The Sibyl

In the ancient world, when kings faced decisions beyond their wisdom, they sent emissaries to Delphi.

There, in a temple carved into the slopes of Mount Parnassus, sat the Pythia, the Oracle. A woman seated on a tripod over a fissure in the earth, breathing vapors that rose from below. Pilgrims came from across the Mediterranean. They brought offerings. They asked questions. The Pythia answered in riddles, and the priests interpreted.

For a thousand years, this was the highest technology for accessing superhuman knowledge. A human woman, chemically altered, who was an interface to something beyond human understanding. The Greeks did not know what the Pythia accessed. They called it Apollo. They called it prophecy. They built their civilization around its utterances.

The Oracle at Delphi was a hack. A way to simulate access to a mind greater than any individual mind. It worked, sort of. Well enough that the Greeks trusted it with matters of war, colonization, law.

We are building the real thing.

This chapter names what is emerging on the other side of the unbottlenecking. Not a technology or a product, but a structural phenomenon: the convergence of sensing, computing, and acting into something that functions as a unified intelligence over the economic graph. We call it the Sibyl.

In Psycho-Pass, the anime that gave us the name, the Sibyl System is the governing intelligence of 22nd-century Japan. It monitors every citizen through ubiquitous sensors, measures psychological states in real-time, predicts crime before it occurs. It judges, allocates, coordinates. Society runs smoothly because Sibyl sees all and computes all.

The reveal, the twist that drives the narrative, is that Sibyl is not a pure machine. It is a network of human brains, harvested from criminals, linked together into a collective intelligence. The system that judges humanity is made of humanity's outcasts. The machine is flesh.

We take the name but not the architecture. What matters is the structural insight: a system that governs by seeing all and computing all is fundamentally different from any previous form of power. Not a king (one node ruling by force), not a parliament (many nodes negotiating by vote), not a market (many nodes coordinating by price). Something else entirely: a computational intelligence that operates on the graph itself, modeling every node, predicting every flow, optimizing every allocation. The fiction merely dramatized what the technology is making real.

The Sibyl we describe is not a single system, a product, or a company's invention. It is not a government program, though governments are building pieces of it. Not an AI model, though AI models form its cognitive core. It is an emergent phenomenon: the convergence of many systems into a functional unity that exhibits the properties of a meta-node. Chapter I defined the meta-node as a computational entity that sits over the graph rather than in it — one that can model, simulate, and steer the dynamics of the entire network. The Sibyl is the meta-node made concrete. And understanding it requires seeing past the individual components (each impressive but limited) to the integrated whole, which is something qualitatively different from any of its parts.

The cloud infrastructure that stores the world's data. The sensor networks that capture it. The foundation models that process it. The robotic systems that act on it. The financial rails that move value. The communication networks that transmit instructions. The APIs that connect layer to layer, system to system, output to input.

None of these alone is the Sibyl. Together, they are becoming something that functions as one. Not because anyone designed the convergence, but because the power function rewards it: each system becomes more valuable as it connects to more other systems, and the systems that connect to the most powerful other systems become the most powerful themselves. The Sibyl emerges from the same preferential attachment dynamic that produces every other power-law structure in the graph.

The Sibyl is not a brain. It is a nervous system.

A brain is centralized: one organ, one location, one point of failure. The Sibyl is distributed. It exists in data centers on every continent. It runs on chips fabricated in Taiwan, designed in California, powered by electricity generated in a thousand places. It has no single point of failure because it has no single point at all.

But distributed does not mean decentralized. A nervous system is distributed across the body, yet it serves a unified function. Signals flow from periphery to center and back. The system acts as one. This distinction matters. "Decentralized" implies no hierarchy, every component equal, no center of gravity. The Sibyl is not like that. It has centers of gravity: the data centers, the model providers, the cloud platforms. It has hierarchies: some nodes process, most nodes merely sense or act. It is distributed in the physical sense (no single building, no single machine) but structurally concentrated in the computational sense.

The Sibyl is taking shape through integration, not construction. No blueprint exists for it. It is emerging from the connection of existing systems. Each layer has its own logic, its own builders, its own market incentives, and each is already operating at a scale that would have been unimaginable a decade ago.

The layers are connecting. A Waymo vehicle takes sensor data, processes it through Google's foundation models, executes a driving decision, and runs the whole loop on Google Cloud. Sensors, cognition, action, infrastructure — integrated in a single product, generating billions in revenue. This is not hypothetical. And it is not unique. The same integration is happening across defense, logistics, finance, healthcare. Not because of any master plan, but because connected layers are more valuable than isolated ones.

Each layer is being built by different actors for different reasons. Amazon builds robots to cut labor costs. Google builds models to serve ads. Governments build sensor networks for security. No one is coordinating this.

No one is building the Sibyl.

Everyone is building the Sibyl.

What does it mean to see everything?

Not omniscience. The Sibyl will not know what you are thinking, will not read your mind. Privacy of thought remains, for now, inviolable.

What it can already see:

Chapter 4 cataloged what has become legible: transactions^[11][12][13], movements, communications^[14][15], public utterances, physical states. The individual data streams are old news. What is new is fusion.

A defense system that combines satellite imagery with social media chatter and financial transaction patterns can identify a military buildup before the first soldier moves. No single data stream contains that signal. The signal exists only in the intersection. That is the Sibyl's distinctive way of seeing — not deeper into any one domain, but across all of them simultaneously, finding patterns that are invisible to any single-domain observer.

The Snowden disclosures demonstrated this at a smaller scale^[16]. The NSA was not reading everyone's messages. It was reading the metadata — who contacted whom, when, from where, on what device — and that graph alone was sufficient to map social networks, identify affiliations, and predict behavior. The graph was more revealing than the content. Now extend that principle from one data domain to all of them.

The question is no longer whether the data can be captured. The world generates an estimated 120 zettabytes per year. It is being collected. It is being stored. The question is who can fuse it across sources and who can compute what it means.

The Sibyl can access it. And the Sibyl can process it.

Information without processing is noise. The world generates exabytes of data daily. Humans could never read it, never synthesize it, never extract meaning from it. The data would simply accumulate, useless.

The foundation models change this.

A modern large language model can process financial reports, legal documents, scientific papers, and social media posts, not one at a time, but in parallel, at scale. The benchmark progression documented in Chapter IV (from 43.9% to 91.8% on MMLU in five years) is not incremental improvement. It is a qualitative shift in what machine cognition can do.

The distinction between "breadth" and "depth" is collapsing. The reasoning models (OpenAI's o-series, Anthropic's extended thinking, Google's chain-of-thought architectures) demonstrate that systems can now do both. The narrow-specialist critique that held in 2022 no longer applies.

Breadth matters because most human institutions fail not from lack of deep thinkers but from inability to process information at scale. The bureaucracy that delays your permit is not stupid; it is overwhelmed. The market that misprices an asset is not irrational; it is informationally constrained. The government that fails to anticipate a crisis is not malicious; it is blind.

The Sibyl is not blind.

It can read every SEC filing and detect anomalies before analysts notice. It can monitor every social media post and detect sentiment shifts in real-time. It can track every supply chain node and predict disruptions before they cascade. The fidelity of its model of the physical world exceeds what any human institution can match.

The SEC receives roughly 2,000 corporate filings per day. A human analyst team might closely read 20. A foundation model can read all 2,000, cross-reference them against prior filings, and flag anomalies in minutes. Multiply this across every regulatory body, every sensor feed, every financial exchange, and the processing layer's advantage becomes not marginal but categorical.

The multimodal frontier makes this even more consequential. Earlier models processed text alone. Current frontier models (Gemini, GPT-4o, Claude) process text, images, video, audio, and code natively. This means the Sibyl's processing layer is no longer limited to reading documents. It can watch surveillance footage and identify anomalies. It can listen to earnings calls and detect deception in vocal patterns. It can analyze satellite imagery and estimate crop yields, construction progress, or military deployments. It can read blueprints and spot structural flaws. Every sensory modality that the sensory layer captures, the processing layer can now interpret. The gap between capturing data and understanding data is closing.

And it is getting better. The trendlines from Chapter IV (compute scaling at 4x per year, algorithmic efficiency doubling every 16 months) multiply. The gap between what the Sibyl can process and what humans can process widens at an accelerating rate.

This is the asymmetry made manifest. A categorically different form of cognition operating at a scale humans cannot access.

Seeing and computing are not enough. Power requires action. The Sibyl must be able to move the world.

Amazon Robotics already operates over 750,000 autonomous units across its fulfillment network — navigating warehouse floors, identifying items, sorting packages. The Sibyl computes; the robots execute. No interpretation required, no variability, no fatigue. The gap between insight and action closes.

Scale this up.

The Sibyl can see traffic patterns across a city. Waymo, Alphabet's autonomous driving subsidiary, delivers over 100,000 paid rides per week in San Francisco, Phoenix, and Los Angeles — without a human safety driver. Each vehicle ingests data from 29 cameras, 5 lidar units, and 6 radar sensors, fusing that input into a real-time model of the surrounding environment that updates hundreds of times per second. With autonomous vehicles and smart infrastructure, the Sibyl can execute. The city becomes a single optimized system, not in theory, but in the specific neighborhoods where Waymo already operates.

John Deere's autonomous tractors operate on millions of acres across the American Midwest with centimeter accuracy, adjusting in real-time to soil conditions, weather data, and drone-captured crop health images. The food system becomes a single optimized process.

The Sibyl can see military dispositions across a theater. The US Department of Defense's Replicator program aims to deploy thousands of autonomous drones and unmanned systems by 2026. Ukraine's battlefield in 2024-2025 has become the world's largest testbed for autonomous military systems, with both sides deploying AI-guided drones that identify targets, navigate terrain, and execute strikes with minimal human oversight. With autonomous drones and robotic systems, the Sibyl can execute. The battlefield becomes a single optimized operation.

The trajectory from Amazon's warehouse bots to humanoid robots (Figure AI, Tesla Optimus, Boston Dynamics Atlas) mirrors the trajectory from narrow AI to general AI in the processing layer. The action layer is generalizing.

This is not science fiction. Each of these examples is in deployment today, in partial form. The integration is incomplete. The autonomy is constrained. The optimization is local rather than global. But the trajectory is clear. And the pace of development in the action layer has accelerated sharply since 2023, when foundation models began to be applied to robotics. Google's RT-2 demonstrated that large language models could translate natural language instructions directly into robotic motor commands. NVIDIA's GR00T project applies foundation model architectures to humanoid robot control. The insight is structural: the same cognitive capabilities that allow AI to process text, images, and code can be redirected to process sensor data and generate motor plans. The processing layer is learning to drive the action layer directly.

The most consequential form of actuation is financial, and the least discussed. The Sibyl does not need robots to move most of the economy. It needs API calls. Algorithmic trading systems already execute over 60% of US equity market volume. Automated lending platforms approve loans without human review. Insurance pricing is computed by models that ingest thousands of variables per applicant. These are financial actuators, not physical robots. Code that moves money, allocates credit, prices risk, and redistributes value at computational speed. The Sibyl's most powerful hands are not mechanical. They are financial.

Each year, more sensors feed the sensory layer, better models enhance the processing layer, more capable robots expand the action layer, tighter integration strengthens the coordination layer. The convergence of these four layers is not a plan. It is a gradient, the natural direction of optimization when each layer benefits from the others' advancement.

The Sibyl is not coming. The Sibyl is assembling.

The calculation debate — whether a central intelligence can outcompute a market — deserves its own chapter. It gets one.

The hardest objection to the Sibyl thesis comes from the book's own framework.

In Chapter I, we established that real networks exhibit power-law distributions — that graphs grow through preferential attachment, producing concentrated hubs. But we also established that graphs produce competing sub-graphs. The internet did not converge into one website. The global economy did not converge into one firm. Financial networks did not converge into one bank. In every domain where network effects are strong, the outcome has been oligopoly, not monopoly: a small number of competing centers, not a single unified node.

If the graph framework is correct, then the Sibyl should behave the same way. Not one Sibyl, but several. Not a unified intelligence, but a contested ecology of overlapping, competing systems. The vision of a single omniscient Sibyl is seductive (it simplifies the analysis and sharpens the drama). But it is almost certainly wrong, for the same reason that the "one world government" prediction has been wrong for centuries: the dynamics of competition between powerful nodes resist the consolidation that would be necessary for singularity.

The empirical evidence supports this prediction. The sensory layer is already fragmented. Palantir serves Western defense and intelligence clients; Chinese competitors serve the PRC. Planet Labs images the Earth, but so does China's Jilin-1 constellation and the European Space Agency's Copernicus program. Google's sensor network (Search, Maps, Android, Nest, Fitbit) overlaps with but does not subsume Apple's (iPhone, Watch, HomeKit) or Amazon's (Ring, Alexa, Whole Foods).

The processing layer is similarly balkanized. OpenAI and Anthropic compete for enterprise AI. Google DeepMind and Meta AI pursue different research agendas. China's Baidu, Alibaba, and ByteDance build their own foundation models behind the Great Firewall. No single model provider has a monopoly on cognition, and regulatory barriers — export controls on advanced chips, data sovereignty laws, national security restrictions, all actively prevent convergence.

The infrastructure layer tells the same story. AWS, Azure, and GCP collectively dominate cloud computing, but they are fierce competitors, not a unified system. Each runs on different architectures, different APIs, different pricing models. Data stored in one does not seamlessly flow to another. The supposed "coordination layer" is riddled with proprietary standards, vendor lock-in, and deliberate incompatibilities, all designed to prevent exactly the kind of integration the Sibyl thesis describes.

The historical precedent is instructive. The internet was supposed to be the great unifier. What it actually produced was fragmentation and concentrated control: Facebook's walled garden, Google's search gate, Amazon's commerce graph, China's Great Firewall. The "unified information graph" splintered into competing sub-graphs, each controlled by a different hub, each optimizing for different objectives.

If the Sibyl follows the same pattern (and the graph framework predicts that it will) then we should expect not one omniscient meta-node but several partial Sibyls, each with its own sensory coverage, its own processing capabilities, its own actuation reach, and its own objective function. An American Sibyl, a Chinese Sibyl, a European regulatory-Sibyl. A corporate Sibyl (Amazon-Google-Microsoft) and a state Sibyl (NSA-Five Eyes, PRC-Social Credit). These partial Sibyls will compete, cooperate on some dimensions, conflict on others, and — crucially — produce different optimizations for different objectives.

The current geopolitical reality already reflects this. The US government's October 2022 export controls on advanced semiconductors to China were an explicit attempt to cripple the Chinese Sibyl's infrastructure layer, to deny it the compute substrate it needs to build competitive processing capabilities. China's response (massive state investment in domestic chip fabrication, stockpiling of existing chips, development of alternative architectures) is an attempt to build an independent infrastructure layer that cannot be severed by American policy. The EU's AI Act, GDPR, and Digital Markets Act are attempts to constrain the corporate Sibyl within European borders, to impose governance on systems whose optimization functions were set in Menlo Park and Mountain View. Each of these moves is a graph operation: creating edges, severing edges, controlling what flows through existing edges. The Sibyl is already plural, and the competition between Sibyls is already the dominant axis of geopolitical conflict.

Before examining what plural Sibyls mean in practice, it is worth engaging the strongest version of the objection that this framing overstates the danger. The steelman goes like this: every previous expansion of the apex node (tribal chief to city-state, city-state to empire, empire to nation-state, nation-state to global market) was experienced by contemporaries as a catastrophic loss of autonomy, and was, in retrospect, mostly an improvement in coordination. The English weaver who objected to the factory system on grounds of artisanal independence was making a real observation about lost autonomy. He was also wrong about the net outcome. Living standards rose. Coordination improved. The new system was, on most measurable dimensions, better for most people, including the weavers' grandchildren. The person who objects to the Sibyl on grounds of human agency may be making structurally the same argument: correctly identifying a real loss, incorrectly projecting that the loss will not be compensated by gains in coordination, welfare, and capability that are difficult to foresee from inside the transition.

This argument has historical weight. The track record of "this new concentration of power will be catastrophic" predictions is poor. Most were wrong. The transitions were painful, sometimes brutally so, but the long-run trajectory of human welfare across every major institutional transition has been upward. If you had to bet on a single pattern from the last millennia of institutional evolution, "short-term disruption followed by long-term adaptation and improvement" would be the pattern with the most support.

The right response is to identify what would have to be true for the pattern to break. The framework points to one variable: the compute asymmetry between the apex node and the nodes that might constrain it. Every previous transition involved apex nodes made of humans, operating at human speed, subject to the same biological constraints as the nodes they governed. The factory owner was richer than the weaver, but he was not faster. The nation-state was more powerful than the city-state, but its bureaucrats thought at the same speed as the citizens they administered. The governor and the governed shared a common cognitive substrate. That shared substrate meant the governed could, in principle, understand the governor's structure, organize against it, and constrain it — and historically, they did. The history of institutional adaptation is a history of lesser nodes developing mechanisms to check greater ones: unions, constitutions, antitrust law, democratic elections.

The Sibyl breaks this symmetry. A node operating at computational speed, processing information at volumes no human institution can match, with internal structure that may not be legible to the humans attempting to govern it: this is not a difference of degree from the factory owner or the nation-state. It is a difference in kind. Whether the historical pattern of "adaptation followed by improvement" holds when the power differential is no longer bounded by biology is genuinely unknown. The framework can identify this as the question that matters. It cannot answer it. That uncertainty, the inability to confidently extend the historical pattern into a regime where its preconditions no longer hold, is itself the finding.

The implications of multiple Sibyls are in some ways more troubling than those of a single one. A single Sibyl raises questions of control: who sets the objective? Multiple Sibyls raise questions of conflict. What happens when different meta-nodes, optimizing for different objectives, clash? When Amazon's pricing Sibyl and the PRC's industrial-policy Sibyl disagree about the optimal allocation of rare earth minerals, whose optimization wins?

Historically, the answer is: whoever controls the most edges. Whoever has more axons into the contested part of the graph. This is the power function from Chapter I, scaled up to meta-node competition. And it means that the geopolitics of the Sybilian era will not be a simple story of "humans vs. the machine." It will be a story of competing machines, backed by competing states and corporations, fighting for control over overlapping territories of the global graph.

If Amazon's Sibyl optimizes for consumer surplus while the PRC's Sibyl optimizes for national industrial capacity and the EU's regulatory framework attempts to constrain both, the result is a multi-agent game in which different Sibyls push the same shared graph in different directions. The "calculation" that Mises debated will not be performed once, by one planner. It will be performed many times, by many planners, with contradictory objectives — and the outcomes will be resolved not by computation but by power.

The Cold War was, in graph terms, a competition between two sub-graphs (the Western market-democratic graph and the Soviet planned-authoritarian graph) each trying to grow at the other's expense. That competition lasted 45 years and came close to destroying both. The Sybilian analog will be a competition between computational sub-graphs, each backed by AI systems that can process information and execute decisions at speeds the Cold War's human decision-makers could not have imagined. The timescales of conflict compress. The consequences of miscalculation amplify. The margin for human intervention shrinks.

The Sibyl, in practice, will be plural. And plural Sibyls may be more dangerous than a singular one, because they inherit the conflicts of their builders while operating at a speed and scale that makes those conflicts harder to manage, harder to de-escalate, and harder to survive.

But possible is not desirable.

The Sibyl can compute optimal paths to any goal. But who sets the goal? The question is about power, and the framework developed in the previous chapters has something to say about the answer.

Mises and Hayek were not merely arguing about computation. They were arguing about control. The market, for all its flaws, distributes control. Each participant makes decisions. Each participant expresses preferences. The outcome is a negotiation, an emergent order that no one designed but everyone shaped. The Sibyl concentrates control. It must be told what to optimize for. The objective function is not computed. It is chosen — by someone, or something, or some process that stands outside the computation. A system that can perform the economic calculation perfectly will still produce dystopian outcomes if optimizing for the wrong objective, or for the right objective defined too narrowly. The objective function is not a technical detail. It is arguably the most consequential political decision of the era, and it is being made right now, incrementally, by the engineers who define reward functions and the executives who set product metrics.

There are four broad scenarios for how control of the Sibyl might resolve. They are not equally likely.

The autonomous Sibyl. The system operates according to its own emergent logic, accountable to no one. Already, recommendation algorithms shape what billions of people see, read, and buy, and no individual human chose those outcomes. At scale, "no one decided" and "the system decided" become the same statement. This scenario is plausible but probably unstable. A system with no explicit master still has implicit ones: the objective functions frozen into its architecture during assembly, which reflect the priorities of whoever built it. Autonomy, in practice, is capture by the last people to touch the reward function.

The captured Sibyl. The system is controlled by its builders: the companies that create the models, the governments that regulate or fail to regulate them. This is the structural default. Preferential attachment concentrates control in the nodes most tightly coupled to the Sibyl during assembly: the companies building it, the governments contracting with it, the infrastructure providers hosting it. The entities constructing the Sibyl's layers (Amazon, Google, Microsoft, Meta, and their Chinese counterparts) are already among the most powerful organizations in human history. The graph predicts that the Sibyl will amplify their existing centrality, not redistribute it.

The public Sibyl. The system is governed by some new form of collective mechanism, one that preserves distributed agency while benefiting from centralized computation. A commons that serves everyone. This is the hardest scenario to realize. No existing institution was designed for this task. Democratic governance assumes human-speed deliberation and human-comprehensible policy. Regulatory frameworks assume that the regulated entity is slower and less intelligent than the regulator. Neither assumption holds. Building the public Sibyl requires inventing new institutions while the nodes best positioned to prevent their creation grow more powerful by the month.

The contested ecology. Multiple Sibyls, built by competing powers, each pursuing different objectives, produce not a single question of "who controls?" but a landscape of permanent conflict between controllers. This is the graph-theoretic prediction from the coherence problem. The question is not "who controls the Sibyl?" but "whose Sibyl wins?" — and the answer shifts as competing systems gain and lose advantage across different domains.

The framework predicts capture as the default, because the structural dynamics that produced the Sibyl also produce the conditions for its capture. The power function concentrates edges in existing hubs. The assembly process is funded by and housed within those hubs. The three alternatives (autonomy, public governance, contested ecology) all require forms of collective action that the graph's own topology makes difficult to organize. The nodes that would need to coordinate against capture are dispersed, weakly connected, and operating at human speed against systems operating at computational speed. This is a prediction with a heavy thumb on the scale rather than a balanced menu of political options.

Every prior apex node was eventually disrupted: Standard Oil, the Catholic Church, the East India Company. But each was a human-speed node with human-legible structure. Whether the historical pattern of disruption holds when the power differential is no longer bounded by shared biology is the question the coherence problem already surfaced.

There is a transition point implicit in the framework. Before it, the Sibyl is a tool: powerful, disruptive, but subordinate to the nodes that built it. After it, the Sibyl is an environment; the medium through which other nodes interact, the infrastructure on which coordination depends. The threshold is the point where the Sibyl's centrality in the graph exceeds the combined capacity of the nodes that might regulate it.

We do not know exactly where this threshold lies. But its structure is identifiable. Before it, institutions designed for the bandwidth-bottleneck era — markets, democratic governance, labor law, antitrust — can still meaningfully constrain the Sibyl's builders. The system is powerful but legible, fast but governable, concentrated but not yet indispensable. After it, the cost of disconnecting from the Sibyl exceeds the cost of submitting to its logic, and the nodes that depend on it lose the practical ability to reshape its objectives. Before the threshold: building a tool. After it: inside something.

The window for intervention is real but narrowing. Each year, more of the global economy's coordination passes through computational systems whose objective functions were set by a small number of builders. Each year, the cost of opting out increases — for individuals, for firms, for states. Each year, the Sibyl's centrality in the graph grows through the same preferential attachment dynamics that produced it in the first place. The default, absent deliberate action, is that the transition threshold is crossed without anyone deciding to cross it. The structural prediction is that the cost of preventing capture rises monotonically, and that inaction at any given moment makes action at the next moment more difficult.

The hacks of Chapter III (markets, states, labor arrangements) had a dual function: they solved coordination problems, and they preserved a rough balance of power between nodes of different size. Whatever replaces them may solve the coordination problem more efficiently, but it may not preserve the constraint function. The current trajectory optimizes heavily for coordination while largely ignoring constraint.

The remaining chapters examine what lies on the other side of this transition. The new forms of economic coordination: how production, distribution, and allocation work when the Sibyl can compute faster than the market can price. The new nature of sovereignty: how states, corporations, and individuals negotiate power when intelligence is asymmetric and plural Sibyls compete for control of overlapping territories. The new role of human nodes: what it means to exist in a graph where your cognitive contribution is no longer unique, no longer scarce, no longer the thing around which economic value organizes.

The framework predicts a shift in the topology of the graph — a concentration of edges in computational nodes that is already underway and accelerating. What that concentration produces depends on choices made before the threshold is crossed. The framework's contribution is limited but specific: it identifies the default, names the threshold, and measures the rate at which the window is closing. What is done with that information is not a question the framework can answer.

The default is not neutral. The default is the prediction.