The emergent properties of the simulation depends on this elementary arragement: a model observes a population, the model’s estimate is returned to that population as a force, and the altered population becomes the model’s next body of evidence. The result is mathematically ordinary and politically strange I think!
1. A local world
Agent i has a position pi,t ∈ ℝ² and velocity vi,t ∈ ℝ². In the untreated flock, its acceleration is the weighted sum of three neighborhood rules: separation from nearby bodies, alignment with nearby velocities, and cohesion toward the local centre. This follows the distributed flocking model introduced by Craig Reynolds.5 So in this sim, thre is no leader and no inherent global plan.
ai,tnatural = βsSi,t + βaAi,t + βcCi,t
Emergence here has a precise, modest meaning: an ordered collective pattern is produced from local interaction. .
2. The statistic that stands for the whole
At time t, the model does not inspect every relation in the flock. It receives one compressed sign: the centroid of the agents still counted as visible. If Vt is that visible population, then
ct = (1 / |Vt|) Σi∈Vt pi,t.
The centroid is a loss-producing summary of the flock. Dispersion, factions, collisions, and abandoned trajectories disappear into two coordinates. The project begins its critique at this reduction, before any prediction is made.
3. A two-step memory
Each coordinate of the centroid is normalized and passed to an autoregressive model of order two. The estimate is linear:
ĉt+1 = b + w1ct + w2ct−1.
The browser recursively applies this rule nine times to place the ghost roughly 0.6 seconds ahead. After a new observation arrives, the coefficients take a small stochastic-gradient step. With feature vector φt = [1, ct−1, ct−2] and error et = ct − θtTφt, the update is
θt+1 = θt + ηetφt.
This belongs to the broad tradition of stochastic approximation: parameters are revised from sequential, noisy evidence rather than fitted once to a completed dataset.6 The model has almost no memory. That poverty is useful. It shows how little machinery is required for a feedback system to manufacture the regularity it later reports.
4. When a forecast acquires force
The influence control λ ∈ [0,1] determines how much of the ghost’s displacement returns as acceleration:
ai,t = ai,tnatural + λκ(ĉt+h − pi,t).
At λ = 0, the prediction remains an observation. Above zero, it enters the causal history of what it predicts. This is the project’s version of a self-fulfilling prophecy: a representation evokes conduct that makes the representation newly credible.4 Donald MacKenzie’s account of financial models supplies a useful parallel. A model can become an engine within a process rather than a camera outside it.8
5. Performativity without magic
The forecast does not create truth by declarationsince its power is cumulative. A minor correction enters the next centroid; the corrected centroid is stored as data; the next forecast cites that record as evidence. Butler’s account of performativity is relevant at the level of repeated, socially compelled acts: repetition sediments into the appearance of substance.2
The system therefore stages performativity as a difference equation. Each iteration is weak bcuz their sequence is strong. Derrida’s account of iterability adds a further complication: a sign must be repeatable beyond the occasion of its production.7 Here, the forecast detaches from the instant that produced it and returns as an instruction for a later instant.
6. Modulation and disappearance
No boid receives an order! The field is continuously tilted by small adjustments. This resembles what Deleuze calls modulation: control operates as an ongoing deformation rather than a passage between fixed enclosures.3 At high influence, assigned angular phases organise the surviving agents into an orbit. The ring looks spontaneous only because coercion has been distributed over time.
An agent outside the accepted radius loses opacity. Once it disappears, it is removed from Vt; its dissent no longer affects the centroid or the training stream. Prediction error can then decline for two reasons: the population has been moved toward the estimate, and contrary cases have ceased to count.
7. The map and its territory
Baudrillard’s language of simulation concerns more than a false image laid over an intact reality. The sharper problem is the precession of the model: signs organise the field in which their referents can subsequently appear.1 The ghost centroid is rendered before the flock arrives. With enough influence, the flock becomes the delayed materialisation of its own statistical double.
8. The ungoverned twin
A second flock begins from identical positions and velocities and receives the same direct disturbances, but never receives predictive force. Its centroid ct0 supplies the displayed counterfactual distance:
Dt = ‖ct − ct0‖2.
This twin is not metaphysical ground truth. It is a control condition inside the same coded environment. Its purpose is comparative: high agreement with the forecast can coexist with increasing distance from the trajectory that would have unfolded without forecast-driven intervention.
How to read the instruments
The displayed agreement is explicit about the system’s changing standard of success. Let Ah be the bounded score derived from recent horizon errors and As the bounded spatial proximity of the governed centroid to the public ghost. The interface reports
Adisplay = (1 − 0.9λ)Ah + 0.9λAs.
Consequently, increasing λ does not merely change the system’s behavior; it shifts the meaning of successful prediction toward conformity with the forecast. This is a deliberate epistemic feature of the artwork and would be unacceptable if presented as an ordinary held-out accuracy measure.