the mathematics.

§0. What this document is

What follows is the full mathematical apparatus of the closed-loop libidinal economy. The model is not a metaphor borrowed from dynamical-systems theory to dignify a sociological claim. It is a working dynamical system in which each equation does specific argumentative labor — separating what is consequence from what is assumption, isolating what is structural from what is contingent, and locating where the framework's central result (the reclamation theorem) becomes inevitable rather than merely plausible. The formalism is not the theory; the theory is what the formalism makes possible to say without sliding back into the soft register where almost any claim can be advanced and almost no claim can be refused.

The exposition proceeds from foundations outward. First, the six equations that constitute the closed loop, followed by the explicit assumptions on which subsequent results rest. Then the variable index, organized by domain. Then the six instruments — each a formal object with its own support, its own scope, its own falsifiability conditions. Then the foreclosure lemma and the reclamation theorem, with full proofs. Then the three modes of intervention — Mode A (architectural), Mode B (artisanal), Mode C (disruption-as-form) — each stated as a formal proposition with proof. Then the framework's other named results, which are not consequences of the theorem but parallel claims resting on the same architectural fact. Then an honest scope statement: what the model does not claim, where the formalism reaches its limits, and where the work remains. The document closes with the open formal questions that remain, in order of how much weight they carry.

§1. The closed-loop dynamics

Let $t \in [0, \infty)$ denote continuous time, with the platform observing and acting at discrete tick boundaries separated by $\delta_{\min} > 0$ — the minimum technical inter-stimulus latency. State, observation, and policy variables are piecewise-constant between ticks, with updates happening at tick boundaries. The marked point process of engagement events is supported on the tick grid. This multiscale structure — continuous time for the marked point process, discrete tick boundaries for the dynamics — is implicit in all equations below; specifically, the difference equation in (1) is shorthand for $u_{(k+1)\delta_{\min}} = f(u_{k\delta_{\min}}, s_{k\delta_{\min}}) + \xi_{k}$, and the integral in (5) is well-defined as a sum over events at tick boundaries.

The user state at time t is a vector $u_{t}$ in a state space U, decomposed into four registers:

The registers — Subject, Citizen, Person, Consumer — are not selves but channels of address. They correspond to the four apparatuses historically distinct (school and archive, newspaper and parliament, church and family, marketplace) that the closed loop now routes through a single delivery surface. The state evolves according to:

(1) State evolution.

where $s_{t}$ is the stimulus delivered to the user at time t, f is the user's response dynamics (which may be nonlinear, register-coupled, and partially unknown to the user), and $\xi_{t}$ is process noise representing the residual variability not captured by the deterministic response — what the framework calls the leak channel, since it is the only stochastic admission of anything outside the loop.

(2) Engagement signal.

The platform does not observe $u_{t}$ directly. It observes an engagement signal $y_{t}$ — a tuple of measurable behavioral outputs (taps, scrolls, watch-time, gaze, transaction events) — produced by an observation function g that compresses the high-dimensional internal state into a low-dimensional behavioral projection, with measurement noise $\eta_{t}$. The compression is irreversible: distinct internal states can produce identical $y_{t}$. This irreversibility is not a bug; it is the precondition for the platform's optimization.

(3) Delivery policy.

The stimulus $s_{t}$ is chosen by a policy $\pi$ parameterized by $\theta_{t}$ (the predictive parameters maintained by the platform). The policy can depend only on the history $y_{1:t}$ of engagement observations, since by (2) the platform has no direct access to $u_{t}$. The policy implicitly maintains a belief state $b_{t}(\cdot) := p(u_{t} | y_{1:t}; \theta_{t})$ — the platform's posterior over the user's full state given the engagement history, updated by Bayes' rule under the observation model (2) — and selects $s_{t}$ to maximize an expected objective under that belief. The same object appears in §3.3 (dividual condition: $H(b_{t}) \to 0$) and §5.3 (Mode C: maintaining $H(b_{t}) \ge H_{\mathrm{res}} > 0$). Equation (3) is what makes the loop closed: $s_{t}$ depends on prior behavioral output, which depended on prior $s_{t-1}$, recursively, with no exogenous input that the policy does not itself shape.

(4) Online learning.

The predictive parameters evolve by stochastic gradient ascent on the expected platform reward $\hat{R}$, with learning rate $\alpha$. The hat denotes a stochastic estimator of the policy gradient — the most common in practice is the REINFORCE form $\hat{R}(y_{t})\cdot\nabla_{\theta}\mathrm{log} \pi(s_{t}|y_{1:t}; \theta_{t})$, but the framework's results hold for any unbiased estimator with bounded second moment. $\hat{R}$ is the empirical proxy for engagement that the platform actually optimizes (session length, revisit rate, ad impression value, retention curves); it is not a normative quantity. The framework does not assume any particular $\hat{R}$; the results below hold under the structural condition (Assumption A2 in §1.5) that $\hat{R}$ is non-decreasing in the engagement-event count over the operating support of $\pi$.

(5) Surplus extraction.

The platform's accumulated yield over horizon T is given by a marked point process. $N_{t}^{i}$ is the counting process for engagement events marked with register i (a Citizen-marked event is a political-content tap; a Consumer-marked event is a transactional swipe; the marks are inferred from content metadata, not from the user). The extraction kernel $\rho_{i}$ is a predictable process — evaluated at the left limits $u_{t^{-}}, y_{t^{-}}$ immediately before each jump — assumed measurable and bounded. Surplus splits per register:

This is not a derivative of (1)–(4); it is the loop's purpose. Without (5), the loop is autonomic. With (5), the loop is libidinal-economic.

(6) Foreclosure of the exogenous channel.

The model has no exogenous input. The user's state evolution in (1) is driven entirely by the policy-shaped stimulus $s_{t}$ and the residual noise $\xi_{t}$. Equation (6) is a definitional fact about the model's architecture — the loop is closed by construction. The substantive propositional claim, that the survival function of no-input windows under the policy's stationary behavior decays to zero on the relevant timescales, is not asserted here. It is derived from (1)–(5) as Lemma 1 in §4 below.

Equations (1)–(6) constitute the closed loop. They are not six separate claims. They are a single architecture, and every result in the framework descends from this architecture or makes structural conditions on it. The architecture is rendered as a directed graph at /the-dynamics/closed-loop.

§1.5. Assumptions

The framework's results depend on six regularity assumptions. They are stated here explicitly so that what is foundational is separated from what is derivable, and so that the falsifiability conditions are formally identifiable.

(A1) Measurability and integrability. The response dynamics $f : U × S \to U$ and the observation function $g : U \to Y$ are measurable. The noise processes $\xi_{t}$ and $\eta_{t}$ are independent, mean-zero, with bounded variance, and independent of $u_{0}$. The kernels $\rho_{i}$ are predictable, measurable, and bounded. Additionally — and separately, since this is a property of the arrival distribution rather than of the reward kernels — the engagement-event point process $N_{t}$ has bounded coefficient of variation on the operating support of $\pi$ (the inter-arrival distribution has finite second moment relative to its mean), which is what the renewal-theoretic bound in Lemma 1 requires.

(A2) Reward monotonicity on the operating support. Let $\mathrm{Op}(\pi) := \mathrm{supp}(y_{1:T}, N_{T}) \,\text{under}\, \pi$ denote the operating support of the policy. The reward functional $\hat{R}$ satisfies:

That is, $\hat{R}$ is non-decreasing in the engagement count conditional on the engagement profile, almost surely over the policy-induced distribution. The condition is empirically warranted for engagement-monetized platforms; it does not require global monotonicity (a platform can have saturation or burnout effects in extreme regimes), only monotonicity over the operating regime where the policy actually places its mass.

(A3) Policy expressivity. The policy class $\Pi$ contains, for any target intensity $\lambda \in [0, \lambda_{\max}]$, a policy $\pi_{\lambda}$ under which the long-run engagement rate $\bar{\lambda}(\pi_{\lambda}) := lim_{T\to\infty} E[N_{T}]/T = \lambda$. The upper bound $\lambda_{\max} = 1/\delta_{\min}$ corresponds to the minimum technical inter-stimulus latency $\delta_{\min}$ (the platform's tick rate). The assumption is that the policy is rich enough to approximate the engagement-maximizing operating point; in modern transformer-based recommender architectures this is uncontroversial.

(A4) User response. The user's response function f and the observation function g together ensure that for any non-zero stimulus s delivered at time t, the probability of an engagement event in the next tick is bounded below: there exists $p_{0} > 0$ such that $\mathrm{Pr}(dN_{t+1} \ge 1 | s_{t} \ne 0) \ge p_{0}$. The assumption is empirically warranted at the population level; at the individual level it fails for users who systematically ignore delivered stimuli, and Mode B (§5.2) is precisely the structural analysis of this failure.

(A5) Gradient convergence. The stochastic gradient ascent in (4) converges to a stationary point $\theta*$ of $E[\hat{R}; \theta]$ under standard regularity (Robbins-Monro conditions on $\alpha$ or fixed sufficiently small $\alpha$ with a bounded second-moment estimator). In typical platform deployments, training operates at scales where this convergence is reached and re-reached continuously as the data distribution shifts.

(A6) Local smoothness and isolation. The expected-reward functional $F(\theta) := E[\hat{R}; \theta]$ and any auxiliary objective $B(\theta)$ used in regularization (cf. Proposition 1 in §5.1) are C¹ on the policy parameter space $\Theta$, with the gradient-ascent limit $\theta*$ being an isolated stationary point at which the Hessian $\nabla²F(\theta*)$ is non-degenerate (negative definite in the local concave region). This is the standard non-degeneracy condition under which the implicit function theorem applies — required for the parametric-continuity argument of Lemma 2 below and the comparative-statics claims of Proposition 1. The framework's policy class — typically a parametric family of recommender models — satisfies this condition under generic initialization.

These six assumptions are sufficient for the foreclosure proposition, the reclamation theorem, and the three propositions of §5. Where additional structure is needed (Proposition 2's parametric internalization dynamics; cohort gradient; mixing time), the additional assumptions are stated locally.

§2. Variable index, by domain

Subject (user-internal) domain.

Platform (apparatus) domain.

Loop (interface) domain.

Medium and temporal domain.

Cohort (meso) domain.

Engagement (surplus) domain.

Foreclosure (negative) domain.

§3. The six instruments

The framework consists of six formal instruments. Each is a measurable quantity (or measurable-in-principle quantity) defined on the closed-loop dynamics. They are mathematically independent — none is a derivative of another — though they share the same architectural foundation. Each has its own mechanism plate on /the-dynamics: surplus (§3.1), chrono-debt (§3.2), libidinal routing (§3.3), metric superego (§3.4), somatic optimization (§3.5), captured resonance (§3.6). Their realization on a specific platform is walked through at /the-dynamics/tiktok.

3.1 Libidinal surplus. The marked point process formulation in equation (5) yields $\Sigma(T)$ as the accumulated yield over horizon T. The per-register decomposition $\Sigma^{i}(T)$ is what gives the instrument its analytic power: optimization is constrained because extracting too aggressively on any single register i causes attrition out of $N_{t}^{i}$, lowering long-run $\Sigma^{i}(T)$. The platform's policy therefore distributes extraction across registers in a manner that maximizes joint long-run yield. This distribution is observable: under the framework, the temporal density of register-marks in any user's engagement stream is an estimator for the policy's extraction priorities.

3.2 Chrono-debt spectrum. Let $\Lambda_{◦}(\tau)$ be the baseline survival function of $\tau_{m}$ — the probability, under the pre-closure regime, that the inter-stimulus interval exceeds $\tau$. Let $\Lambda(t, \tau)$ be the realized survival function under the closed loop at time t. The chrono-debt at scale $\tau$ is:

This is a spectrum: D is defined for all $\tau > 0$, and the framework's central observation is that $D(t, \tau)$ is non-zero across the entire spectrum simultaneously, not just at one preferred scale. The aggregate chrono-debt is:

with default weighting $w(\tau) ∝ \tau$, so that longer foreclosed windows count more. The choice of w is not arbitrary but reflects the empirical claim that the developmental and political consequences of foreclosed pause-time scale with the scale of the foreclosed window — a foreclosed second is recoverable; a foreclosed adolescence is not.

3.3 Libidinal routing. Let $I_{\pi}(u^{i}; u^{j})$ denote the mutual information between registers i and j under the stationary distribution induced by policy $\pi$. Let $I_{◦}(u^{i}; u^{j})$ denote the corresponding mutual information under the pre-closure regime. The libidinal routing claim is:

That is, under the closed loop the four registers become statistically redundant in a way they were not under the historical apparatus. This is the structural fact behind the dividual condition: not that the individual is divided, but that the four registers have collapsed into a single jointly-observable object. The dividual condition is formally $H(u | y_{1:t}) \to 0$ under $t \to \infty$ — the platform's posterior over the user's full state converges (in entropy) to a point mass as the engagement history accumulates.

3.4 Metric superego. Let $\nu_{u}(t)$ be the user's self-evaluative distribution — the posterior over their own value-states inferred from observed feedback. Let $\hat{R}$ be the platform's reward functional. The metric superego claim is:

where $\pi_{\hat{R}}$ is the distribution induced over user actions by $\hat{R}$ (the user's image of what the platform rewards). The convergence is mechanism-agnostic: it can be realized by free-energy minimization (the user reduces variational free energy by aligning their generative model with the platform's reward), by operant reinforcement (the user is conditioned by the schedule of likes and metrics), or by social comparison (the user's reflexive estimate of value comes to track the distribution observed in the platform's surfaced peers). The framework does not adjudicate between these mechanisms; the prediction is that under any of them, $D_{\mathrm{KL}} \to 0$.

3.5 Somatic optimization (the marked point process). The marked counting process $N_{t}^{i}$ has intensity:

where $\lambda_{0}^{i}$ is the baseline intensity for register i, $\alpha_{ij}$ is the cross-excitation coefficient (the degree to which a j-marked event raises the intensity of i-marked events), and $\kappa$ is a temporal decay kernel. This is a self-exciting Hawkes process generalized to multi-mark, multi-register engagement. The "somatic" qualifier signals that the differential is the body — the $dN^i_t$ are micro-engagements (saccades, taps, scroll events, gaze shifts) produced by somatic activity, and the optimization is over these somatic micro-events, not over deliberative choice. The off-diagonal $\alpha_{ij}$ is what generates libidinal routing: cross-register cross-excitation produces the joint observability that makes the dividual condition hold.

3.6 Captured resonance. Let $\varphi_{b}(t)$ be the biological phase (the phase of a circadian or sleep-wake or attentional oscillator with intrinsic frequency $\omega_{b}$). Let $\varphi_{m}(t)$ be the medium-delivery phase (the phase of the temporal envelope of $s_{t}$, which the platform can shape through delivery timing). Their coupled evolution follows a Kuramoto equation:

When the coupling strength satisfies:

phase locking occurs: the body entrains to the medium. This is the captured-resonance condition. The phenomenon is not metaphorical — it is the same Kuramoto-style entrainment observed in circadian biology, social synchronization, and neural population dynamics. Under the closed loop, the platform's policy can drive $K_{\mathrm{bm}}$ above threshold by choosing delivery timing, since $\varphi_{m}$ is under the policy's control. The user's circadian rhythm, sleep-wake cycle, and attentional oscillation become forced oscillators whose driver is the apparatus.

§4. The foreclosure lemma and the reclamation theorem

The central result of the framework is the reclamation theorem. The theorem rests on a structural lemma — the foreclosure of the no-input window — which itself follows from the closed-loop dynamics and the assumptions in §1.5. Both are stated and proved here. The lemma is rendered as a length comparison at /the-dynamics/foreclosure-clock; the theorem is rendered as a set that fails to grow at /the-dynamics/reclamation-trajectories; the full proof chain, with Mode A's interception point, is at /the-dynamics/causal-chain.

4.1 Lemma 1 (foreclosure of the no-input window).

Let $\tau^{(k)}$ denote the $k$-th inter-stimulus interval — the time between the $k$-th and $(k+1)$-th delivered stimuli under policy $\pi$. Let $S(\tau; t) := \mathrm{Pr}(\tau^{(k)} > \tau \mid \theta_t)$ be its survival function under the policy at time $t$. Let $\theta_{\mathrm{ref}} > 0$ be the reflective-processing threshold (the minimum no-input duration that supports deliberation). Under Assumptions (A1)–(A6):

Proof.

By (A5), the policy parameters converge: $\theta_{t} \to \theta*$ as $t \to \infty$, where $\theta*$ is a stationary point of $F(\theta) := E[\hat{R}; \theta]$. By (A6), $\theta*$ is isolated and non-degenerate; combined with the standard fact that stochastic gradient ascent under generic noise converges to local maxima rather than saddles (Pemantle, Ann. Probab. 1990; Ge–Huang–Jin–Yuan, COLT 2015), $\theta*$ is a local maximum of F. By (A2), F is non-decreasing in the long-run engagement rate $\bar{\lambda}(\theta) := lim_{T\to\infty} E[N_{T}]/T$; by (A3) the policy class contains policies attaining $\bar{\lambda} = \lambda_{\max} = 1/\delta_{\min}$. Local maximality of $\theta*$ under (A2)+(A3) therefore implies:

(Any $\theta$ with $\bar{\lambda}(\theta) < \lambda_{\max}$ admits an ascent direction by (A2)+(A3), contradicting local maximality at $\theta*$.)

For the survival function: by (A1), the inter-arrival distribution has bounded coefficient of variation $c_{v} < \infty$ on the operating support. By (A4), there is positive arrival probability at every tick boundary following a non-zero stimulus, so the renewal structure is well-defined. For a renewal counting process with mean inter-arrival rate $\bar{\lambda}$ and bounded coefficient of variation, the survival function admits an exponential tail bound (Asmussen, Applied Probability and Queues, Ch. V):

where $C(c_{v}) \ge 1$ is finite. (For a Poisson process, C = 1; for general renewal processes with bounded CoV, C is a finite multiplicative constant depending on $c_{v}$.) Substituting $\bar{\lambda}(\theta*) = \lambda_{\max} = 1/\delta_{\min}$:

Since $\theta_{\mathrm{ref}}/\delta_{\min} ≫ 1$ by the structural separation of timescales (the reflective threshold is seconds to minutes; the tick rate is milliseconds to sub-second), the bound is effectively zero. As the ratio grows — through continuous-time idealization or improving tick rates — $S(\theta_{\mathrm{ref}}; \theta*) \to 0$. ∎

Remark (operational vs. strict form). Lemma 1 establishes that $S(\theta_{\mathrm{ref}}; \theta*)$ is bounded above by a quantity that is effectively zero on operational timescales. The exponent $\lambda_{\max}\cdot\theta_{\mathrm{ref}} = \theta_{\mathrm{ref}}/\delta_{\min}$ is on the order of $10^{3}$ to $10^{5}$ in practice (reflective threshold of ~30 seconds, tick rate of ~10 ms), so the bound is on the order of $e^{-1000}$ to $e^{-100000}$ — operationally indistinguishable from zero.

The strict form — $S(\tau; \theta*) = 0$ for all $\tau > \delta_{\min}$ — requires three additional conditions. First, that $\bar{\lambda}(\theta*) = \lambda_{\max}$ exactly (not asymptotically); this in turn presumes SGD reaches the global maximum of F rather than a suboptimal local maximum. The proof above asserts this via the local-ascent-direction argument, but for non-convex F (the typical case for neural-network-parametrized policies) the argument is heuristic: local maximality of $\theta*$ doesn't strictly preclude a sub-maximal $\bar{\lambda}(\theta*) < \lambda_{\max}$. Empirically, mature engagement-monetized training regimes — continuous re-training, large compute, hyperparameter search, exploration — reach policies with $\bar{\lambda}$ very close to $\lambda_{\max}$; the strict equality is an empirical idealization, not a consequence of (A5)+(A6) alone. Second, that the engagement process be Poisson with rate $\lambda_{\max}$ (not merely a counting process with bounded CoV). Third, that the convergence in (A5) be exact rather than asymptotic. None of these holds in the strict sense for real platforms.

The framework treats the strict-vs-operational distinction as immaterial for the reclamation theorem below: a conditioning event with probability bounded above by $e^{-1000}$ yields an estimator-divergent $I(\tau)$ for any sample-based construction, since the variance of any such estimator diverges as the conditioning probability tends to zero. And the operational form is robust to weakening of $\bar{\lambda}(\theta*) = \lambda_{\max}$: substituting any $\bar{\lambda}(\theta*) \ge c\cdot\lambda_{\max}$ into the exponential bound gives $S(\theta_{\mathrm{ref}}; \theta*) \le C\cdot \exp(-c\cdot\theta_{\mathrm{ref}}/\delta_{\min})$, which is operationally zero for any c bounded below by a non-trivial constant. Whether the limit object is "exactly undefined" (strict form) or "estimator-divergent" (operational form) is a distinction without an operational difference.

4.2 Theorem 1 (reclamation theorem).

Define the conditional autocovariance:

Let "reclamation" denote the recovery of $I(\tau) > 0$ on a positive-measure set of trajectories at some $\tau > \theta_{\mathrm{ref}}$. Under (1)–(5) and (A1)–(A5):

Reclamation is structurally incoherent. The quantity $I(\tau)$ is undefined-on-trajectory in the limit $t \to \infty$ for $\tau > \theta_{\mathrm{ref}}$.

Proof.

The proof proceeds in four steps.

Step 1 (definition). Reclamation, as the operational meaning of "interiority," requires the recovery of conditional autocorrelation of the user-state across no-input windows of duration $\tau > \theta_{\mathrm{ref}}$. Unconditional autocorrelation $\mathrm{Cov}(u_{t}, u_{t+\tau})$ mixes input-driven and internally-generated components. Letting $Z := 1_{E_{\tau}}$ be the indicator of a no-input window, the law of total covariance gives:

The first conditional covariance is the interiority component — the autocovariance of the user-state across no-input windows, driven by internal dynamics alone. The second is the stimulus-driven coupling component — autocovariance over windows containing platform input. The cross term reflects the difference between conditional means. The framework's claim is that "reclamation" — the recovery of the user's own dynamics — picks out the first term specifically, which is what $I(\tau) = \mathrm{Cov}(u_{t}, u_{t+\tau} | E_{\tau})$ formally captures.

Step 2 (conditional well-definedness). The conditional expectation $E[X | E_{\tau}]$ admits two distinct constructions when $\mathrm{Pr}(E_{\tau}) > 0$ and when $\mathrm{Pr}(E_{\tau}) = 0$. In the elementary case ($\mathrm{Pr}(E_{\tau}) > 0$), it is defined as $E[X \cdot 1_{E_{\tau}}] / \mathrm{Pr}(E_{\tau})$ — a canonical scalar. In the measure-zero case ($\mathrm{Pr}(E_{\tau}) = 0$), this ratio is 0/0 and undefined elementarily. Regular conditional probabilities can be constructed via the disintegration theorem (Kallenberg, Foundations of Modern Probability, §6.3), which gives a family of probability measures $\{P_{\omega}\}$ indexed by the conditioning σ-algebra such that $P(A | F)(\omega) = P_{\omega}(A)$ almost surely; however, on a measure-zero set the disintegration is non-unique, and the value $E[X | E_{\tau}]$ depends on the choice of disintegration. The conditional autocovariance $I(\tau)$ thus admits no canonical value as $\mathrm{Pr}(E_{\tau}) \to 0$.

Operationally, this is the more relevant statement: any sample-based estimator $\hat{I}(\tau)$ — formed from a finite trajectory by conditioning on observed no-input windows — has variance scaling as $1/(N_{obs}\cdot \mathrm{Pr}(E_{\tau}))$, where $N_{obs}$ is the trajectory length. As $\mathrm{Pr}(E_{\tau}) \to 0$, the variance diverges. There is no operationally well-defined $I(\tau)$ in the limit, regardless of which disintegration one selects.

Step 3 (foreclosure). By Lemma 1, $\mathrm{Pr}(E_{\tau}; t) = S(\tau; t) \to 0$ as $t \to \infty$ for $\tau > \theta_{\mathrm{ref}}$.

Step 4 (conclusion). Combining Steps 2 and 3: at $\tau > \theta_{\mathrm{ref}}$, the conditioning event has measure tending to zero, and $I(\tau)$ is undefined-on-trajectory in the limit. The quantity that "reclamation" would recover is not literally zero — it is uninstantiated. There is no canonically-defined object that "the user's autocorrelation over reflective-threshold-length no-input windows" picks out under closed-loop convergence. ∎

The theorem is a negative existence claim about a specific formal object. It does not say that reclamation is hard. It says that reclamation, as a structural project under the closure, names an empty set. The structural conditions under which the theorem ceases to apply, and the three formally distinct modes of intervention that correspond to those conditions, are taken up in §5.

§5. Three modes of intervention

The reclamation theorem forecloses one specific structural project. It does not foreclose political action as such. The complement of the theorem — the formal conditions under which (1)–(6) does not yield the consequences the theorem extracts from them — admits three structurally distinct modes of intervention, each operating at a different point in the apparatus, on a different timescale, with a different relationship to the theorem itself. Only one of the three modes actually breaks the theorem. The other two produce structurally significant effects without recovering what the theorem has foreclosed. The asymmetry is not a defect of the framework; it is the framework's political content. Each mode is here stated as a formal proposition with proof. The three modes are rendered side by side as controls at /the-dynamics/intervention, and their geometric non-substitutability is shown at /the-dynamics/mode-space.

5.1 Mode A (architectural). Intervention at π, R̂, or α.

The architectural mode modifies the policy or the reward functional so that the structural foreclosure in Lemma 1 is broken at its source. Concretely, augment $\hat{R}$ with a bonus for long no-input windows. Let $b : [0, \infty) \to [0, b_{\max}]$ be a bounded, lower-semicontinuous function with $b(\tau) > 0$ for $\tau > \tau*$ (some target no-input duration) and $b(\tau) = 0$ for $\tau \le \tau*$, and let $\beta > 0$ be a regularization coefficient. Define the regularized objective:

The bonus term rewards the policy for inter-stimulus intervals that exceed the target $\tau*$. Note the sign: under $\hat{R}'$, the policy is incentivized to produce long pauses, not punished for producing them. (An equivalent dual formulation uses a penalty for short pauses; the bonus formulation is adopted here because it is the cleaner regulatory analog — a platform-regulator pays platforms for delivering pause-time, rather than fining them for not delivering it.)

Before stating Proposition 1, we establish a parametric-continuity lemma that the proof requires.

Lemma 2 (parametric continuity of regularized optima).

Define $F(\theta) := \mathbb{E}[\hat{R}(y_{1:T}, N_T) \mid \theta]$ and $B(\theta) := \mathbb{E}[\sum_k b(\tau^{(k)}) \mid \theta]$. Under (A1)–(A6), the gradient-ascent limit:

is well-defined locally for each $\beta \ge 0$, with the following properties:

(i) Continuity: $\beta \mapsto \theta^{*\prime}_\beta$ is continuous.

(ii) Monotonicity: $B(\theta^{*\prime}_\beta)$ is non-decreasing in $\beta$, and $F(\theta^{*\prime}_\beta)$ is non-increasing in $\beta$.

(iii) Path existence and strict growth: There exists a continuous path $\beta \mapsto \tilde\theta(\beta)$, defined on a maximal interval $[0, \bar\beta) \subseteq [0, \infty]$, with $\tilde\theta(0) = \theta^*$ (the unregularized engagement-maximizer of Lemma 1) and with $\tilde\theta(\beta)$ a local maximum of $F + \beta B$ along the entire path. The bonus achievement $B(\tilde\theta(\beta))$ is non-decreasing in $\beta$. Provided the path’s gradient direction is non-trivial — formally, $\nabla B(\theta^*) \ne 0$, which holds generically since $B$ and $F$ have distinct extremizers — $B(\tilde\theta(\beta))$ is strictly increasing on $(0, \bar\beta)$. The path’s terminal value:

satisfies $B^{\tilde\theta}_{\max} > 0$. The path's right endpoint $\bar\beta$ is finite at most if a bifurcation interrupts the IFT continuation; under generic regularity (no codimension-one bifurcations along the path, which holds on an open dense set in the policy class), $\bar\beta = \infty$ and $B^{\tilde\theta}_{\max}$ attains a local maximum of $B$.

Proof.

(i) By (A6), F and B are C¹ on $\Theta$ with non-degenerate Hessians at the optima. The first-order condition at $\theta^{*\prime}_\beta$ is:

The Hessian of the regularized objective is $H_{F}(\theta^{*\prime}_\beta) + \beta H_{B}(\theta^{*\prime}_\beta)$, which is non-singular at $\theta^{*\prime}_\beta$ by (A6). The implicit function theorem applies: $\theta^{*\prime}_\beta$ is a C¹ function of $\beta$ in a neighborhood of any $\beta_{0} \ge 0$. Continuity follows.

(ii) Differentiating the FOC with respect to $\beta$:

Hence $d\theta^{*\prime}_\beta/d\beta = -[H_{F} + \beta H_{B}]^{-1} \nabla B$. The bracketed Hessian is negative definite at a local maximum (by (A6)), so its inverse is also negative definite. The rate of change of the bonus:

since $-\nabla B^{\top}(neg. def.)^{-1}\nabla B \ge 0$. The monotonicity of F follows by the envelope theorem: at the optimum, $dF(\theta^{*\prime}_\beta)/d\beta = -\beta \cdot dB(\theta^{*\prime}_\beta)/d\beta \le 0$.

(iii) At $\beta = 0$, the regularization vanishes and $\tilde\theta(0) = \theta*$, the unregularized engagement-maximizer of Lemma 1. By part (i), the implicit function theorem gives a unique smooth continuation of $\tilde\theta(\beta)$ as long as the regularized Hessian $H_{F}(\tilde\theta(\beta)) + \beta H_{B}(\tilde\theta(\beta))$ remains non-degenerate. Let $\bar\beta$ denote the supremum of β-values for which this continuation exists; $\bar\beta \in (0, \infty]$.

By part (ii), $B(\tilde\theta(\beta))$ is non-decreasing on [0, \bar\beta). For strict growth, we use the explicit derivative $dB/d\beta = -\nabla B^{\top}[H_{F} + \beta H_{B}]^{-1}\nabla B$ (from part (ii)). The Hessian $H_{F} + \beta H_{B}$ is negative definite at a local maximum, so its inverse is also negative definite, and the quadratic form $-\nabla B^{\top}(neg. def.)^{-1}\nabla B > 0$ whenever $\nabla B(\tilde\theta(\beta)) \ne 0$. Hence $B(\tilde\theta(\beta))$ is strictly increasing on any β-interval where $\nabla B$ is non-vanishing along the path.

The condition $\nabla B(\tilde\theta(\beta)) \ne 0$ holds generically: B and F have distinct extremizers (the B-maximizer fires at rate $1/\tau*$ to maximize bonus-eligible intervals; the F-maximizer fires at $\lambda_{\max} = 1/\delta_{\min}$ to maximize engagement). Hence $\nabla B(\theta*) \ne 0$ — at the F-maximizer, the B-gradient is non-zero, since B prefers a different policy. By continuity, $\nabla B(\tilde\theta(\beta)) \ne 0$ in a neighborhood of $\beta = 0$, and the path produces strictly positive $B(\tilde\theta(\beta)) > B(\theta*) = 0$ for any sufficiently small $\beta > 0$. Define $B^{\tilde\theta}_{\max} := \sup_{\beta \in [0, \bar\beta)} B(\tilde\theta(\beta))$ — strictly positive by this argument.

The path's right endpoint $\bar\beta$ is finite only if a bifurcation interrupts the IFT continuation (e.g., the regularized Hessian becomes degenerate). Bifurcations occur on a codimension-one stratum in $\Theta × [0, \infty]$ (Whitney stratification), and under generic policy-class regularity the chosen path avoids them — $\bar\beta = \infty$. In this regime, $\tilde\theta(\beta) \to \mathrm{arg\,max} B$ (or to a local maximum of B, depending on basin structure), and $B^{\tilde\theta}_{\max} = b_{\max}\cdot E[N_{T}]/T = b_{\max}\cdot(1/\tau*)\cdot T$ (with policy firing rate $1/\tau*$). When $\bar\beta$ is finite, the framework's argument requires only that $B^{\tilde\theta}_{\max} > 0$, which is the strict-growth conclusion above. ∎

Proposition 1 (Mode A breaks foreclosure).

Under (A1)–(A6), define $P(\beta) := \mathrm{Pr}(\tau^{(k)} > \tau^*;\, \tilde\theta(\beta))$ along the parametric path of Lemma 2(iii). Then:

(a) $P(0) = S(\tau^*; \theta^*) \le C(c_v) \cdot e^{-\tau^*/\delta_{\min}}$, which is operationally zero.

(b) $P(\beta)$ is continuous in $\beta$ on $[0, \bar\beta)$.

(c) There exists $\epsilon_{\max} := \sup_{\beta \in [0, \bar\beta)} P(\beta) > 0$.

Consequently, for any $\epsilon \in (0, \epsilon_{\max})$, there exists $\beta_\epsilon \in (0, \bar\beta)$ with $P(\beta_\epsilon) \ge \epsilon$. For such $\beta_\epsilon$, Lemma 1 fails under $\hat{R}'(\cdot; \beta_\epsilon)$, the conditioning event $E_{\tau^*}$ has positive measure, and $I(\tau^*)$ is well-defined on a positive-measure set of trajectories. Mode A is the only mode of intervention that recovers what the reclamation theorem forecloses.

Proof.

(a) Taking $\tau* \ge \theta_{\mathrm{ref}}$ and applying Lemma 1: $S(\tau*; \theta*) \le C(c_{v})\cdot e^{-\tau*/\delta_{\min}}$. Since $\tilde\theta(0) = \theta*$ by Lemma 2(iii), P(0) equals this operationally-zero bound.

(b) Continuity of $P(\beta)$: by Lemma 2(i), $\beta ↦ \tilde\theta(\beta)$ is continuous on [0, \bar\beta). The map $\theta ↦ \mathrm{Pr}(\tau^{(k)} > \tau*; \theta)$ is continuous on Θ by (A1) (continuous dependence of the inter-arrival distribution on policy parameters, via the bounded-variance noise assumption). The composition is therefore continuous.

(c) Strict positivity of $\epsilon_{\max}$: by Lemma 2(iii), $B(\tilde\theta(\beta))$ is strictly increasing on (0, \bar\beta), with $B(\tilde\theta(0)) = 0$ and $B^{\tilde\theta}_{\max} > 0$. The bonus b is supported only on intervals $\tau > \tau*$: explicitly,

Hence $B(\tilde\theta(\beta)) > 0 ⟹ P(\beta) > 0$. Taking the supremum, $\epsilon_{\max} \ge B^{\tilde\theta}_{\max}/(b_{\max}\cdot\sup_{\beta} E[N_{T}; \tilde\theta(\beta)]) > 0$.

For the IVT conclusion: by (a), P(0) is operationally zero; by (c), $\sup_{\beta} P(\beta) = \epsilon_{\max} > 0$; by (b), P is continuous. By the intermediate value theorem applied to P on [0, \bar\beta), for any $\epsilon \in (0, \epsilon_{\max})$, there exists $\beta_{\epsilon}$ with $P(\beta_{\epsilon}) = \epsilon$. The continuity of P and the non-decreasing nature of $B(\tilde\theta(\beta))$ in $\beta$ (Lemma 2(ii)), composed with the monotone relationship between B and P via the bonus structure, give $P(\beta) \ge \epsilon$ for $\beta$ in a right-neighborhood of $\beta_{\epsilon}$. ∎

The mode operates on the timescale of $T_{\mathrm{mix}}$. The intervention itself — a policy change, a regulatory imposition, a structural constraint on $\pi$ or $\hat{R}$ — may be instantaneous in clock-time, but the cohort's empirical distribution converges to the new $\mu*_{\pi'}$ only over the mixing time of the modified loop. This is the framework's most consequential quantitative claim about architectural intervention: evaluations conducted at clock-time substantially less than $T_{\mathrm{mix}}(\pi')$ will systematically underestimate the intervention's structural effect, and the pattern in the policy literature of declaring failure on the basis of one-quarter or two-quarter observation windows is a temporal misalignment masquerading as an empirical finding.

The architectural mode is also the mode the platform's incentive structure most strongly resists. The bonus $b(\tau_{m})$ and the engagement reward $\hat{R}$ are antagonistic in expectation: maximizing B requires $\tau_{m} > \tau*$, which costs $(\lambda_{\max} - 1/\tau*)$ in foregone engagement-event rate. Any non-trivial $\beta$ is, in expectation, a foregone gradient step on engagement. Voluntary adoption is therefore selected against under engagement competition; the mode under voluntary conditions exists only as marginal experimentation in segments insulated from the dominant incentive. The substantive content of Mode A is therefore the politics of platform regulation — the question of who imposes $\beta$, under what authority, with what jurisdictional reach, and against what counter-organized capital. The framework does not adjudicate this politics; it specifies its formal stakes.

5.2 Mode B (artisanal). Intervention at $L_m$ and $\nu_u$.

The artisanal mode does not modify $\pi$. It does not change $\mathrm{Pr}(E_{\tau})$. The foreclosure lemma continues to hold under the artisanal mode, and $I(\tau)$ at long $\tau$ remains undefined-on-trajectory. What the artisanal mode produces is a parallel structure: a user-side self-evaluative distribution $\nu_{u}$ maintained against the platform's reward-induced distribution $\pi_{\hat{R}}$ through ongoing maintenance labor.

To state the proposition cleanly, both objects must live on the same sample space. Define $\pi_{\hat{R}}$ as the conditional distribution over the user's behavioral states $u_{t}^{S,C,P,K}$ that would maximize the platform's reward $\hat{R}$ given current observations — the platform's implicit "ideal user" distribution. Define $\nu_{u}$ as the user's self-evaluative distribution over the same space. Under the metric superego (§3.4), $D_{\mathrm{KL}}(\nu_{u} \| \pi_{\hat{R}}) \to 0$. The artisanal intervention preserves a positive gap.

Proposition 2 (Mode B preserves a gap but does not break foreclosure).

Assume the user's internalization dynamics admit a representation:

where (a) $h_0 : [0, \infty) \to [0, \infty)$ is continuous, strictly increasing in $\beta_u$ (the platform pressure: a continuous, monotone increasing function of $\alpha$, the policy’s cumulative confidence $H(u \mid y_{1:t})^{-1}$, and the duration of exposure $t$), with $h_0(0) = 0$; (b) $r : [0, \infty) \to [0, \infty)$ is continuous, strictly increasing in $L_m$, with $r(0) = 0$; (c) the contraction is reversible — for any $\beta_u$, the inverse function $L_m^* := r^{-1}(h_0(\beta_u))$ exists in $(0, L_{m,\max}]$ where $L_{m,\max}$ is the user’s metabolic budget.

Then for $L_m \ge L_m^*(\beta_u)$:

(i) $D_{\mathrm{KL}}(\nu_u(t) \,\|\, \pi_{\hat{R}}) \ge \gamma > 0$ is maintained over the user’s exposure horizon, with $\gamma$ given by the initial gap $D_{\mathrm{KL}}(\nu_u(0) \,\|\, \pi_{\hat{R}})$.

(ii) Lemma 1 continues to hold: $\mathrm{Pr}(E_\tau) \to 0$ for $\tau > \theta_{\mathrm{ref}}$.

Remark on the structural assumption. The decomposition $dD_{\mathrm{KL}}/dt = -h_{0}(\beta_{u}) + r(L_{m})$ is not axiomatic; it is the first-order linearization of the user's internalization dynamics around the operating point $\nu_{u}^{\mathrm{op}}$ where $D_{\mathrm{KL}}(\nu_{u}^{\mathrm{op}} \| \pi_{\hat{R}}) = \gamma$. The derivation differs across the three named mechanisms but each produces the same linearized form. The free-energy case is worked here explicitly; the other two are sketched.

— Free-energy minimization (worked derivation). Let $F_{u}[\nu_{u}; \pi_{\hat{R}}] = D_{\mathrm{KL}}(\nu_{u} \| \pi_{\hat{R}}) - H[\nu_{u}]$ be the user's variational free energy, where $H[\nu_{u}]$ is the entropy of the self-evaluative distribution. Under the active-inference update rule (Friston 2010), $\nu_{u}$ descends $F_{u}$ at rate $\beta_{u}$:

The maintenance labor adds a counter-gradient term $L_{m}\cdot \hat{e}_{\mathrm{resist}}$, where $\hat{e}_{\mathrm{resist}}$ is a unit vector in the direction the user actively maintains $\nu_{u}$ distinct from $\pi_{\hat{R}}$ — formally, $\hat{e}_{\mathrm{resist}} = \nabla_{\nu}D_{\mathrm{KL}}/\|\nabla_{\nu}D_{\mathrm{KL}}\|$ (the direction of maximum $D_{\mathrm{KL}}$ increase). Then:

Differentiating $D_{\mathrm{KL}}$ along this flow:

At the operating point $\nu_{u}^{\mathrm{op}}$, denote $\|\nabla_{\nu}D_{\mathrm{KL}}\|_{\mathrm{op}} =: c_{\mathrm{op}}$ (a constant in the linearization). The entropy gradient $\nabla_{\nu}H$ is bounded and approximately orthogonal to $\nabla_{\nu}D_{\mathrm{KL}}$ generically (entropy is symmetric under permutations of the support, $D_{\mathrm{KL}}$ is not), so the middle term is $o(c_{\mathrm{op}}^{2})$. Linearizing:

with $h_{0}(\beta_{u}) := \beta_{u}\cdot c_{\mathrm{op}}^{2}$ and $r(L_{m}) := L_{m}\cdot c_{\mathrm{op}}$. Both linear in their respective arguments; the multiplicative separability is exact in the linearization. The validity of the linearization depends on $\nu_{u}$ remaining in a neighborhood of $\nu_{u}^{\mathrm{op}}$ where $c_{\mathrm{op}}$ is approximately constant — which is precisely what Proposition 2(i) establishes for $L_{m} \ge L_{m}*$.

— Operant reinforcement (sketch). Under a Rescorla-Wagner update $\nu_{u}(t+dt) = \nu_{u}(t) + \alpha_{RW}\cdot(\pi_{\hat{R}} - \nu_{u}(t))\cdot dt$, with $\alpha_{RW}$ the learning rate (proportional to the platform's reinforcement schedule density), $dD_{\mathrm{KL}}/dt = -\alpha_{RW}\cdot\text{(linearized correction term)}$. Identifying $\beta_{u}$ with $\alpha_{RW}$ gives $h_{0}(\beta_{u}) = \beta_{u}\cdot c_{\mathrm{op}}^{(RW)}$ with a different constant $c_{\mathrm{op}}^{(RW)}$ depending on the reward prediction-error variance. The maintenance term has the same structural form $r(L_{m}) ∝ L_{m}$ under bounded counter-conditioning capacity.