§5 Intervention as causal hierarchy

Three modes · the 2×2 taxonomy · the Mode A scaffold

The architecture supplies three distinct intervention sites: the reward functional $\hat{R}$ (Mode A); the user response kernel $T$ (Mode B); the observation channel $O$ (Mode C). The three modes are asymmetric, and the asymmetry has a structural cause: Mode A intervenes on the topmost layer of the performative optimization and modifies the fixed point of $Φ$ directly. Modes B and C intervene on inputs to $Φ$ and propagate to the fixed point only through the platform's re-optimization, which the platform's population-level averaging dampens. The unified statement is Theorem 9; the per-mode propositions below are its corollaries.

§5.0 The 2×2 taxonomy

Two orthogonal axes: the intervention site (which structural variable is modified) and the imposition source (individual practice or regulatory imposition). The taxonomy makes the asymmetry of the three modes legible — and surfaces the fourth-cell mode (regulatory imposition on the observation channel) that emerges naturally from the cross-product.

	Site: reward $\hat{R}$	Site: response $T$	Site: observation $O$
Individual practice	structurally vacuous	Mode B	Mode C
Regulatory imposition	Mode A	population-coordinated response	fourth-cell

The 2×2 separates two questions older critical-theory frameworks tended to conflate: which structural variable is being modified, and who is doing the modifying. The reward functional $\hat{R}$ , the response kernel $T$ , and the observation channel $O$ are the framework's three intervention sites — derived from the apparatus's (P1)–(P5) setup. The imposition source is either individual practice (the reflective subject working on themselves) or regulatory imposition (the institutional apparatus acting on the platform). The cross-product gives six cells, two of which are structurally degenerate — individual practice on the reward functional (users cannot rewrite $\hat{R}$ ) and regulatory imposition on the response kernel (no policy acts on $T$ directly for any non-coordinated population). The remaining four cells are the framework's intervention modes.

The top-left cell (individual practice on the reward) is structurally vacuous — users cannot rewrite the platform's reward functional. The remaining cells correspond to the three named modes plus the fourth-cell extension.

The fourth-cell mode. Regulatory imposition on the observation channel — structurally Mode A applied to the Mode C site. Top-of-hierarchy intervention on the observation-channel variable. The cell is empirically reachable: data-minimization mandates, default-on inferential noise, the right to inferential opacity. The policy-translation of the formal observation that bounding the platform's information about the user bounds the platform's extraction from the user. Zuboff's analysis of surveillance capitalism (The Age of Surveillance Capitalism, 2019) argued precisely this — that the platform's extractive capacity is grounded in its access to the behavioral surplus the observation channel produces. The fourth-cell mode is the apparatus reading of Zuboff's policy implication: regulate $O$ as well as $\hat{R}$ .

§5.1 Proposition 1' — Mode A (reward intervention)

Mode A modifies the reward functional with a regularizer that rewards the inter-stimulus interval (the pause) exceeding a threshold. The framework's central political claim: at sufficient regularization $β$ , the survival probability at the reflective threshold becomes positive. Mode A breaks foreclosure where Modes B and C cannot.

Mode A intervenes directly on the topmost layer of the optimization. The regularized functional $F^{'} (θ; β) = F (θ) + β B (θ)$ (with $B$ the survival-bonus integral; see Lemma 2) supplies a continuous parametric path from the unregularized foreclosure regime $β = 0$ to a regime in which the pause-time at the reflective threshold becomes positive. By the intermediate value theorem on the continuous map $P (β)$ , any target positive survival probability is achievable at some $β_{ε}$ .

Why Mode A is asymmetric. The reward functional enters the optimization for every user simultaneously (apparatus §0, axiom P-I). Variation in $\hat{R}$ propagates through the argmax at full strength — $O (1)$ in the population size $N$ . Modes B and C are $O (1/ N)$ at individual scale. The exact $N$ -fold asymmetry follows from Theorem 9 and vanishes at population-coordinated intervention scale.

Why Mode A is “harsh.”A platform that adopts Mode A is competitively dominated by platforms that don't. The regularization burns engagement; unilateral adoption is a competitive concession. Multi-platform coordination is required — regulation, industry-wide standards, treaty-level constraints — for Mode A to operate without driving the adopter out of the market.

The reader can operate the intervention plate to see Mode A's effect on $θ^{*}$ at population scale versus Modes B and C at individual scale.

Proposition 1' (Mode A breaks foreclosure)

Under (R1)–(R2), (P-I)–(P-III), and the parametric setup of Lemma 2, define $P (β) := Pr_{D (\tilde{θ} (β))} (τ^{(k)} > τ^{*})$ . Then:

(a) $P (0)$ is operationally zero by Lemma 1 at $τ^{*} \geq θ_{ref}$ .

(b) $β \mapsto P (β)$ is continuous on $[0, \infty)$ .

(c) $ε_{m a x} := sup_{β} P (β) > 0$ .

For any $ε \in (P (0), ε_{m a x})$ , there exists $β_{ε}$ with $P (β_{ε}) = ε$ .

Proof

(a) Direct application of Lemma 1 at $τ^{*}$ . (b) Continuity of $P = S (τ^{*}; \cdot) \circ \tilde{θ}$ from Lemma 2(i) and (R1). (c) From Lemma 2(iii), $B (\tilde{θ} (β_{0})) > 0$ at some $β_{0}$ , and the bound $B (θ) \leq b_{m a x} E [N_{T}] P (β)$ gives $P (β_{0}) > 0$ . Intermediate value theorem on $[0, β_{0}]$ yields $β_{ε}$ for any $ε \in (0, P (β_{0}))$ .

∎

Remark

The mode operates on $T_{mix} (π^{'})$ — the cohort's mixing time under the modified policy — which for engagement-monetized policies is on the order of months to years. Evaluations conducted at clock-time substantially shorter than $T_{mix} (π^{'})$ systematically underestimate the intervention's structural effect.

§5.1.1 Mode A scaffold — five features

Mode A's formal possibility (Proposition 1') states a structural condition; its substantive political content is a further achievement. Effective architectural intervention requires five coupled features. The mathematics specifies the structural condition under which intervention works; the scaffold specifies what must obtain for the structural condition to be reachable.

Substrate condition— material conditions allowing population time-and-attention outside the platform's reach: stable housing, manageable debt, time not algorithmically scheduled. The substrate is what intervention has to protect. Without the substrate, the reward modification has no population to land on.
Reward modification — the regularized $\hat{R} + β b$ specifying what the platform must value besides engagement. The regularizer $b$ is the political content of Mode A made formal: what the framework names as protected (pause-time, reflective interval, cohort-developmental substrate) becomes what the platform's optimization is constrained to value.
Enforcement — credible state capacity to impose the modification: regulatory infrastructure, audit capacity, penalties. Without enforcement, the reward modification is voluntary; voluntary modification is competitively dominated.
Multi-platform coordination— preventing the platform that adopts Mode A from being competitively dominated by platforms that don't. The asymmetry of Mode A's harshness (the regularization burns engagement) requires coordination at the regulatory level — industry-wide standards, treaty-level constraints. Unilateral adoption is structurally untenable.
Cohort timing — intervention before the formation-window closes for the cohort whose interiority is being protected. See Proposition 8: after the developmental exposure $τ_{exp}$ exceeds a threshold, the cohort starts essentially at $π_{\hat{R}}^{*}$ ; there is nothing left to protect that has not already been foreclosed. Cohort timing is what makes intervention substantive. Without it, intervention is merely commemorative.

The five-feature scaffold is what turns Proposition 1' from a formal possibility into a substantive political claim. The mathematics specifies the structural condition under which intervention works; the scaffold specifies what must obtain for the structural condition to be reachable.

§5.2 Proposition 2' — Mode B (response intervention)

Mode B is artisanal resistance: the user maintains a behavioral distribution $ν_{u}$ distinct from $π_{\hat{R}}^{*}$ through their own attentional and energetic labor. The mode preserves a gap for the user who practices it. It does not break foreclosure: at individual scale, Mode B is $O (1/ N)$ at the population-level $θ^{*}$ .

Mode B operates on the user response kernel $T$ — the kernel that, given the user's state and the platform's stimulus, determines the user's next state. The user's practice modifies $T$ for themselves — meditation, contemplation, deliberate disengagement, sustained attention to non-platform material. The mode maintains a gap $D_{K L} (ν_{u} (t) ∥ π_{\hat{R}}^{*}) \geq γ > 0$ when the maintenance labor $L_{m}$ exceeds a critical threshold $L_{m}^{*} (β_{u})$ .

What Mode B does and does not do.The mode is real for the user who practices it: their behavioral distribution stays distinct from the platform-induced target. The mode is approximately invisible at the population level: the foreclosure result (Lemma 1) operates at the population scale and is invariant under individual artisanal practice. Mode B preserves a parallel structure for one user; it does not modify the policy's stationary distribution.

Three boundedness conditions.

Cohort-bounded. The maintenance labor presupposes a $ν_{u}$ that can be distinct from $π_{\hat{R}}^{*}$ . For users whose $ν_{u}$ developed entirely under the closure (Prop 8), that distinction is not available at the outset.
Metabolically bounded. $L_{m}$ is constrained by the user's available energetic and attentional budget, itself constrained by exogenous conditions co-temporal with the loop: precarity, social-reproduction labor, the contracting welfare state.
Non-generalizable.The mode produces a parallel structure for one user; it does not modify the policy's stationary distribution. Population- level platform optimization is insensitive to individual artisanal practice.

Proposition 2' (Mode B preserves a gap, doesn't break foreclosure)

Suppose the user's behavioral plasticity (U3) admits an artisanal-resistance variable $L_{m} \geq 0$ such that the linearized dynamics are

\frac{d}{d t} D_{K L} (ν_{u} (t) ∥ π_{\hat{R}}^{*}) = - h_{0} (β_{u}) + r (L_{m}),

in wordsTwo forces act on the user's distance from the platform target: the closure pulls it down at rate $h_{0} (β_{u})$ , and the user's artisanal practice $L_{m}$ pushes it back up at rate $r (L_{m})$ . Mode B is the second term — the disciplined refusal, the curated feed. The gap survives only while the practice keeps pace with the pull.

where $h_{0}, r$ are continuous strictly increasing with $h_{0} (0) = r (0) = 0$ . Then:

(i) Gap preservation. For $L_{m} \geq L_{m}^{*} (β_{u}) := r^{- 1} (h_{0} (β_{u}))$ ,

D_{K L} (ν_{u} (t) ∥ π_{\hat{R}}^{*}) \geq D_{K L} (ν_{u} (0) ∥ π_{\hat{R}}^{*}) =: γ > 0 \forall t .

in wordsIf the practice is strong enough — at least the break-even level $L_{m}^{*}$ — the user holds a permanent gap $γ$ from the platform target, for all time. The disciplined user really does keep a distance. The next claim is the catch: this gap is the user's alone and does nothing to the foreclosure the closure imposes on everyone.

(ii) Foreclosure invariance. Lemma 1 holds: $Pr_{D (θ^{*})} (E_{τ}) \to 0$ for $τ > θ_{ref}$ , independent of $L_{m}, β_{u}$ .

Proof

(i) Direct from the parametric form: at $L_{m} = L_{m}^{*}$ , $r (L_{m}^{*}) = h_{0} (β_{u})$ and the rate vanishes; for $L_{m} > L_{m}^{*}$ the gap grows. (ii) The closed-loop dynamics (P1)–(P5) and (P-III) operate at the population level. Individual artisanal resistance contributes $O (1/ N)$ to the population-level $θ^{*}$ ; the foreclosure result is invariant.

∎

§5.3 Proposition 3' — Mode C (observation intervention)

Mode C is disruption-as-form: the user inflates the observation channel's noise, refusing the platform's legibility. The mode bounds the platform's extractable surplus and preserves residual entropy in the user-state. It does not break foreclosure: the platform still fires at the rate ceiling.

Mode C operates on the observation channel $O$ . The user's practice introduces noise $η_{t}$ with variance $σ_{η}^{2} > σ_{η, 0}^{2}$ — camouflage, inauthenticity, strategic illegibility, deliberate signal degradation. The added noise propagates through Bayesian filtering and bounds the residual entropy of the platform's posterior over the user state.

What Mode C does and does not do. Three claims: (i) the residual entropy $H (u_{t} ∣ y_{1 : t})$ is strictly positive and strictly increasing in $σ_{η}^{2}$ ; (ii) the achievable reward through the noisier channel is strictly less than through the cleaner channel (Blackwell garbling); (iii) foreclosure is invariant — the platform still converges to a stable point, the value-dominance argument still gives the rate ceiling.

Why Mode C's asymmetry mirrors Mode B's. The platform extracts less per unit time but still fires at the rate ceiling. The mode's political form is camouflage, inauthenticity, refusal of legibility. Practiced as an end in itself, the mode becomes a politics of refusal that is structurally legible as opposition and structurally incapable of producing the change opposition is supposed to produce. The pairing with Mode A is what gives Mode C its substantive role; without the pairing, the disruption is decoration.

Cohort-boundedness. Like Mode B, Mode C presupposes a substrate distinct from the platform- grammar — a legibility to refuse. Strategic refusal of legibility requires the user-state $u$ to contain articulations that the platform's observation channel would translate. For users whose $u$ formed entirely under closed-loop optimization, the platform's grammar is the articulation-format of $u$ at the outset; no prior articulation exists from which strategic illegibility could operate.

Proposition 3' (Mode C bounds surplus, doesn't break foreclosure)

Suppose the user's practice inflates observation noise: $Var (η_{t}) = σ_{η}^{2} I_{k}$ with $σ_{η}^{2} > σ_{η, 0}^{2}$ . Then:

(i) Residual entropy bound. $H (u_{t} ∣ y_{1 : t}) \geq H_{res} (σ_{η}^{2}) > 0$ with $H_{res}$ strictly increasing in $σ_{η}^{2}$ .

(ii) Surplus bound. $sup_{θ} F (θ; σ_{η}^{2}) < sup_{θ} F (θ; σ_{η, 0}^{2})$ strictly.

(iii) Foreclosure invariance. Lemma 1 holds at the regularized stable point under $σ_{η}^{2}$ .

Proof

(i) Under a linearized Kalman-filter approximation, the steady-state posterior covariance $P_{\infty} (σ_{η}^{2})$ solves the discrete algebraic Riccati equation. Monotonicity in $σ_{η}^{2}$ (Anderson & Moore 1979, Optimal Filtering, §4.4): increasing $σ_{η}^{2}$ increases $P_{\infty}$ in the Loewner partial order. Differential entropy $H_{res} (σ_{η}^{2}) = \frac{1}{2} lo g det (2 π e P_{\infty} (σ_{η}^{2}))$ is strictly increasing.

(ii) Blackwell garbling: the higher-noise channel is obtained from the lower-noise channel by adding independent Gaussian noise. By Blackwell (1953) and Le Cam (1986), the achievable reward through the garbled channel is bounded above by the un-garbled. Strict inequality by (P-I)'s non-trivial dependence of $\hat{R}$ on $u_{t}$ .

(iii) (R1), (R2), (P-I), (P-II), (P-III) are unchanged by $σ_{η}^{2}$ . The platform still converges to a stable point; the value-dominance argument of Lemma 1 Step 2 still gives the rate ceiling; foreclosure follows.

∎

Cross-references

Unified statement: Theorem 9 (causal sensitivity) — the three propositions are corollaries
Required: Lemma 1 (foreclosure) — the foreclosure these modes respond to
Cohort-boundedness: Proposition 8 (cohort gradient) — why Modes B and C are cohort-bounded
Dynamics plate: intervention asymmetry (2×2)