[↗] GIDS series

preprint · part 2 · 3 of 7

Part 2: Deriving the Transcendental Embedding

Part 1 described how evolution can carve a finite repertoire of distinctions out of God’s Infinite Dimensional Space. That account explains why an organism has a “world” at all. It does not yet explain why one human inhabits that world differently from another human, even when both inherit roughly the same species-level template. Part 2 answers that question.

The Technical Scope (because otherwise I’ll accidentally lie to you)

Before I keep descending into Kantian hell, I need to pin down the scope so I do not smuggle Kant into the parts that are supposed to be engineering.

There are really three layers running through the rest of this paper.

First, there is the interpretive layer: the noumenal/phenomenal story that motivates why an observer should have structured experience at all.

Second, there is the predictive layer: the formal object the mathematics is actually allowed to touch. That object is not the whole ineffable mush of phenomenal life in itself, but a predictive representation that preserves the parts of an actor’s response surface needed for the tasks we care about.

Third, there is the control layer: once such a state can be estimated, we can rank or search over candidate propositions by their predicted effect on the actor’s next state and downstream objective. That is the whole point. But causal claims there require intervention-grade data, not just retrospective logs and me getting excited.

For this formalism, an actor is a bounded system for which we can specify persistent state, channels through which the environment reaches it, a rule or distribution by which state changes, and outward traces by which those changes become observable. Consciousness is not required for operational actorhood. It matters to the philosophical claim about experience; it is not required to ask whether a system has enough persistence and response structure to be predicted.

The draft up to this point used one name, Transcendental Embedding, for several different things at once. From here onward I separate them. The transition is easier to read under somewhat dubious pretension than by dumping everything on you, the reader, at once:

First, there is the inherited template: the repertoire of distinctions a human organism can in principle host inside their mental interior.

Second, there is the realized individual embedding: the weighting, coupling, and organization of that template in one person after development, language, culture, memory, and repeated experience.

Third, there is the phenomenal state at a time: the full lived condition of the organism now.

Fourth, there is the general predictive actor-state: the response structure that determines how the actor is expected to change under a family of possible propositions.

Fifth, there is the task-conditioned predictive state: the projection of that general response structure needed for one task and horizon.

Sixth, there is the estimated state: the low-resolution object we can compute from outward traces.

Part 2 is the transition from the first object to the other five.

One more boundary matters. The first actor class in this research program is the individual human. A corporation can later be treated as an actor, but it is not generated by copying the human equations and changing a label. A corporation has to be assembled from the people inside it, the organization of authority between them, its institutional memory, its incentives, its recorded facts and statistics, and the environment in which it exists. The same outer logic of state, proposition, transition, and trace can apply after the actor has been constructed. The construction itself is different.

Behold; You! The Chimera

Call the person-in-role object a Chimera. The term is mnemonic only. The theoretical work is done by the fact that a person is never encountered in the abstract, but always as a person under a role, inside an institution, in a regime, in a place, at a time. It would be computationally challenging if we did not introduce this categorization now, even though it amounts to a shortcut and a bastardization of the philosophical thesis.

If an alien biologist watched human outputs only, human life would look repetitive. Much of what humans do can be reduced, at the level of gross behavior, to self-preservation, courtship, reproduction, kin-bonding, status competition, alliance formation, and resource control. The outer patterns recur. The difficulty lies elsewhere. Human beings often arrive at similar outputs by different internal routes.

Practically, for what I, the author, am interested in—GTM engineering—imagine one founder rejects a product because of caution. Another rejects it because of fear. Another because the price signals weakness. Another because the pitch activated a prior bad memory. Another because the role they occupy requires public skepticism. Same output, different internal geometry.

That difference is the point of this section.

Let \(G_i\) denote the inherited seed available to person \(i\). This is an individual realization inside the species-compatible repertoire of possible distinctions.

Let \(T_{i,t}\) denote the realized individual Transcendental Embedding of person \(i\) at time \(t\): the relatively durable, slowly changing organization produced when the inherited seed develops through one life.

Let \(\phi_{i,t}\) denote the total phenomenal state of person \(i\) at time \(t\).

Let \(c_{i,t}\) denote the active role-and-institution context at time \(t\): founder, buyer, parent, employee, soldier, friend, plus the relevant company, market, group, and local demands.

We can then define the person-in-role object as

\chi_{i,t}=(T_{i,t},c_{i,t}).

This says something simple. The same person can yield different outputs across settings not because the person changes species, but because a different context activates a different organization of salience, inhibition, and available action. The person remains one person. The local geometry changes.

Those role-specific masks are highly specific to the individual and the regime rather than generic archetypes. They can preserve opposed dispositions long enough for the model to ask whether the opposition is a true contradiction or whether it resolves along a deeper axis activated by the regime.

Psychology and Factor Analysis

Psychology already contains rough tools for decomposition. Factor analysis took piles of correlated observations and produced constructs that were easier to name, compare, and compute. In a standard common-factor model,

\mathbf x_i = \boldsymbol\mu_x + \boldsymbol\Lambda_{\mathrm{FA}}\mathbf f_i + \boldsymbol\varepsilon_i, \qquad \mathbb E[\mathbf f_i]=\mathbf 0, \qquad \operatorname{Cov}(\mathbf f_i)=I, \qquad \mathbb E[\boldsymbol\varepsilon_i\mid\mathbf f_i]=\mathbf 0,

where \(\mathbf x_i\) is a vector of observed measurements, \(\boldsymbol\mu_x\) is the observed-variable mean, \(\mathbf f_i\) is a lower-dimensional latent-factor vector, \(\boldsymbol\Lambda_{\mathrm{FA}}\) contains factor loadings, and \(\boldsymbol\varepsilon_i\) is residual variation. If the measurements are centered, \(\boldsymbol\mu_x=\mathbf 0\).

Under the conventional standardized-factor identification above, the factor coordinates are not uniquely oriented. For an orthogonal matrix \(R\) satisfying \(R^{\top}R=I\),

\boldsymbol\Lambda_{\mathrm{FA}}\mathbf f_i = (\boldsymbol\Lambda_{\mathrm{FA}}R)(R^{\top}\mathbf f_i).

The common component is unchanged, and the standardized factor covariance remains the identity because \(R^{\top}R=I\). So the same fitted covariance structure can survive a rotation of the latent coordinates. Named constructs require identification conventions and interpretation; the mathematics does not hand us one sacred psychological axis.

This was useful. It also produced crude conglomerations.

A named personality construct is generally not one primitive feature of the mind. It is a statistical compression over many subtler tendencies: what the person notices, what they count as threatening, how quickly they generalize, how they respond to uncertainty, how they imagine other minds, how they trade status against safety, how strongly they carry prior events into the present, and so on. The factor is real in the practical sense that it summarizes repeatable covariance. It is not necessarily a fundamental axis of experience.

That is the basic way to understand what GIDS is trying to do here. We are doing something similar to the factor-analytic method, but we are trying to push beneath the crude named constructs and recover more granular vectors of how people understand and operate in the world. We also refuse to freeze the whole person into one timeless score. Some structure is durable. Some is activated only in one role. Some changes after one event. Some exists only as an interaction between the person and the proposition.

Write a psychometric proxy for person \(i\) as

\boldsymbol\psi_i\in\mathbb R^k,

where the coordinates may include IQ-like measures, psychometric traits, moral scales, behavioral factors, or related standardized summaries.

This object is useful, but it is not the Transcendental Embedding itself.

A factor score is not a memory field.
It is not a role.
It is not a present state.
It is not a proposition.
It is not a transition rule.

What it does provide is a coarse prior. It places a person inside a region of likely behavior. That is enough to matter, but not enough to solve micro-interaction. Factor analysis may tell us that a person is threat-sensitive, novelty-seeking, rigid, verbal, impulsive, or dutiful. It does not tell us whether those coordinates are active now, under this framing, with this memory already cued, inside this role.

A candidate feature deserves to be called more fundamental only after it repeatedly pays rent. It should remain useful across time when the person is stable, transfer across related tasks and contexts, interact sensibly with fast state, improve prediction beyond simpler constructs, and—when intervention data exists—respond to propositions in the direction the model predicts. Even then, the coordinate chart is not sacred. Rotations and other reparameterizations may preserve the same predictive information.

So psychology enters Part 2 as one source of approximation, not as the final ontology. To make the problem tractable from an engineering standpoint, we also need to append source-aware categorical traces to the estimation of a person’s Transcendental Embedding: role labels, objection classes, recurring topics, counterpart identities, action-types, firm-state tags, and other discrete observations that accumulate over time.

Those categorical traces are not the Transcendental Embedding in itself. They are an operational bridge from outward history to the estimated observer-side object. The important engineering choice is to avoid collapsing biography, stated language, observed behavior, and third-party inference into one immediate average merely because they sound semantically related. Some channels later become comparable; they should not be assumed comparable at ingestion.

this is very up to you on specific implementation details, tbh, be creative

Dimensionality Reduction, Attention, and Relevance

The issue is not that every decision has one permanent principal axis in the person. The problem compounds in difficulty because the individual organization contains many coordinates and relations, while any given transition is usually governed by a weighted subset of them.

Let \(x_t\) be the present proposition or stimulus. At the ideal level, define a task-conditioned relevance map

\Lambda_{\tau} \!\left( T_{i,t}, \phi_{i,t}, c_{i,t}, w_t, x_t \right) \in\mathbb R^{d_\tau},

where \(\tau\) indexes the task under study. The symbol \(\pi\) is deliberately not used here because it later denotes a decision policy.

Let attention or salience be represented by

\boldsymbol\lambda_{i,t}^{(\tau)}(x_t) \in[0,1]^{d_\tau}.

Then the ideal active proposition-conditioned slice is

\mathbf r_{i,t}^{(\tau)}(x_t) = \boldsymbol\lambda_{i,t}^{(\tau)}(x_t) \odot \Lambda_{\tau} \!\left( T_{i,t}, \phi_{i,t}, c_{i,t}, w_t, x_t \right),

where \(\odot\) denotes elementwise weighting.

This is the conceptual claim: the person may occupy a large space, but the next transition often depends on a smaller weighted slice of that space. The task is therefore not to discover one universal “main factor” of the person. The task is to discover which coordinates and interactions carry signal for a transition under a task, a proposition, and a context.

Attention matters because not all dimensions are weighted equally at every moment. The environment does not strike the whole embedding uniformly. A sentence, a person, a price, or a memory cue activates some coordinates and leaves others inert. That is why identical prompts can produce different outputs at different times.

The ideal expression is not directly computable because \(T_{i,t}\) and \(\phi_{i,t}\) are not observed. Later in this part, the learned approximation replaces them with the slow vector \(\widehat{\mathbf t}_{i,t}\), the fast vector \(\mathbf z_{i,t}\), and the explicit context carried by the operational state. The symbol \(\mathbf z_{i,t}\) is reserved for the fast state only; it is not reused for this proposition-conditioned slice.

The Notion of State

Philosophically, everything belongs inside state.

The phenomenal state includes what is perceived, what is remembered, what is felt, what is attended to, what is being done, what bodily changes are underway, and what action tendencies are presently live. In that sense, the state is total.

Let that total state be

\phi_{i,t}\in\Phi_i.

That is the full object.

If I leave the formal section aimed directly at \(\phi_{i,t}\), however, I start overclaiming almost immediately. So from here on I distinguish the motivating object from the objects the mathematics is actually allowed to touch.

Define the ideal pre-proposition information state

\mathsf I_{i,t} := \bigl( \mathcal H_{i,<t}, T_{i,t}, c_{i,t}, w_t \bigr),

where \(\mathcal H_{i,<t}\) contains the history available before the proposition at decision time \(t\), \(T_{i,t}\) is the slowly changing realized person structure, \(c_{i,t}\) is active context, and \(w_t\) is the relevant external world state. This is the ideal information bundle relative to which the prediction problem is defined; it is not the complete actor–world state \(\Sigma_{i,t}^{\star}\), because it does not grant the predictor direct access to the full phenomenal state. Assume the relevant state and outcome spaces are standard Borel spaces so that the conditional laws below exist in the ordinary regular sense.

The first predictive object is a general predictive actor-state, denoted

Q_{i,t}.

Intuitively, \(Q_{i,t}\) represents the actor’s response surface: everything in \(\mathsf I_{i,t}\) that still matters for how the actor would change under a family of admissible future propositions.

For a finite decision horizon \(H\), let

\mathbf x_{t:t+H-1} = (x_t,x_{t+1},\ldots,x_{t+H-1})

be an admissible open-loop proposition sequence. Let \(O_{i,t+1:t+H}\) denote the corresponding future observable trace sequence. Let \(\boldsymbol\Xi_{t+1:t+H}\) denote the random exogenous path and \(\boldsymbol\xi_{t+1:t+H}\) a supplied scenario value, including market, personnel, company, and role changes.

Choose a version of the observational regular conditional response kernel

\mathscr R_{i,t}^{(H),\mathrm{obs}} \!\left( \cdot \mid \mathsf I_{i,t}, \mathbf x_{t:t+H-1}, \boldsymbol\xi_{t+1:t+H} \right) := \mathcal L \!\left( O_{i,t+1:t+H} \mid \mathsf I_{i,t}, \mathbf X_{t:t+H-1}=\mathbf x_{t:t+H-1}, \boldsymbol\Xi_{t+1:t+H}=\boldsymbol\xi_{t+1:t+H} \right).

This conditional kernel is defined only up to almost-sure equality and is operationally relevant on the support of the proposition and scenario process. If future propositions are adaptive, the factual joint law under the historical policy must include the policy and candidate-set process that generated them. Merely inserting a new policy into the conditioning index does not identify its counterfactual response law from observational data, even when its individual actions have support. A counterfactual adaptive policy requires controlled transition kernels plus an identification argument, or an explicit extrapolation assumption. A causal response surface uses the corresponding interventional laws under proposition interventions rather than merely conditioning on historically selected propositions.

Two ideal information states are equivalent when they induce the same declared family of response laws for every admissible proposition sequence or policy, scenario regime, planning horizon, and measurable future event in the family being studied. This equivalence relation defines the ideal predictive information object abstractly. It is not automatic that the quotient by this relation inherits a convenient standard-Borel measurable structure. Whenever a usable measurable realization exists, denote that realization by \(Q_{i,t}\). Otherwise, the indexed family of response kernels itself is the predictive object, and no finite or conveniently measurable quotient is claimed. The scenario path is explicit because a recursive model does not earn the right to freeze the future environment merely for mathematical convenience.

This object may be infinite-dimensional. That is fine. The ideal response surface can be much larger than the finite approximation we learn.

The second predictive object is a task-conditioned summary of that state. For task \(\tau\) and elapsed-time outcome horizon \(\Delta\), write

q_{i,t}^{(\tau,\Delta)} = \Pi_{\tau,\Delta}(Q_{i,t})

when a measurable map with the required sufficiency property exists. \(\Pi_{\tau,\Delta}\) need not be linear or orthogonal. It is called a projection only in the loose sense that it keeps what one task and horizon require while discarding the rest.

Let

Y_{i,t}^{(\tau,\Delta)}

denote the outcome associated with the decision at time \(t\), evaluated at or by horizon \(\Delta\): reply, meeting, objection class, delay bucket, stage advance, sentiment shift, or whatever the task actually cares about. Writing the outcome this way prevents a multi-horizon decision row from pretending that every target occurs at one common timestamp.

For every measurable outcome set \(B\), the task-conditioned state is sufficient when, for each admissible proposition value \(x\) in the predictive support,

\mathbb P \!\left( Y_{i,t}^{(\tau,\Delta)}\in B \mid \mathsf I_{i,t}, X_t=x \right) = \mathbb P \!\left( Y_{i,t}^{(\tau,\Delta)}\in B \mid q_{i,t}^{(\tau,\Delta)}, X_t=x \right).

Read that in English: once the same proposition value is conditioned on, everything in the ideal information state that still matters for this future has already been compressed into \(q_{i,t}^{(\tau,\Delta)}\). The causal version replaces this observational conditioning statement with an intervention-indexed one.

At this point the expression is predictive, not automatically causal. If propositions in the historical data were selected in a biased way, the learned conditional law is a forecasting law. A causal predictive state would have to preserve the corresponding interventional laws under \(\operatorname{do}(X_t=x)\), which requires the experimental machinery introduced later.

Observable Predictive State

For engineering work, the key move is to define the formal state in terms of observable consequences rather than inaccessible total interiority.

The hierarchy is:

\phi_{i,t} \quad\text{full phenomenal state},

Q_{i,t} \quad\text{general predictive response state},

q_{i,t}^{(\tau,\Delta)} \quad\text{task-and-horizon summary},

and

\widehat s_{i,t} \quad\text{measurable operational approximation}.

The paper gets much more honest the second these stop being treated as the same thing.

The operational approximation I actually want to estimate is

\widehat s_{i,t} = \left( \widehat{\mathbf t}_{i,t}, \mathbf z_{i,t}, \mathbf c_{i,t}, \mathbf w_t \right),

where \(\widehat{\mathbf t}_{i,t}\in\mathbb R^{d_T}\) is the slow estimated person vector, \(\mathbf z_{i,t}\in\mathbb R^{d_Z}\) is the fast latent vector inferred from records available before \(t\), and \(\mathbf c_{i,t}\) and \(\mathbf w_t\) are finite encodings of context and world state.

The old notation \(\widehat s_{i,t}\approx q_{i,t}^{(\tau,\Delta)}\) was too vague. Approximation needs an operational meaning, and representation error must be separated from model-fitting error.

If \(\widehat s_{i,t}\) is a measurable function of \(\mathsf I_{i,t}\), define the state-compression gap

\epsilon_{\tau,\Delta}^{\mathrm{state}}(\widehat s) := I \!\left( Y_{i,t}^{(\tau,\Delta)}; \mathsf I_{i,t} \mid \widehat s_{i,t},X_t \right).

Equivalently,

\epsilon_{\tau,\Delta}^{\mathrm{state}}(\widehat s) = \mathbb E \!\left[ D_{\mathrm{KL}} \!\left( \mathbb P(Y\in\cdot\mid\mathsf I_{i,t},X_t) \;\middle\|\; \mathbb P(Y\in\cdot\mid \widehat s_{i,t},X_t) \right) \right],

where \(Y=Y_{i,t}^{(\tau,\Delta)}\). This term is zero exactly when the operational state is sufficient under the joint observational law and its supported propositions, up to null sets. Uniform or causal sufficiency across interventions requires the corresponding family of interventional laws.

A fitted conditional law \(P_{Y,\theta,\tau,\Delta}(\cdot\mid\widehat s,X)\) introduces a separate model-estimation gap:

\epsilon_{\theta,\tau,\Delta}^{\mathrm{model}}(\widehat s) := \mathbb E \!\left[ D_{\mathrm{KL}} \!\left( \mathbb P(Y\in\cdot\mid \widehat s_{i,t},X_t) \;\middle\|\; P_{Y,\theta,\tau,\Delta}(\cdot\mid \widehat s_{i,t},X_t) \right) \right].

To state the log-score identity without quietly assuming a discrete outcome, suppose the relevant conditional laws are dominated by one fixed reference measure on the outcome space. Let

p^{\star}_{\tau,\Delta}(y\mid\mathsf I_{i,t},X_t)

be a version of the full-information conditional density or mass function, let

p_{\widehat s,\tau,\Delta}(y\mid\widehat s_{i,t},X_t)

be the true conditional density or mass function after compression, and let

p_{Y,\theta,\tau,\Delta}(y\mid\widehat s_{i,t},X_t)

be the fitted density or mass function. Define the full-information Bayes log risk

\mathscr R_{\log}^{\star} := \mathbb E\!\left[ -\log p^{\star}_{\tau,\Delta} \!\left( Y_{i,t}^{(\tau,\Delta)} \mid \mathsf I_{i,t},X_t \right) \right].

Whenever the displayed expectations and divergences are finite, the exact decomposition is

\begin{aligned} &\mathbb E\!\left[ -\log p_{Y,\theta,\tau,\Delta} \!\left( Y_{i,t}^{(\tau,\Delta)} \mid \widehat s_{i,t},X_t \right) \right] - \mathscr R_{\log}^{\star} \\[4pt] &\qquad= \epsilon_{\tau,\Delta}^{\mathrm{state}}(\widehat s) + \epsilon_{\theta,\tau,\Delta}^{\mathrm{model}}(\widehat s). \end{aligned}

For a discrete outcome with counting measure, \(\mathscr R_{\log}^{\star}=H(Y_{i,t}^{(\tau,\Delta)}\mid\mathsf I_{i,t},X_t)\). For continuous outcomes it is a conditional expected negative log density, not an invariant “entropy of the state.” The first gap is information discarded by compression; the second is error in the fitted law conditional on that compressed state. Ordinary held-out log loss estimates total predictive risk, which also contains the irreducible full-information Bayes risk; differences against sufficiently strong references and targeted ablations are needed to diagnose whether failure came from state compression or from the predictor built on top of it.

The general state \(Q_{i,t}\) is supposed to support many task summaries. That is how the claim that “the ontology is the product” becomes testable. A person representation that predicts only one narrow label may simply be a task shortcut. A better actor-state should transfer across related outcomes, horizons, roles, and proposition families.

Memory as a Series of Vectors

For present purposes, memory need not be treated as narrative first. It can be modeled as a field of traces with weights.

Let

\mathbf m_{i,t}^{\mathrm{mem}} = \sum_{j=1}^{N_i} \varpi_{ij,t}\mathbf h_{ij}^{\mathrm{mem}},

where \(\mathbf h_{ij}^{\mathrm{mem}}\) is a stored trace representation and \(\varpi_{ij,t}\) is its weight at time \(t\).

Some traces are weak. Some are strong. Some decay. Some reactivate under similarity, emotion, role, or repetition. The point is not that memory is literally a vector sum in the brain; the point is that weighted trace structure gives us a tractable model of persistence and retrieval.

A present proposition does not encounter the whole memory field evenly. Before the response is observed, proposition-conditioned retrieval can be written schematically as

\widetilde\varpi_{ij,t}(x_t) = \mathcal R_{\mathrm{ret},\theta} \!\left( \varpi_{ij,t}, \mathbf h_{ij}^{\mathrm{mem}}, \widehat s_{i,t}, x_t \right).

The proposition-conditioned retrieved memory is then

\widetilde{\mathbf m}_{i,t}^{\mathrm{mem}}(x_t) = \sum_{j=1}^{N_i} \widetilde\varpi_{ij,t}(x_t) \mathbf h_{ij}^{\mathrm{mem}}.

After the actual trace \(o_{i,t+1}\) arrives, the memory state can be updated by

\mathbf m_{i,t+1}^{\mathrm{mem}} = U_{\mathrm{mem},\theta} \!\left( \mathbf m_{i,t}^{\mathrm{mem}}, x_t, o_{i,t+1}, \mathbf c_{i,t} \right).

This gives memory two jobs.

First, it stores prior traces.
Second, it changes which parts of the person-space are active now.

This also explains why repeated prompts can produce different outputs. The second encounter is not with the same person-state as the first. The first encounter has already changed the trace structure. A prior positive or negative experience can therefore make the second prediction harder, not easier, if the model fails to represent that update.

Recommendation systems provide a useful analogy here. A view history is not a mind. But a recommender does show the core move: repeated traces can be compressed into a latent representation that improves prediction. In the present framework, biography, language, preferences, recurrent actions, role history, and prior interactions play the role of trace data from which a person-level state can be estimated.

Categorical Trace Pooling as an Operational Memory Estimator

A large part of what we observe about people arrives in categorical form: role labels, recurring topics, objection classes, counterpart identities, action-types, domain tags, product themes, price postures, and other discrete markers. Rather than treat these as dead one-hot tables or discard them into prose, we can embed them and pool them over time.

Operationally, the implementation assumes a fixed global registry of categorical families, source channels, and admissible regimes, with learned null vectors and mask bits for absent cells so the resulting representation stays fixed-width across people and time.

Let \(f\in{1,\ldots,F}\) index categorical families and let \(\sigma\in\mathcal S_{\mathrm{src}}\) index source channels such as biography, stated language, observed behavior, and third-party or inferred traces. For person \(i\), the notation \(r<t\) below means that record \(r\) was timestamped and available before decision epoch \(t\). Let

\mathcal B_{i,r}^{(f,\sigma)}

be the multiset of raw category tokens in family \(f\) from source \(\sigma\).

Before pooling, surface labels should be contextually typed. A token that looks contradictory in the raw may become perfectly consistent once we mark whether it concerns self versus other, in-group versus out-group, own-interest versus third-party interest, formal stance versus enacted stance, or another asymmetry carried by the regime. Write this contextual lifting as

\widetilde{\mathcal B}_{i,r}^{(f,\sigma)} = \operatorname{Lift}_{\mathrm{ctx}}\!\left( \mathcal B_{i,r}^{(f,\sigma)},c_{i,r} \right).

Only the opposition that remains after this lifting deserves to be treated as a genuine contradiction.

Let

E_{f,\sigma}:\mathcal V_{f,\sigma}\to\mathbb R^{d_{f,\sigma}}

be the embedding table for that family and source, and let \(\mathbf e_{\varnothing}^{(f,\sigma)}\in\mathbb R^{d_{f,\sigma}}\) be a learned empty-bag vector. The event-level pooled representation is

\mathbf u_{i,r}^{(f,\sigma)} = \begin{cases} \displaystyle \frac{1}{|\widetilde{\mathcal B}_{i,r}^{(f,\sigma)}|} \sum_{\upsilon\in\widetilde{\mathcal B}_{i,r}^{(f,\sigma)}} E_{f,\sigma}(\upsilon), & |\widetilde{\mathcal B}_{i,r}^{(f,\sigma)}|>0,\\[10pt] \mathbf e_{\varnothing}^{(f,\sigma)}, & |\widetilde{\mathcal B}_{i,r}^{(f,\sigma)}|=0. \end{cases}

Define the binary availability mask

M_{i,r}^{(f,\sigma)} = \mathbf 1\{|\widetilde{\mathcal B}_{i,r}^{(f,\sigma)}|>0\}.

Let \(W_{f,\sigma}^{\mathrm{cat}}:\mathbb R^{d_{f,\sigma}}\to\mathbb R^{d_*}\) be a learned alignment map that gives every slot a common output width while preserving family and source identity. Concatenate the cells:

\mathbf e_{i,r}^{\mathrm{cat}} = \big\|_{(f,\sigma)} \left[ W_{f,\sigma}^{\mathrm{cat}}\mathbf u_{i,r}^{(f,\sigma)}, M_{i,r}^{(f,\sigma)}, \log\!\left(1+|\widetilde{\mathcal B}_{i,r}^{(f,\sigma)}|\right) \right].

This is the recommender move in its simplest form: sparse categorical IDs are mapped to dense vectors and multivalent bags are pooled into fixed-width representations. The count term prevents an event containing one token from becoming automatically identical to an event containing the same token many times. But I do not want the next step to be a naive global average across every source and every role. Categories that arrive through speech, biography, and behavior are not automatically the same thing just because they share a label. They can later become comparable; they should not be forced into comparability at ingestion.

Let

\rho_r:=\operatorname{role}(c_{i,r})

denote the regime or role at event \(r\). Let the slow weights satisfy \(\beta_{i,r,t}^{\mathrm{slow}}\ge 0\), and define the slow evidence mass

\beta_{i,\rho,+,t}^{\mathrm{slow}} := \sum_{r<t} \mathbf 1\{\rho_r=\rho\} \beta_{i,r,t}^{\mathrm{slow}}.

The slow categorical memory for regime \(\rho\) is

\mathbf g_{i,\rho,t}^{\mathrm{slow}} = \begin{cases} \displaystyle \frac{ \sum_{r<t} \mathbf 1\{\rho_r=\rho\} \beta_{i,r,t}^{\mathrm{slow}} \mathbf e_{i,r}^{\mathrm{cat}} }{\beta_{i,\rho,+,t}^{\mathrm{slow}}}, & \beta_{i,\rho,+,t}^{\mathrm{slow}}>0,\\[14pt] \mathbf e_{\varnothing,\rho}^{\mathrm{slow}}, & \beta_{i,\rho,+,t}^{\mathrm{slow}}=0, \end{cases}

where \(\mathbf e_{\varnothing,\rho}^{\mathrm{slow}}\) is a learned empty-regime vector. Define the availability mask

M_{i,\rho,t}^{\mathrm{slow}} = \mathbf 1\{\beta_{i,\rho,+,t}^{\mathrm{slow}}>0\}.

The complete slow bank retains the pooled content, explicit availability, and total evidence mass:

\mathbf g_{i,t}^{\mathrm{slow}} = \big\|_\rho \left[ \mathbf g_{i,\rho,t}^{\mathrm{slow}}, M_{i,\rho,t}^{\mathrm{slow}}, \log\!\left(1+\beta_{i,\rho,+,t}^{\mathrm{slow}}\right) \right].

For task-conditioned retrieval from recent categorical history, let the nonnegative relevance weights satisfy \(\alpha_{i,r,t}^{(\tau)}\ge 0\), and define

\alpha_{i,+,t}^{(\tau)} = \sum_{r<t}\alpha_{i,r,t}^{(\tau)}.

The normalized content summary is

\overline{\mathbf g}_{i,t}^{\mathrm{fast},\tau} = \begin{cases} \displaystyle \frac{1}{\alpha_{i,+,t}^{(\tau)}} \sum_{r<t} \alpha_{i,r,t}^{(\tau)} \mathbf e_{i,r}^{\mathrm{cat}}, & \alpha_{i,+,t}^{(\tau)}>0,\\[12pt] \mathbf e_{\varnothing,\tau}^{\mathrm{fast}}, & \alpha_{i,+,t}^{(\tau)}=0. \end{cases}

Then retain both content and accumulated evidence:

\mathbf g_{i,t}^{\mathrm{fast},\tau} = \left[ \overline{\mathbf g}_{i,t}^{\mathrm{fast},\tau}, \log\!\left(1+\alpha_{i,+,t}^{(\tau)}\right) \right].

This is a task-conditioned retrieval from recent categorical history, not the fast state itself. The shared fast state \(\mathbf z_{i,t}\) is updated chronologically from observed events. The retrieved pool may then enter the task-relevance map that asks which parts of that shared state matter now. Keeping the relevance mass separate is necessary because a normalized average alone cannot distinguish one weak exposure from the same weak exposure repeated twenty times.

The weighting laws should not treat every trace equally. They may depend on recency, task relevance, regime relevance, action intensity, repeated weak exposure, susceptibility, and source reliability. Decisive action traces should often outrank passive exposure traces, while repeated weak exposures should still accumulate over time according to the person’s susceptibility.

This is also where strategic self-presentation enters the model: stated concern, biographical prior, and observed behavior are allowed to disagree without being collapsed at ingestion.

The slow categorical bank is therefore best read as what the person is generally like now, at this stage and across regimes, rather than as a timeless essence. The fast pool captures what is currently active for a task. Neither is the phenomenal state itself.

Minimality, Identifiability, and Slow/Fast Factorization

Now comes the part that keeps the whole thing from dissolving into vibes.

A predictive state is not interesting merely because it is sufficient. A gigantic archive is sufficient too. The point is to retain the information the future still cares about without dragging the whole archive behind every prediction forever.

Say that a task-conditioned predictive state \(q_{i,t}^{(\tau,\Delta)}\) is minimal when, for any other sufficient state \(r_{i,t}^{(\tau,\Delta)}\), there exists a measurable map \(h\) such that

q_{i,t}^{(\tau,\Delta)} = h\!\left(r_{i,t}^{(\tau,\Delta)}\right)

almost surely.

This is the right level of humility. I am not claiming there is one mystical coordinate chart for the soul. I am claiming that, for a task, there may be a smallest predictive information object. Any two minimal sufficient states generate the same predictive sigma-field up to null sets; under standard Borel regularity they can be related by measurable inverse maps on full-measure subsets. The manuscript does not assume that such a finite-dimensional minimal state exists for every possible task; that is an empirical and structural question.

The operational state is split into slow and fast pieces:

\widehat s_{i,t} = \left( \widehat{\mathbf t}_{i,t}, \mathbf z_{i,t}, \mathbf c_{i,t}, \mathbf w_t \right).

Let \(\mathbf u_{i,t}^{\mathrm{dur}}\) collect durable person-side proxies, including the slow categorical bank. The slow estimate is

\widehat{\mathbf t}_{i,t} = E_{T,\theta}\!\left( \mathbf u_{i,t}^{\mathrm{dur}}, \mathcal H_{i,<t}^{\mathrm{durable}} \right).

Let \((\mathsf h_n)_{n\ge1}\) be the chronological record sequence, with \(v_n:=\operatorname{time}(\mathsf h_n)\). When the stream is shared across actors, let \(\mathbf d_{i,n}^{\mathrm{rec}}\) record whether record \(n\) pertains to actor \(i\) and which actor-specific fields are available. Starting from \(\mathbf z_{i,0}\), the fast state after processing record \(n\) is

\mathbf z_{i,n} = U_{z,\theta}\!\left( \mathbf z_{i,n-1}, \widehat{\mathbf t}_{i}(v_n^-), \mathbf c_i(v_n^-), \mathbf w(v_n^-), \mathsf h_n, \mathbf d_{i,n}^{\mathrm{rec}} \right), \qquad n\ge1.

The applicability vector is not a memory vector and not an outcome-availability mask. It contains a binary applicability flag, and when that flag is zero the update is required to leave the actor’s fast state unchanged.

At decision epoch \(t\), let

N(t):=\#\{n:v_n<\operatorname{time}(\mathsf d_t)\}, \qquad \mathbf z_{i,t}:=\mathbf z_{i,N(t)}.

This converts the asynchronous record clock into the decision-aligned state used by the model.

The reason for the split is not aesthetic. Different things change on different timescales. A founder does not become a different founder because of one email. But their local state can absolutely change because of one email. The slow term is supposed to carry durable person structure. The fast term is supposed to carry within-window state needed to preserve the predictive content of recent history.

If this split is real, removing \(\mathbf z_{i,t}\) should hurt short-horizon prediction. Removing \(\widehat{\mathbf t}_{i,t}\) should hurt cold-start performance and cross-context generalization. If neither happens, I do not get to pretend the decomposition was profound. Part 4 will force that issue.

Identifiability remains limited. A learned latent space can be rotated, rescaled, or otherwise reparameterized without changing its predictions. The empirical target is therefore not one sacred coordinate system. It is stable predictive information, recoverable structure, transfer, and intervention behavior.

Deriving the Transcendental Embedding

We can now write the distinction that Part 1 left implicit.

Let \(G_i\) be the inherited seed: the individual lineage-compatible repertoire and organization available at the beginning of development.

Let

T_{i,t} = \mathcal E_{\mathrm{ind}}\!\left( G_i, \mathbf u_{i,<t}^{\mathrm{lang}}, \mathcal H_{i,<t}^{\mathrm{life}} \right)

be the realized individual Transcendental Embedding at time \(t\), where \(\mathbf u_{i,<t}^{\mathrm{lang}}\) denotes language, culture, and socialization recorded before decision \(t\), and \(\mathcal H_{i,<t}^{\mathrm{life}}\) is life history recorded before that decision.

Life history is fundamentally a sequence, not a bag. For a crude summary, one may write

\overline{\mathbf h}_{i,t} = \sum_{r<t} \beta_{i,r,t}\mathbf v_{i,r}^{\mathrm{event}}, \qquad \beta_{i,r,t}\ge 0,

where \(\mathbf v_{i,r}^{\mathrm{event}}\) is an event representation and \(\beta_{i,r,t}\) is its later weight. The sequence model should still retain order when order matters.

Some events contribute little.
Some events bend the later space of response.

This is the theoretical object.

But we do not observe \(T_{i,t}\) directly. What we observe are traces and proxies. Let

\(\boldsymbol\psi_i\) be psychometric and cognitive summaries;
\(\mathbf b_i\) be biography and background;
\(\mathbf u_{i,t}^{\mathrm{lang}}\) be the language and cultural-position features available before decision \(t\);
\(\mathbf v_{i,t}^{\mathrm{role}}\) be role and institution history features;
\(\overline{\mathbf h}_{i,t}\) be a weighted summary of observable life-event structure;
\(\mathbf g_{i,t}^{\mathrm{slow}}\) be the slow source-aware and regime-aware categorical trace bank.

Collect these in

\mathbf u_{i,t}^{\mathrm{dur}} = \left[ \boldsymbol\psi_i, \mathbf b_i, \mathbf u_{i,t}^{\mathrm{lang}}, \mathbf v_{i,t}^{\mathrm{role}}, \overline{\mathbf h}_{i,t}, \mathbf g_{i,t}^{\mathrm{slow}} \right].

Then a first operational estimate is

\widehat{\mathbf t}_{i,t}^{(0)} = E_{T,\theta}^{(0)}\!\left( \mathbf u_{i,t}^{\mathrm{dur}} \right).

At first pass, this initial encoder can be additive:

\widehat{\mathbf t}_{i,t}^{(0)} = W_\psi\boldsymbol\psi_i + W_b\mathbf b_i + W_{\mathrm{lang}}\mathbf u_{i,t}^{\mathrm{lang}} + W_{\mathrm{role}}\mathbf v_{i,t}^{\mathrm{role}} + W_h\overline{\mathbf h}_{i,t} + W_g\mathbf g_{i,t}^{\mathrm{slow}}.

Each matrix maps its input into the common slow-state dimension \(d_T\). If one channel suppresses or inverts another, that sign belongs inside the learned operator rather than being hard-coded as a minus sign on the whole source.

This weighted estimate is not the Transcendental Embedding itself. It is the tractable object from which we begin. It is a low-resolution approximation built from standardized signals because those are the signals computation can access at scale.

As chronological records become available, update the record-indexed fast state by the rule above and use

\mathbf z_{i,t}=\mathbf z_{i,N(t)}

as the state available at decision epoch \(t\).

Once the slow estimate and fast state are available, the local operational object is

\widehat s_{i,t} = \left( \widehat{\mathbf t}_{i,t}, \mathbf z_{i,t}, \mathbf c_{i,t}, \mathbf w_t \right).

Given a proposition \(x_t\), the learned proposition-conditioned active representation is

\widehat{\mathbf r}_{i,t}^{(\tau)}(x_t) = \boldsymbol\lambda_{\theta,\tau} \!\left( \widehat s_{i,t},x_t,\mathbf g_{i,t}^{\mathrm{fast},\tau} \right) \odot \Lambda_{\theta,\tau} \!\left( \widehat s_{i,t}, x_t, \mathbf g_{i,t}^{\mathrm{fast},\tau} \right),

where the learned salience map returns a vector in \([0,1]^{d_\tau}\) and the learned relevance map returns a vector in \(\mathbb R^{d_\tau}\). This active slice is an interaction feature. It is not the fast state \(\mathbf z_{i,t}\), and it is not the theoretical sufficient statistic \(q_{i,t}^{(\tau,\Delta)}\).

Let \(\boldsymbol\Xi_{t+1}\) denote the random exogenous change between decision epochs and \(\boldsymbol\xi_{t+1}\) a realized or supplied scenario value. Evaluating the parameterized transition kernel at \(\boldsymbol\xi_{t+1}\) defines a scenario forecast; it does not require a positive-probability singleton event. An unconditional forecast must instead integrate the conditional transition kernel against a declared law for \(\boldsymbol\Xi_{t+1}\). The recursively usable learned model evolves the operational state:

\widetilde s_{i,t+1} \sim K_{s,\theta} \!\left( \cdot \mid \widehat s_{i,t}, X_t=x_t, \boldsymbol\Xi_{t+1}=\boldsymbol\xi_{t+1} \right).

A delayed task outcome has a separate predictive law,

P_{Y,\theta,\tau,\Delta} \!\left( \cdot \mid \widehat s_{i,t}, X_t=x_t \right).

For a one-step task this law may be read from the next state and immediate observation. For \(\Delta\) extending beyond the next decision epoch, the law is defined only relative to a declared continuation policy, future candidate-set process, exogenous-path regime, and censoring convention, unless those variables are conditioned on directly. It may be implemented by a rollout or by a direct head calibrated to that same regime. It must not be presented as if a 90-day outcome were an immediate emission from one next-state sample.

Part 2 stops here on purpose.

We cannot yet claim to know the true form of the transition kernel.
In no way am I claiming to have solved qualia or consciousness; I have just, maybe, created a representation that keeps me from talking nonsense while trying to predict human transition.
It does not claim that the estimate and the reality itself are identical.

What it does claim is smaller and enough for the next step: a person can be represented as a latent structure derived from an inherited seed, development, language, culture, memory, and repeated life events; that structure can be estimated from outward traces; and the estimate can serve as the person-side state in a transition model whose recursively modeled target is future operational state and whose task heads predict observable behavior.

Part 3 can now ask the narrower question: once a person has been represented by \(\widehat{\mathbf t}_{i,t}\), once recent dynamics have been represented by \(\mathbf z_{i,t}\), and once the person-in-role state has been represented by \(\widehat s_{i,t}\), how do we represent the proposition without pretending the proposition is the same kind of thing as the person, and how do we learn a transition map that predicts what happens next with increasing fidelity?