The Silent War Is a Control Loop

In Cognition Becomes Terrain I argued that politics has extended war beyond geography. Once a population can be measured, sorted, acted upon, and measured again, war acquires another address. This essay gives that address a form.

It is a regulator. At time $t$, a population state is measured into a representation, a policy acts on that representation, and the action enters a transition that produces the population encountered at the next turn:

$$\Omega_t \;\xrightarrow{\;\varphi\;}\; X_t \;\xrightarrow{\;f\;}\; T_t \;\xrightarrow{\;\pi\;}\; A_t \;\xrightarrow{\;G\;}\; \Omega_{t+1}$$

Here $G$ is the transition from action to the next population state:

$$\Omega_{t+1}=G(\Omega_t,A_t,\xi_t).$$

$\xi_t$ gathers changes not produced by the policy alone: events, other institutions, collective responses, and everything else acting on the population during the interval.

Terrainization builds $\varphi$. The Proxy is produced by $f$. Segmentation is the partition induced by $T_t$. Distribution is what $\pi$ does. Feedback is the return from $\Omega_{t+1}$ to its next measurement. The diagram is a synthesis, not an organizational chart. The instruments already exist. They fit together, and their limits tell us where to look.

The experiments below use synthetic populations. They establish no deployment claim. Their purpose is narrower: to make the objects computable and expose the failure modes before we mistake a system’s internal account of itself for the territory.

Terrainization builds a σ-algebra

Let $\Omega$ be a population and $\varphi : \Omega \to X$ a measurement carrying each person into a feature space. The measurement fixes the σ-algebra

$$\mathcal{F}_\varphi = \{\varphi^{-1}(B): B \in \mathcal{B}_X\},$$

the distinctions available to any policy that receives only $X$. The full route from person to action is $\pi \circ f \circ \varphi$. If $\varphi(\omega)=\varphi(\omega')$, every deterministic policy that factors through this route assigns both people the same action. Randomization may give them different draws, but only from the same action distribution. Nothing downstream can condition on a difference that the measurement never recorded.

This is why measurement is not a neutral prelude to decision. It decides which differences can become actionable and which remain outside the map. Before the map there is no target of this particular kind. The construction of $\varphi$ is already part of the operation.

The map can be detailed and still be unfaithful. More columns do not resolve the problem if the columns were chosen for another purpose, if the trace is a poor stand-in for the thing named, or if the categories force unlike lives into one administrative kind. Fidelity is not the same as resolution. The question is who drew the map, what they needed it to do, and which decisions will treat its coordinates as though they belonged to the territory itself.

The Proxy is a decision-relative compression

The measurement sets the ceiling. The encoder $f$ works beneath it. Let $T=f(X)$ be sufficient for a decision problem $(Y,\ell)$ when acting through $T$ costs nothing against acting through $X$:

$$\inf_{\pi:T\to A}\mathbb{E}[\ell(\pi(T),Y)] = \inf_{\pi:X\to A}\mathbb{E}[\ell(\pi(X),Y)].$$

Exact sufficiency is one endpoint. Most operational representations are lossy. The Proxy is better understood as a point on a decision-relative compression frontier: how much of $X$ can be discarded before loss, as the institution defines it, exceeds a chosen budget?

One way to write the penalty is excess conditional risk,

$$d_\ell(x,t)=\mathbb{E}[\ell(a_t,Y)\mid X=x]-\min_a\mathbb{E}[\ell(a,Y)\mid X=x],$$

where $a_t$ is the best action available from the compressed state $t$. But $a_t$ depends on the encoder $p(t\mid x)$, so the distortion changes with the object being optimized. The expression below is therefore a self-consistent decision-relative construction, not a standard rate–distortion problem with a fixed distortion measure. Standard convexity, Blahut–Arimoto, and operational coding guarantees do not follow unless the decoder is fixed in advance.

Within that qualification, the construction asks for the least information rate compatible with an expected distortion budget,

$$R(D)=\inf_{p(t\mid x):\,\mathbb{E}[d_\ell]\le D}I(X;T).$$

The politics enters through $Y$, $\ell$, and the budget $D$. The discarded material is not without value. It is material the chosen decision has priced below the cost of carrying it. A life can disappear at that margin while the system remains perfectly faithful to its objective.

Naftali Tishby, Fernando Pereira, and William Bialek make a related construction in The Information Bottleneck Method. Their objective is

$$\min_{p(t\mid x)} I(X;T)-\lambda I(T;Y),$$

under the Markov chain $Y-X-T$. At a self-consistent solution, this has the form of rate–distortion with the induced distortion $D_{\mathrm{KL}}[p(y\mid x)\|p(y\mid t)]$. That is a precise correspondence, not permission to identify information bottleneck with every decision-loss problem. It also contains the recursion that matters here: $p(y\mid t)$ depends on the encoder whose quality it is being used to judge.

Change $Y$ or the loss and the representation changes. Two institutions can hold the same traces on the same people and nevertheless produce incompatible Proxies.

I ran that possibility on twenty thousand synthetic people with eight attributes. Three fitted scoring rules used the same rows and columns for three different objectives: repayment risk, engagement, and persuadability, the last modeled as a treatment effect. Each score partitioned the population into quintiles:

labels = [np.digitize(s, np.quantile(s, np.linspace(0, 1, 6)[1:-1])) for s in scores]
ari = adjusted_rand_score(labels[i], labels[j])

The largest Adjusted Rand Index between any two partitions was 0.0018, where zero indicates agreement no better than chance after adjustment. No feature appeared in the top three for all objectives, though engagement and persuadability both used social_degree. The run does not prove that these are unique or globally optimal representations. It shows something simpler: even fitted rules over identical traces can partition one population almost independently when their objectives differ.

Answerability is a reachability problem

The earlier essay located the injury in consequence without answerability. Here answerability becomes reachability under the person’s own authority.

Give a person an action set $\mathcal{A}(x)$ and a budget $B$ under cost $c$. Their reachable set is

$$\mathcal{R}_B(x)=\{x\oplus a:a\in\mathcal{A}(x),\;c(a)\le B\},$$

and recourse exists when this set intersects the region assigned the desired decision. Berk Ustun, Alexander Spangher, and Yang Liu formalized this problem in Actionable Recourse in Linear Classification. Their general method handles discrete and constrained actions through integer programming and can certify infeasibility relative to a declared action set.

My synthetic case asks a narrower and stronger feasibility question: can any continuous move within the declared box bounds cross the decision boundary, even without a cost ceiling? It is not a budgeted recourse calculation. For a linear score, the answer depends only on whether the total beneficial score reduction available within those bounds can close the gap:

gain = {j: -coef[j] * direction[j] for j in actionable}
moves = [j for j in actionable if gain[j] > 0]
feasible = sum(gain[j] * ceiling[j] for j in moves) >= s_x - tau

The script computes effort cost for cases that can cross, but neither cost nor $B$ enters the 53.42 percent feasibility count.

With income allowed to rise by one fifth, tenure by one year, and device count by two, 53.42 percent of five thousand rejected synthetic cases had no feasible recourse within those bounds, even at unlimited cost. Because the model was fitted on standardized features, 26.86 percent of absolute coefficient weight sat on features the action model declared immutable. Change the bounds and those numbers change. Their force lies in the declared relation between score and action set: predictive performance and reachability are different properties of a rule, and optimizing one does not supply the other.

Availability and exposure are different operators

Publication places an item in a set. Exposure gives it a route through that set. Under a position-based examination model,

$$P(C_i=1\mid k_i)=P(E_i=1\mid k_i)\,P(C_i=1\mid E_i=1,i).$$

Under the stronger noise-free assumption that examination plus relevance produces a click, the second factor becomes the item’s relevance probability. Censorship removes an item from the published set. Modulation changes its rank while leaving it technically available. An audit that enumerates the set tests the first operation and can miss the second entirely.

I simulated two thousand items in two groups. Both drew relevance from the same distribution, so they were equal in expectation. No item was removed; one group received a one-logit ranking penalty. In this realization the unpenalized group’s mean latent relevance was 0.2504 and the penalized group’s was 0.2519. Total examination probability differed by 4.327×; because the realized group sizes were 986 and 1,014, mean examination probability differed by 4.450×. The naive click-rate ratio was 10.28×.

Before the Bernoulli draw, the expected click-rate ratio was 6.27×. Its excess over mean exposure came from stronger positive covariance between examination and relevance in the unpenalized group: its high-relevance items sat where the position curve was steep, while the penalized group’s occupied the shallow tail. The remaining jump to 10.28× came from finite sampling—ten clicks for the unpenalized group and one for the penalized group. The ranking process produced the expected gap; the draw amplified it and returned both as a property of the items.

Thorsten Joachims, Adith Swaminathan, and Tobias Schnabel’s Unbiased Learning-to-Rank with Biased Feedback shows how inverse propensity weighting can produce an unbiased ranking-risk estimate under a declared observation model with positive propensities. In the simpler noise-free item model used here, $C_i/e_i$ is unbiased for latent relevance because $\mathbb{E}[C_i]=e_i\rho_i$. Its variance is

$$\operatorname{Var}(C_i/e_i)=\frac{\rho_i(1-e_i\rho_i)}{e_i},$$

which grows as examination probability falls. With $e_i=k_i^{-0.9}$, the last rank has $e_i\approx 1.1\times10^{-3}$. The run made that instability visible: raw group means after inverse weighting were 0.1815 for the unpenalized group and 0.5561 for the penalized one, reversing the naive ordering and grossly overshooting the magnitude of the latent difference. At those per-item variances, the sign from a single draw cannot support an inference about the latent gap.

The script then floored propensity at rank 150 and self-normalized within each group. That operation produced differently normalized quantities, so their former 1.87 ratio was not a residual-bias measure and cannot be read as movement toward parity. For a group mean, the per-item clipped IPS estimator is

$$\widehat{\rho}_g=\frac{1}{n_g}\sum_{i\in g}\frac{C_i}{\max(e_i,e_{150})}.$$

On the same draw, its per-item clipped IPS means were 0.1570 for the unpenalized group and 0.0896 for the penalized one. The floor biases tail items downward, and the sample contains only one Bernoulli draw per item; neither the levels nor their ratio establish parity. Part of this was visible before the audit ran: the largest weight is approximately $1/e_i$, so the tail announces its own variance problem.

The floor that tames that variance is not neutral. Its expected bias is asymmetric: the group with less exposure sits deeper in the ranking, absorbs more clipping, and is pushed further below its latent relevance—the same direction the ranking penalty pushed it. In this draw the two estimators fail in opposite directions. The raw one overshoots the penalized group and hides the penalty. The clipped one undershoots it, reverses the latent ordering, and returns the modulation as a property of the items, which is the failure this section opened by naming. Choosing where to floor the propensity is choosing how much of the modulated tail the audit is willing to see.

The fixed point moves

Ordinary risk minimization assumes that the data distribution stays put. Deployment can break that assumption when the distribution becomes a function of the deployed parameter, $\mathcal{D}(\theta)$. Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt distinguish two solutions in Performative Prediction. A parameter is performatively stable when retraining on the distribution it induces returns it:

$$\theta_{\mathrm{PS}}\in\arg\min_{\theta'}\mathbb{E}_{Z\sim\mathcal{D}(\theta_{\mathrm{PS}})}\ell(Z;\theta').$$

It is performatively optimal when it minimizes performative risk across deployments:

$$\theta_{\mathrm{PO}}\in\arg\min_{\theta}\mathbb{E}_{Z\sim\mathcal{D}(\theta)}\ell(Z;\theta).$$

The points need not coincide. Repeated risk minimization targets stability, not necessarily performative optimality.

Let $\mathcal{D}$ be $\varepsilon$-sensitive in Wasserstein distance, and let the loss be $\beta$-jointly smooth and $\gamma$-strongly convex. Perdomo and his coauthors prove that repeated risk minimization is contractive and converges to a unique stable point when

$$\varepsilon<\gamma/\beta.$$

$\varepsilon$ measures how strongly the population distribution can move when the rule moves. Crossing the threshold does not prove that every system will cycle. It removes this convergence guarantee; the same paper constructs divergence once the conditions are relaxed. The distinction matters. A sufficient bound tells us where certainty ends, not what every process must do beyond it.

The simulation below carries the proxy state from one round into the next. It therefore sits outside Perdomo and his coauthors’ memoryless map $\mathcal{D}(\theta)$. Gavin Brown, Shlomi Hod, and Iden Kalemaj call the broader model class Performative Prediction in a Stateful World: the population’s response depends on both the deployed model and its current state. This code illustrates stateful performativity; it is not an empirical test of either paper’s convergence results:

for _ in range(ROUNDS):
    theta = fit(proxy, outcome)
    observed.append(score(theta, proxy, outcome))
    reference.append(score(theta, x_reference, y_reference))
    proxy = proxy + response(theta, proxy)

Across twelve rounds, accuracy on the model-influenced data fell from 0.7971 to 0.6773. Accuracy on a fixed population sampled before the response process fell from 0.7900 to 0.5050. The correlation between the moving proxy and its latent source fell from 0.9705 to 0.7028.

The cumulative update makes the proxy drift away from the fixed reference by construction. The approach toward chance accuracy demonstrates that chosen drift, not a general law of deployed prediction.

The fixed population is not “truth.” It answers a particular counterfactual question: how does the retrained rule perform against the pre-response distribution? Randomized holdouts, shadow policies, delayed outcomes, and independent measurements can open a view beyond the immediate training log, each under its own assumptions and costs. A system evaluated only on the distribution it has helped produce can improve, decay, or stabilize according to its own account while its relation to an external reference changes in another direction.

Ashby’s bound, and what remains outside the map

W. Ross Ashby’s An Introduction to Cybernetics gives one constraint on the regulator. In his formulation, disturbance can be reduced only by the variety available to the regulator: “variety can destroy variety.” The data-processing inequality gives another constraint, from information theory. For the Markov chain $Y-X-T$,

$$I(T;Y)\le I(X;Y),$$

with equality exactly when $I(X;Y\mid T)=0$, meaning that $T$ preserves all information in $X$ about $Y$. These are related in the argument I am making, but they are not the same theorem. Ashby concerns the regulator’s capacity against disturbance. Data processing concerns information that cannot increase through an encoding.

The encoder also induces fibres, $f^{-1}(t)=\{x:f(x)=t\}$. Any policy that receives only $T$ must assign every point in one fibre the same action distribution. That does not make the people unreachable. It makes their differences unavailable as grounds for differential action. An enforcement policy can still reach all of them at once, precisely because the map cannot tell them apart.

This locates one part of the counter-practice described in the earlier essay. Symbols, slang, and altered spelling may disturb a classifier, a ranking rule, or a moderation system. A political argument placed among shitposts can pass first through a shared idiom. The same position moving across many pages formed around different scenes can keep another route open when one page disappears. None of this guarantees that two forms remain inside a fibre. Whether they are equivalent depends on the actual encoder. Sometimes altered language crosses the boundary the classifier uses and opens a temporary passage. Sometimes it produces exactly the signature the system was trained to capture. The question is empirical: which representation is operating, which difference does it register, and which action follows?

Those routes can also make the network easier to identify. Ambiguity buys time, not invisibility. The point is not to romanticize evasion but to understand the contest over representation: one side tries to acquire enough variety to classify and act; the other changes form, multiplies routes, and forces the map to be redrawn.

The title names a loop, but the formal treatment is uneven. $\varphi$, $f$, and $\pi$ each receive a theorem or an experiment. $G$ receives a symbol, and $\xi_t$ a sentence. This essay does not formalize that closure. It specifies no loop-level setpoint or error signal, and therefore gives Ashby’s regulator no objective against which its variety can be counted. That omission is load-bearing: choosing whose objective governs, over what horizon, and with what error signal is where the political claim acquires its force. Naming that gap is the boundary of this essay; resolving it is a different argument.

This is the control loop I mean by silent war. Its measurement determines which differences can enter decision. Its Proxy retains what an objective rewards. Its policy shapes exposure and recourse. Its subjects respond, so its future data are partly the residue of its earlier actions. None of the cited theorems proves the whole political claim. Together they identify constraints that recur wherever the circuit is built.

The fibres are not a sanctuary. They are temporary limits in a contest whose better-resourced participant can usually rebuild the sensor. Still, they tell us what a real investigation must find: who drew the map, which traces became actionable, which differences it discarded, which decisions used it, and how the people living in the territory changed once the map began acting upon them.