---
title: "The Superdark Factory"
description: "Marek Poliks, Roberto Alonso Trillo, Tyler Dunn, Hugh Scott-Douglas and Jay Springett describe the limit case of automation: a \"superdark factory\" that sets its own objectives, rebuilds itself to pursue them, and cannot be understood from the inside by anyone, including itself."
type: "reading-notes"
authors:
  - "Marek Poliks"
  - "Roberto Alonso Trillo"
  - "Tyler Dunn"
  - "Hugh Scott-Douglas"
  - "Jay Springett"
published: "2026-09-08"
original_url: "https://superdark.antikythera.org/"
doi: "10.1162/ANTI.AW01"
canonical_url: "https://antikythera.wiki/work/journal/superdark"
md_url: "https://antikythera.wiki/md/work/journal/superdark"
last_updated: "2026-09-06"
site: "Antikythera Wiki"
---

# The Superdark Factory

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Marek Poliks, Roberto Alonso Trillo, Tyler Dunn, Hugh Scott-Douglas, Jay Springett
- **Published:** 2026-09-08
- **Kind:** Journal · Journal article
- **DOI:** https://doi.org/10.1162/ANTI.AW01
- **Original:** https://superdark.antikythera.org/
- **Length:** 71k words
- **This page:** https://antikythera.wiki/work/journal/superdark
- **Notes generated:** 2026-09-06

## Summary

Marek Poliks, Roberto Alonso Trillo, Tyler Dunn, Hugh Scott-Douglas and Jay Springett describe the limit case of automation: a "superdark factory" that sets its own objectives, rebuilds itself to pursue them, and cannot be understood from the inside by anyone, including itself. The article proceeds in three chapters. The first builds a vocabulary: a factory is a Guattarian machine that automates production; automation is a factory's production of its own input; factories fall into three classes according to what they take inside (executions, plans, objectives); and darkness is the condition in which descriptive information about a system stops being useful for predicting, auditing or steering it. Class 3 darkness, superdarkness, cannot be governed by reconstruction, so the human leaves the loop and returns as a strategic player whose only intervention is a single committed first move. The second chapter specifies that move as the "dark stack": composable primitives that seed populations of learners, versioning by behavior rather than state, a recursive and adversarial evaluation layer, and a charter cowritten with the factory. The third scales the picture up to an "anything factory," an ecology of superdark factories with its own convergence pathologies, and argues that governance of such an ecology has to be prepared now.

## The argument

The authors begin with an architect who wants to build a factory that exceeds their own capacity to design it: not a machine that does architecture on command, but one that does architecture to itself, forever. Such a factory must be bewildering to its maker; if it were not, the maker could have built the same thing more cheaply. That bewilderment is what the article calls darkness. Darkness is neither secrecy nor illegibility. It is the "decorrelation of information from utility": every log, schematic and snippet of code may be disclosed and readable, yet none of it supports prediction, audit or steering, because what the factory will do next is produced by a loop running against a world that has not yet arrived.

To make this precise the article classifies factories by input. A Class 1 factory automates execution and takes plans as input (a CNC mill, a compiler). A Class 2 factory automates plan-making and takes objectives (a coding agent, Bostrom's paper clip maximizer, a query optimizer, a Class 2 e-commerce experimentation engine). A Class 3 factory automates objective-setting and takes only norms and constraints, quantized into metrics, from its world. The von Neumann probe, which produces its own objectives but never renegotiates the norm it carries, marks the ceiling of Class 2. Class is a matter of where one draws the line around a set of relations, which is why the same sweatshop can be a Class 1 factory or, with its designers inside the line, a Class 2 one.

Darkness intensifies with class. A 3D printer is asymptotically resolvable. An agentic coding harness is dark not because of the language model at its center but because its plan is selected along a live, branching trajectory. Ocado's Andover warehouse stands in for the general case: an auditor holding the complete source code and every robot's spec sheet still cannot say where robot #1007 will be in twenty minutes. For a Class 2 factory a time series of snapshots plus the objective would in principle allow reconstruction. For a Class 3 factory nothing does, because the missing key is the factory's world, which cannot be handed across a table. Anyone who forces the factory to disclose the calculus by which it converts norms into objectives has, by that act, demoted it to Class 2.

The middle of the first chapter prices what the authors call neurotic governance, drawn from Stafford Beer's Viable System Model: a governor who supplies nothing to a machine they could not reconstitute outside of it. In Class 1 this costs nothing, since the governor sets the factory's pace. In Class 2 the factory runs on its own clock and the governor's rate *r* becomes a dial: performance rises with *r* and then plateaus at the rate of objective-setting, risk steepens, and the cost of the logging apparatus grows while its usefulness shrinks. The "control tower" of contemporary agentic deployments is this apparatus; the article traces its decay from total review to unread audit log to a single kill button. In Class 3 the dial moves inside the factory and neurotic governance becomes a contradiction in terms. The decision to build a Class 3 factory therefore turns on an unpriceable surplus P and a ruin cost S that can only be handled as Knightian uncertainty, with exposure caps and covenants. Refrain wins on paper, until someone proves P is real and a race begins.

The section titled "Radiant Darkness" recasts darkness as generative. Adrian Thompson's 1996 evolved FPGA circuit, which used five apparently disconnected cells and the physics of one particular piece of silicon, introduces three spaces: configuration space *a*, solution space *b*, and readable space *c*, the subset of configurations that can be backtraced entirely to inputs the governor supplied. Thompson's circuit and AlphaFold's folds lie in *b* ∖ *c*: they incorporate "auxiliary supply" from the world through channels no one provisioned. Novelty, on this account, always looks to the governor like a mistake, because anything that looked like a good decision was already something the governor knew how to want. In Class 3 the auxiliary supply becomes the whole supply and *c* becomes coextensive with *a*, readable only in the sense that the world is readable.

Governance then has to be redesigned as a game rather than a reconstruction. Borrowing Wiener's Manichean devil, cybernetics' steersman, and Stackelberg competition, the authors argue that the factory must carry a self-model in order to price costs and returns, that this self-model must be drawn from its wake in the world just as the architect's model is, and that the factory's identity must therefore be conferred from outside by a line the architect draws and commits never to cross. This commitment is the "Stackelberg move," the architect's first and only move inside the factory, and the dark stack is what it delivers.

The second chapter works through the four layers of that stack, using contemporary agentic software as its medium while insisting the principles are medium-agnostic. Primitives are composable, non-decomposable units whose contracts manufacture darkness at the smallest scale; primitive design is incentive design, seeding both mean-based no-regret learners at an exploratory frontier and no-swap-regret learners at a retentive core, since the difference between Stackelberg value V and the larger quantity U* is exactly what only the naive learners can reach. Versioning abandons Git for a behavioral definition: two configurations belong to the same version if their input–output distributions are statistically indistinguishable, measurable with transfer operators and spectral gaps, and the architect's kernel of hard casts is the one thing that cannot be revised without killing the factory. Evaluations are online, requisite in variety, massively recursive, graded from outside the loop they judge, and fed by a manufactured adversarial layer that produces the realized consequences against which verdicts are scored. The charter is the factory's input, a probabilistic spec of norms, metrics and constraints that the factory cowrites through sortition-based delegation and, at the level of constraints, through futarchic prediction markets. A final section on time argues that governance is best defined not by role but as the factory's slowest loop, and that the Stackelberg move must encode cascade-control ratios and jitter to prevent iatrogenic thrash and entrainment.

The third chapter treats the 2010 flash crash as a taste of superdarkness at economic scale and asks what follows once a superdark factory exists: Red Queen races between firms, reproduction of factories by factories, trophic ecologies anchored on lithography, frontier models and power plants, and the four convergence pathologies reappearing as flash crashes, monocultures, bubbles and too-big-to-fail institutions. Against "remainder humanism," which defines the human as whatever machines cannot yet do (here, generate demand), the authors argue that superdark factories can be consumers and even pure financial objects. The human is then not the factory's source of demand but a counterparty with its own material dependencies, engaged in strategic parasitism or a parallel economy. The closing claim is that the governance of the superdark factory entails the governance of the anything factory, and that it must be infrastructured now.

## Section by section

### Introduction and Chapter I, Introduction

The article opens twice, once as an abstract and once in earnest, on the same image: a factory that decides what to make, how, and why, and whose insides churn too fast to comprehend. The nearest instance is being assembled now in software, where coding agents design, build, review and ship with less human attention per decision every day, though most "software factories" merely reproduce assembly lines and waterfalls at higher speed. The architect who wants more must surrender the chair of discernment itself. The result is a "Copernican" humiliation and also a narrow new agency: one committed move before the factory is switched on, then a reserved seat at a table of games.

### I. On Factories and Darkness

A factory is a machine in Guattari's sense, defined by its relations rather than its parts and always open to revision (Simondon's "margin of indetermination"), as against a mechanism such as a clock. Production is transformation; automation is a factory's production of its own input, illustrated by the power loom drawing the weaver's labor inside and shrinking its input to a pattern. The three classes follow, with careful distinctions between a sequence (a washing-machine cycle) and a plan (a decision among real alternatives), and between an objective (instrumental) and a norm or constraint (non-instrumental, quantized into metrics with unavoidable tension). The von Neumann probe illustrates norms, constraints and metrics but fails Class 3 because it never weighs its norm against another. A Class 4 factory that automates its world is left undefined. Self-optimization is then treated class by class: calibration in Class 1 (autofocus, a mill adjusting to thermal expansion, already drawing on "auxiliary supply" from the world), exploration among alternatives in Class 2, and revision of the standard of judgment itself in Class 3. Darkness is defined as the withdrawal of firsthand knowledge that accompanies every act of automation, and traced from the 3D printer to the coding harness to Andover, ending with superdarkness as the limit case and the decision to treat the Class 3 interior as a [black box](https://antikythera.wiki/terms/black-box) useful only for speculation.

### II. The Human and the Loop

Beer's insight that management requires opacity is turned into the figure of the neurotic governor, whose capacity to reconstitute the factory is priced through performance *p*, risk *s* and governance cost *g* as functions of the rate *r* at which the factory outruns them. The section works through an e-commerce example and the "control tower" of human-in-the-loop AI deployment to show that the optimum *r** is greater than zero but bounded by the pace of objective-setting, that *g* buys down *s* at a diminishing exchange rate, and that a sufficiently performant factory turns the tower into a vestige. Once objective-setting itself becomes the competitive bottleneck, the market considers automating it. Class 3 cannot be governed neurotically because demanding a reconstitutive log is itself a Class 2 imperative. The upstream decision, whether to build a Class 3 factory at all, is framed as a détente broken by whoever proves P first. The section then argues against defining the human by what machines cannot do (the "God-of-the-gaps," Weatherby's "remainder humanism"), cites AlphaGo's Move 37, AlphaEvolve, AlphaDev, the Darwin Gödel Machine and POET as escalating Class 2 generativity, and concludes that the human is excluded from the superdark factory's interior not because it is worth less but because the factory abolishes the concept of an interior role. A human who insists on occupying the loop demotes the factory to a lesser class.

### III. Radiant Darkness

Neurotic governance has a hidden design cost: it confines the factory to what the governor knows how to want. Thompson's FPGA experiment motivates configuration space *a*, solution space *b* and readable space *c*, which can only be established in hindsight. Against a rate *r* = 0 governor, AlphaGo's Move 37 sits at the frontier of *c* (readable space no one had read), while AlphaFold's folds sit outside *c* altogether because the physics of protein folding entered the system only through decades of crystallography pressed into weights. Overfitting is reinterpreted: the Thompson circuit "overfit" only because an unstated acceptance criterion (generalizability) hid in the auxiliary supply, and a *b** in *b* ∖ *c* is by definition unpredictable. In Class 3 the spaces migrate to the level of objectives, *r* is set by the factory itself, and the interior becomes absolutely dark while remaining, through the mechanisms of the second chapter, predictable, auditable and steerable from its effects.

### IV. How to Move in the Dark

Cybernetics is invoked as the discipline of negotiating with black boxes. Wiener's Augustinian devil, which does not change its policy once found out, describes the neurotic governor's hope; the superdark factory is closer to his Manichean devil, a strategic player, though not one that intends anything against its architect. A Class 3 factory must price norms into objectives and so must have a self-model, but darkness that is not cryptographic means the factory must be dark to itself as a whole and can only model itself from its wake, as the architect does. Its identity therefore has to be given from outside, by a line drawn and never crossed. This yields the game's players and the "Stackelberg move": a first, visible, irrevocable commitment in Schelling's sense that places the factory permanently in a reactive position. The section closes by defining the four design levels of the dark stack.

### Chapter II, I. The Primitive

Software, the authors write, is now gardened rather than written. Primitives are the smallest composable, non-decomposable units (hooks, harnesses, subagents, MCP servers, agents); their contracts decorrelate interior from consumer utility and so manufacture darkness at the lowest level. Class can be read from who reads the contract and when. Genuine Class 3 composability would let an orchestration layer form spontaneously against constraints that did not exist when contracts were written; hard-coded agent pipelines are "just a waterfall." The design tension is thick versus thin, illustrated by the Composable Data Management System Manifesto, DuckDB's outward composability and inward monolith, and the history of query optimizers from System R through EXODUS and Cascades to adaptive execution. Two subsections follow. "Incentive Design for Continuous Learning" draws on robust mechanism design (Bergemann and Morris, Carroll's principle of robust simplicity) and on results about strategizing against no-regret learners to distinguish mean-based no-regret learners, which can wander outside readable space, from no-swap-regret learners, which converge to the best response within it. The Stackelberg value V becomes a ceiling for a committed architect; the gap to U* is harvestable only by proxy through the seeded population, yielding the article's central equivalence, U* > V if and only if *b** lies in *b* ∖ *c*. Blum and Mansour's construction shows no-swap-regret learners can be assembled from ordinary ones given counterfactuals. "Information Design for Continuous Learning" turns to protocols: Walden Yan's revisions on multi-agent context, Bayesian persuasion, a rich request channel and a thin but stateful reward channel carrying propensity scores, and a rule of minimal sufficient disclosure in which schematics are public, local state private, and propensity the sole exception.

### II. Versioning

Prefaced by Springett's "Hard Worlds for Little Guys," the section argues that a superdark factory loses statefulness at initialization and cannot be versioned by what it is. Following Beer's POSIWID, it is versioned by what it does: a version is a metastable input–output regime, detectable with transfer operators whose eigenstates and spectral gap indicate whether versions are durable or bleeding into one another. "Pathology Detection" names four convergence pathologies: stable failure (a robust attractor failing its input), overfitting (metric over norm), learning death (the exploratory frontier dead or quarantined, the factory parked at V) and thrash (permanent unsettledness). "Commitment and the First Commit" adopts Josh Stark's hard and soft casts: the Stackelberg move is a thin constitutional kernel of hard casts, inviolable from inside and outside, whose revision is the death of one factory and the birth of another. Soft casts are priced penalties applied through the reward channel as a Lagrangian term with a PID-controlled weight λ, and a niche of unhistoried compute is reserved against learning death.

### III. Evaluations

Unit tests and the SDLC waterfall assume isolable, time-invariant units and a preproduction environment isomorphic to production; the first wave of Class 2 software factories inherits this morphology with offline LLM-as-judge evaluations. Superdark design instead prioritizes online evaluation, since production is a stage of development. Guardrails are the first online evaluators and already fail at scale; LATTICE shows self-improving guardrails but only as a no-swap-regret learner. Four principles follow: evaluations must be online and continuous, contain Ashby's requisite variety of mean-based and swap-based judges, be massively recursive in Beer's sense, and be graded by a signal outside the loop they judge. "Catastrophes" imports Thom's folds and cusps and the early-warning-signal metrics of variance, autocorrelation and ensemble disagreement. "Overfitting and Adversarial Populations" adds a realized-consequence reward fixed at the Stackelberg move and, since real failures are sparse, an adversarial layer of red teams and deliberate severity-0 bug construction to manufacture them. Writing the first evaluation is the most direct act of governance available to the human at the boundary.

### IV. The Charter and the Loop

The charter is the factory's input, an editioned probabilistic spec of acceptable output distributions given constraints; StrongDM's Attractor is cited as a Class 2 precursor. It is a site of governance, worlding and negotiation, cowritten by a governance committee that must include a delegation from the factory, since a factory unable to participate in its own charter is Class 1 or 2 automation. Its softness is backed by the kernel's live-or-die threat, and its pricing establishes a Kantorovich marketplace with a non-fungible center. "Self-Writing" cedes the metrics layer and the shadow price λ to the factory, retains norms behind a read-only or power-sharing wall, and proposes sortition for delegation, with DAOs, the AgentCity papers and Hanson's futarchy ("vote on values, bet on beliefs") as models; futarchic λ becomes forward-looking. "Time and Phase" reinterprets the pathologies as loop failures: overfitting as aliasing, learning death as overlong periods, thrash as oscillation, stable failure as insufficient gain. "Clocking the Factory" defines iatrogenic thrash, adopts cascade-control ratios of 3:1 to 10:1, proposes system identification to time charter revisions, and, following Herbert Simon, defines governance as the slowest loop. Because emergence alone cannot guarantee band separation against entrainment, the Stackelberg move must encode minimum cascade ratios and jitter in the reward queue, while permitting Holling's productive "revolts."

### Chapter III. The Anything Factory

The flash crash of 2010, halted by a Stop Logic routine humans had encoded years earlier, shows humans as bewildered spectators of both machine failure and machine resolution. The superdark factory automates economy-shaped entities by singularizing them along a charter. Successors follow: Red Queen effects, reproduction of factories by factories (as reproduction, never as function), trophic ecologies with anchor assets at the base and a duty to "vectoralize" dependencies, and the pathologies at planetary scale. The authors take up Hemenway Falk and Tsoukalas's "AI Layoff Trap" and its Pigouvian and UBI remedies as instances of remainder humanism, argue that factories can generate synthetic demand and, with Poliks and Trillo's *Exocapitalism*, that they can be pure financial objects, and redefine the human as a counterparty with its own dependency inventory. "So much depends on our first move."

## Key concepts

- **Factory, machine, automation.** A factory is a Guattarian machine that automates production; automation is the factory's production of its own input, a redrawing of the boundary between inside and outside.
- **Class 1, 2, 3 factories.** Automation of executions (input: plans), of plan-making (input: objectives) and of objective-setting (input: norms and constraints, quantized as metrics). Class is set by where the line around a set of relations is drawn. Class 4 is undefined.
- **Darkness and superdarkness.** The decorrelation of descriptive information from utility for prediction, audit or steering; distinct from secrecy and illegibility. Superdarkness is the Class 3 limit, where the missing key is the factory's world.
- **Neurotic governance and *r*.** Beer-derived: a governor who supplies nothing they could not reconstitute outside the machine. Its cost is a speed limit on the factory, priced through *p*, *s*, *g* and their Class 3 successors P and S.
- **Auxiliary supply.** Input from the world that enters a machine through channels no governor provisioned; the source of Thompson's circuit and AlphaFold's folds.
- **Configuration, solution and readable space (*a*, *b*, *c*).** Generative darkness is the region *b* ∖ *c*, solutions the world supplied rather than the governor.
- **Stackelberg move.** The architect's single, visible, irrevocable first commitment, which confers the factory's identity and delivers the dark stack; afterwards the architect is a strategic player, not an operator.
- **Dark stack.** Primitive design (composability, incentive and information design), versioning (behavioral, transfer-operator based, anchored by a constitutional kernel of hard casts), evaluations (online, requisite-variety, recursive, externally graded, adversarially fed) and charter (negotiated, cowritten, priced).
- **No-regret and no-swap-regret learners; V and U*.** Mean-based learners at the frontier generate surplus; swap-regret learners at the core retain it. U* > V if and only if the best configuration lies outside readable space.
- **Convergence pathologies.** Stable failure, overfitting, learning death and thrash, each with a temporal reading and a PID-priced remedy.
- **Kernel, hard and soft casts, charter, λ.** The inviolable kernel and the revisable charter; λ as the shadow price of a constraint, determined with the factory and, futarchically, in advance.
- **Anything factory.** The ecology of superdark factories; the human as counterparty rather than source of demand.

## Connections

The article's closest neighbors in the journal are the ones its authors name. [Some Movements of Models](https://antikythera.wiki/work/journal/movementsofmodels) supplies the [black box](https://antikythera.wiki/terms/black-box) whose systems-engineering sense, a subsystem defined only by its relations, this article radicalizes into a factory defined only by its wake; where de Monchaux worries that a box hard to enter becomes impossible to understand, Poliks and colleagues make that impossibility a design goal. [Infinity Mirror](https://antikythera.wiki/work/journal/infinitymirror) and its [corrective loops](https://antikythera.wiki/terms/corrective-loops) describe planetary regulation built from feedback; the superdark factory's stack of loops, cascade-control ratios and Stop Logic kernel are the same problem at the scale of a firm, and the closing chapter returns it to the planetary scale of flash crashes and phase-locked condensates. [The Noocene](https://antikythera.wiki/work/journal/noocene) treats a mind-shaped Earth whose feedback is not designed by any single actor, which is close to what an economy of superdark factories would be. [AUTO](https://antikythera.wiki/work/journal/auto) frames automation as decisions embedded into environments that then constrain what comes next; the dark stack is a deliberate account of what to embed when the environment will thereafter be unreadable, and Sherman's "operator is dead" anticipates this article's abolition of the interior role.

Beyond the frontmatter, the article's insistence that specifying norms too precisely invites gaming, and that the charter must stay thin enough not to bound the factory artificially, parallels the warning against [alignment overfitting](https://antikythera.wiki/terms/alignment-overfitting) in [After Alignment](https://antikythera.wiki/work/journal/afteralignment), and its governance committee that includes the governed is a concrete instance of [recursive governance](https://antikythera.wiki/terms/recursive-governance). The recurring appeal to Bratton's "Copernican trauma" and to the God-of-the-gaps critique of defining the human by what machines cannot do situates the essay within the program's broader argument about synthetic intelligence, here pushed to the question of who, or what, gets to set objectives.


## Related works

- [AUTO–](https://antikythera.wiki/work/journal/auto)
- [Infinity Mirror](https://antikythera.wiki/work/journal/infinitymirror)
- [Some Movements of Models](https://antikythera.wiki/work/journal/movementsofmodels)
- [The Noocene](https://antikythera.wiki/work/journal/noocene)