that records alone solve the experience of time.