The System Gets a Brake One Way or Another

1. The driver who cannot slow down

You are a passenger. The driver is accelerating, and has been for some time. You ask them to slow down.

They say they can't. If they slow down, the car in the next lane will be ahead.

Notice what the driver has not said. They have not said the current speed is safe. They have not disputed that the road is getting narrower or that the tires are getting warm. They have conceded the danger entirely and answered a different question. You asked about speed; they answered about position.

Now consider the driver of the other car, who is having the same conversation with their own passenger and giving the same answer. Both drivers would prefer to slow down. Each says they cannot unless the other does first. Neither believes the other will. So both keep their foot down, each for a reason that is intelligible, each aware that the outcome they are producing together is one neither of them wants.

At some point the speed will stop increasing. That much is certain, because no physical system accelerates indefinitely. What is not certain is how it stops. If the two drivers reach an understanding, the deceleration is uneventful. If they do not, the deceleration still arrives: fuel runs out, a tire fails, a bend arrives faster than reaction time allows, someone loses control, a police car appears, or the two cars meet. The correction is no longer a decision anyone makes. It is something that happens to them.

That asymmetry is the subject of this essay. The candidate thesis, to be tested rather than assumed:

An unsustainable competitive trajectory does not require its participants to agree to stop. It requires only that something fail. What coordination decides is whether the correction is applied by the participants or imposed by the system's limits.

The car is a thought experiment, and thought experiments are cheap. What makes this one worth pursuing is that the driver's answer, "I can't slow down or they'll be ahead," is not a rationalization. Under a large class of payoff structures it is correct. That is the disturbing part, and it is where the analysis has to begin.

2. The accelerator is not the constraint

"Just slow down" fails as advice because it treats speed as the variable the driver controls. The driver's answer reveals that, from their point of view, the free variable is relative position, and speed is merely the instrument they use to hold it.

This is a distinction between absolute and relative outcomes, and it changes everything about what counts as a reasonable choice. In absolute terms, 130 mph is worse than 60 mph for both drivers: more fuel, more risk, less margin. In relative terms, 130 mph while the other car does 130 is the same position as 60 while the other does 60. If what the driver is being evaluated on is position, the two situations are equivalent, and the only bad outcomes are the two in which they slow down and the other does not.

Economists have a vocabulary for this. Hirsch (1976) described positional goods, whose value to the holder depends on how much others have; Frank (2011) extended the idea into what he calls positional arms races, in which competitors escalate expenditure on something that confers advantage only relative to one another, so that when everyone escalates, nobody gains and everyone pays. Tournament theory (Lazear and Rosen, 1981) formalizes a related structure: when rewards are attached to rank rather than to absolute output, the rational level of effort is set by what competitors do, not by the intrinsic value of the work. Akerlof (1976) gave the pattern its most memorable name, the rat race, in a model where workers signal their quality by working harder than is efficient because the reward depends on comparison.

Three distinctions now do the analytical work for the rest of the essay.

Collective preference versus unilateral incentive. Both drivers prefer mutual restraint to mutual escalation. Neither has an incentive to restrain unilaterally. The collective preference is real and does not by itself produce action, because no individual actor is positioned to choose the collective outcome.

Voluntary restraint versus coordinated restraint. Voluntary restraint is one driver slowing and hoping. Coordinated restraint is both drivers slowing under some arrangement that makes each confident the other will. The first is exposed to exploitation; the second is not, if the arrangement holds. The word coordination will be used in this narrower sense throughout: not goodwill, but an arrangement that changes what each party can expect of the other.

Local rationality versus system sustainability. Each driver's decision, taken alone, survives scrutiny. The trajectory they jointly produce does not. Nothing in the local reasoning refers to the trajectory, which is why the trajectory is not corrected by local reasoning.

Competition, in this configuration, performs a specific conversion. It takes an outcome that everybody would prefer to avoid and turns it into an outcome that nobody believes they can individually prevent. The mechanism of conversion is the relative-position payoff. Remove that, and the driver's objection dissolves. Keep it, and no amount of shared understanding about the danger will move the accelerator, because the danger was never in dispute.

3. A rational race toward an unwanted outcome

The two-car scenario invites the label "Prisoner's Dilemma." Sometimes that is right. Often it is not, and the difference matters because the remedies differ.

3.1 Which game are the drivers playing?

Strip the situation to two players, each choosing Slow or Fast. Four outcomes: both slow, both fast, or one of each. The game is determined by how each player ranks these outcomes, and the driver's words tell us most of what we need.

The driver says: I would slow if they would. That is a statement that (Slow, Slow) is preferred to (Fast, Fast). The driver also says: I cannot slow unless they do. That is a statement that (Slow, Fast), being left behind, is the worst outcome. What the driver has not told us is how they rank (Fast, Slow): being ahead while the other restrains. Everything turns on that cell.

If the driver would prefer mutual restraint to being ahead, so that pulling ahead of a restrained competitor is not actually the goal, the game is a Stag Hunt (Skyrms, 2004), also called an assurance game. It has two equilibria. Both slow is one; both fast is the other. Mutual restraint is the better equilibrium for both, and the only thing preventing it is each player's fear that the other will not hold to it. The barrier is confidence, not incentive. Give each driver a credible reason to believe the other will slow, and both slow, because both wanted to. This is the structure in which the phrase "coordination problem" is exactly right: there is an outcome both prefer, and the problem is arriving at it together.

If the driver would prefer being ahead to mutual restraint, the game is a Prisoner's Dilemma. Fast is now the dominant strategy: better if the other slows (you pull ahead), better if the other stays fast (you don't fall behind). Both fast is the unique equilibrium, and it is worse for both than mutual restraint. Here confidence is not enough. Even a driver who is certain the other will slow has reason to stay fast. The barrier is incentive, and the remedy has to change payoffs, not beliefs.

The distinction is not academic. An assurance problem yields to communication, verification, and reputation. In a one-shot dilemma, confidence alone is insufficient: even certainty that the other party will restrain leaves unilateral escalation individually advantageous, and only enforcement, compensation, or a restructuring of what counts as winning changes that. Repeated interaction, however, can change the calculation by attaching future costs to present defection, which is how reputation and reciprocity come to sustain cooperation in dilemmas that recur (Section 8.1). Most real competitive races are mixed, with some actors in one structure and some in the other, and with the same actor moving between structures as circumstances change. Diagnosing which barrier is binding, fear or greed, is the first practical question.

3.2 Chicken, and why the game is not stationary

There is a third structure, and it enters the story late. In Chicken (Hawk-Dove in the biological literature), each player prefers to hold course if the other yields, but mutual holding is the worst outcome for both, worse than yielding alone. Two cars approaching each other on a narrow bridge: whoever swerves loses face, but if neither swerves, both lose the car.

Early in the race, the two drivers are not playing Chicken. Mutual escalation at 70 mph is uncomfortable, not catastrophic; being left behind still ranks lower. But as speed rises and margin is consumed, the (Fast, Fast) cell gets worse, and the payoff ordering itself can change. Chicken requires a complete ordinal structure, not merely a bad mutual-escalation cell: unilateral escalation against a yielding rival is best, mutual restraint next, yielding to an escalating rival third, and mutual escalation worst. If mutual continuation deteriorates far enough that yielding becomes preferable to it, while pulling ahead of a yielding rival remains the most attractive outcome, the interaction has moved into a Chicken-like region, with Chicken's characteristic pathology: each player now has an incentive to appear unable to yield, so that the other yields first. The transition is not automatic. Whether it occurs, and at what point, depends on how the underlying payoffs change with escalation, which is an empirical property of the particular race.

This is the first place the analysis produces something the metaphor did not contain. The payoff matrix need not be fixed. Escalation can alter it. A race that begins as a coordination problem, solvable by assurance, can become a dilemma as lock-in, switching costs, irreversible commitments, continuation rewards, and organizational dependence on the escalated state raise the marginal value of continuing, and can then enter a Chicken-like region as the margin runs out. (Sunk costs as such do not do this; costs already incurred should not affect the marginal choice. What changes the game is what continuation now buys and what stopping now forfeits.) Where this happens, remedies that would have worked at one stage are ineffective at the next. This is a conditional dynamic model proposed by this essay, not a theorem that races proceed through these stages, and it is marked here as synthesis.

3.3 Arms races, security dilemmas, and collective action

The two-player games generalize in two directions.

The security dilemma (Herz, 1950; Jervis, 1978) describes states that arm defensively and, in doing so, make neighbors less secure, prompting them to arm, which makes the first state less secure. Jervis's analysis is precise about when this produces a race: it depends on whether offensive or defensive capabilities dominate and whether they can be distinguished. When they cannot, restraint is unverifiable; in this essay's terms, though not Jervis's, that is the condition under which an assurance structure collapses toward a dilemma. Richardson's (1960) arms-race equations model the dynamics directly, with each side's spending driven partly by the other's. The essential feature for this essay is that both sides may be sincere in wanting stability and still produce instability, because each side's defensive acquisition is indistinguishable, to the other, from preparation for attack.

Collective-action theory (Olson, 1965) generalizes to many players, where the problem is that the benefit of restraint is diffuse and the cost concentrated on whoever restrains. Hardin's (1968) tragedy of the commons is the special case where the shared resource is depletable: each herder gains the full benefit of an additional animal and bears only a fraction of the cost to the pasture. Gordon (1954) had already shown the same logic for open-access fisheries. The commons framing fits the two-car problem only partially. The cars are not depleting a shared resource in the ordinary sense; what they are depleting is safety margin, which behaves like a commons in one respect (both draw on it; neither replenishes it) but not in others (it is not a stock either can privately capture). The commons literature is invoked here for its structure, not as a claim that every escalation race is a resource-depletion problem. Ostrom's (1990) study of commons that did not collapse is more relevant to Section 8 than Hardin's model of ones that did.

The point of surveying these models is not to choose one. It is that the two-car scenario survives formalization. It maps onto well-studied structures, the mapping is not forced, and the structures agree on the central claim: participants can want the restrained outcome, be correct that unilateral restraint is costly, and jointly produce the escalated outcome without any of them wanting it. No villain is required. That is what makes the problem hard, and what makes "competition is bad" or "people should cooperate" inadequate as diagnoses.

4. Why danger is not enough

If both drivers can see the danger, and both prefer to slow, why does the visible danger not resolve the problem? The game-theoretic answer is that danger changes the absolute payoffs, not the relative ones, and the drivers are optimizing on the relative ones. But that answer is too clean. Real systems contain specific mechanisms that keep escalation running even when its participants privately dislike it, and they are worth naming because each one is a target for intervention.

First-mover disadvantage and fear of exploitation. Whoever restrains first bears the cost immediately and alone. If the other party does not follow, the first mover has paid for nothing. This is the Stag Hunt's fear equilibrium operating in time: it is not that restraint is undesirable but that going first is.

Unverifiable restraint. The driver cannot see the other car's speedometer, only its position. If the other driver eases off the accelerator, it looks identical to a car that is about to be overtaken. Jervis's point about indistinguishable offensive and defensive capability is the general form: when the other party's restraint cannot be distinguished from their weakness or their feint, restraint cannot be reciprocated with confidence, and assurance is unavailable even to parties who want it.

Short horizons and organizational incentives. The individuals making acceleration decisions are rarely evaluated on the long-run trajectory. They are evaluated on this quarter's position, this cycle's performance, this year's ranking. Relative-performance metrics inside organizations reproduce the two-car structure at every level: a manager who unilaterally slows their team is not rewarded for the system's long-run stability, which is diffuse and unattributable, but is penalized for the team's relative decline, which is visible and attributable. The system can contain many people who dislike its direction and cannot, from where they sit, afford to be the one who deviates.

Lock-in and committed investments. Capital, careers, operating models, and identity get committed to the current speed. The organization that has built its structure around escalation can face real switching costs and forfeited continuation benefits when it slows, and this dependence can grow as escalation continues.

Competitive selection. This mechanism operates without anyone deciding anything. If actors that voluntarily restrain themselves are systematically outcompeted, then over time the population of actors is composed of those who did not. Smaldino and McElreath (2016) model this for scientific publishing: methods that produce more publishable results propagate even when they produce worse science, because labs are selected on output, not accuracy. The individuals may all prefer rigor. The population trends toward whatever wins. The driver who slows down is not persuaded to speed up again; they are simply no longer in the race, and the race continues without them.

Shifting baselines. Pauly (1995) coined "shifting baseline syndrome" for the way each generation of fisheries scientists took the depleted stocks of their own early careers as the normal reference point, so that the long-run decline was invisible from any single vantage. Vaughan's (1996) analysis of the Challenger launch decision describes the organizational version, which she called the normalization of deviance: each anomaly that did not cause disaster became part of the accepted operating envelope, so that the envelope expanded step by step until it contained the conditions that did. Applied to the car: 90 mph felt reckless at the start of the journey and feels ordinary an hour later, not because the physics changed but because the reference point did. The drivers are no longer comparing their speed to a safe speed. They are comparing it to yesterday's.

These mechanisms are distinct, and several can sustain escalation even when the others are absent. Together they explain the phenomenon that makes competitive escalation so resistant to the obvious objection: a system can be populated by people who see the danger, dislike the trajectory, and continue to reproduce it because the choices available to them still reward continuation more reliably than unilateral restraint.

The alternative reading, that the participants are simply foolish, greedy, or in denial, is often incomplete. The more uncomfortable possibility is that they are responding intelligibly to the structure they are in, and that changing behavior therefore requires changing that structure rather than merely condemning the people inside it.

5. Speed consumes margin

"Unsustainable" is usually deployed as a moral adjective. It has an operational meaning that is more useful.

A trajectory is unsustainable with respect to some resource when continuing it requires drawing down that resource faster than the resource is replenished, so that continuation has a finite horizon whether or not anyone chooses to stop. The resource can be almost anything: fuel, liquidity, ecological carrying capacity, attention, labor, institutional legitimacy, public tolerance, error tolerance, supply-chain slack, mechanical tolerance. What the definition requires is only that the resource be finite in the relevant sense and that the trajectory consume it.

Under this definition, unsustainable does not mean about to collapse. A system can operate beyond its prudent range for a long time. The definition says only that the horizon exists, not where it is. That distinction, between "cannot continue indefinitely" and "will stop soon," is the one the essay's argument most depends on, and it is where careless versions of the argument go wrong.

5.1 Buffers

The resource being consumed in competitive escalation is often not the obvious one. It is margin: the distance between current operating conditions and the conditions at which something fails.

Rasmussen (1997) gave this a diagram that has become foundational in safety science. He described a system's operating point as moving within a space bounded by three limits: economic failure, unacceptable workload, and unacceptable risk. Two gradients push the operating point around that space. Management pressure toward efficiency pushes it away from the economic boundary. The individual tendency toward least effort pushes it away from the workload boundary. Nothing pushes it away from the safety boundary with equivalent force, because the safety boundary is invisible until it is crossed. The operating point therefore drifts, gradually and without any decision to accept more risk, toward the edge of the safe envelope. Rasmussen's point was that this drift is a systematic product of ordinary pressures, not of negligence. Dekker (2011) later called the general phenomenon drift into failure.

The two-car race is a Rasmussen diagram with the competitor supplying the gradient. Each increment of speed is small, locally justified, and moves the operating point closer to a boundary nobody can see. The margin consumed in each step is unremarkable. The cumulative consumption is the story.

Woods (2015), writing from the resilience-engineering tradition, distinguishes robustness (the capacity to handle anticipated disturbances) from what he calls graceful extensibility: the capacity to stretch performance when a surprise exceeds the envelope. Systems operating near their boundaries have used up their capacity to stretch. They can still perform, often very well, but they have no reserve for the disturbance they did not anticipate. Every buffer converted into speed is a buffer no longer available to absorb a shock.

5.2 The system adapts to its own speed

There is a second effect, subtler than depletion. As escalation continues, the surrounding system adapts to the escalated state. Suppliers size for the higher demand. Downstream processes assume the higher throughput. Capital structures assume the higher valuation. Staffing assumes the longer hours. The escalated state becomes the state on which other things depend.

This means that the cost of slowing down rises over time for a reason independent of competition. Early, slowing down is a reversion. Later, it is a disruption of everything that has been built on the assumption of speed. The buffer being consumed is not only the safety margin of the cars but the adaptability of the road.

5.3 Not every race runs out of road

The argument must not overreach. Many competitive escalations stabilize without either coordination or failure, and the mechanisms are worth naming because they are the null hypothesis against which the essay's thesis has to be tested.

Some races hit diminishing returns: past a point, additional speed buys no additional position, and escalation stops on its own. Some hit technological limits that arrive gently, as an S-curve rather than a wall. Some are ended by substitution, when the contested resource stops mattering. Some by market exit, when enough competitors leave that the remaining ones face weaker pressure. Some by changing preferences among the people the competitors are competing for. And in some, the "speed" being escalated is genuinely valuable in absolute terms, so that the race, however uncomfortable for the racers, is not consuming margin at all but producing something.

The thesis therefore applies to a bounded class: escalations in which the contested variable is primarily positional, in which escalation consumes a finite margin or sustaining resource, and in which no stabilizing mechanism intervenes before the margin is exhausted. That class is smaller than "all competition" and larger than is comfortable.

6. When slowing down stops being optional

The candidate thesis said that an unsustainable system needs only something to fail, not agreement to stop. Section 5 sharpened what "unsustainable" means. This section examines what "fail" means, because the word carries cinematic connotations the argument does not need and should not rely on.

6.1 Deliberate and forced correction

A deliberate correction is one applied by the participants under some coordination arrangement: a treaty, a rule, a mutual understanding, a shared circuit breaker. It is chosen, timed, and typically gradual. The drivers agree to slow, and slow.

A forced correction is one imposed on the participants by a constraint they did not choose to invoke. The category is broad, and most of it is not a crash. It includes: one participant failing and leaving the race (bankruptcy, exit, breakdown); a resource being exhausted (fuel, capital, credit, a fish stock); an infrastructure bottleneck asserting itself (the road narrows); an external authority intervening (a regulator, a court, a police car); a social or political reaction that changes the payoffs from outside (backlash, loss of legitimacy); a technical failure that removes a capability (the engine); a cascading failure in something the race depended on; or an abrupt repricing, in which the shared belief that sustained the escalation collapses and the escalated state is revealed as unsupported.

What these share is not their severity but their involuntariness. Some forced corrections are mild. One car runs low on fuel and pulls over; the other, no longer pressured, slows. Nothing has crashed. The point is not that forced correction is catastrophe. It is that forced correction is not a decision, and therefore its timing, form, and severity are not under the participants' control.

This distinction is not, so far as the literature reviewed here shows, standard terminology under these names. Adjacent concepts exist: resilience engineering distinguishes graceful degradation from brittle failure; economics distinguishes soft and hard landings; ecology distinguishes gradual and catastrophic regime shifts. The deliberate/forced distinction is proposed here as a general framing that these domain-specific pairs instantiate, and is marked as synthesis.

6.2 Thresholds, cascades, and repricing

Three bodies of work describe how forced corrections arrive, and each is invoked here only for the specific mechanism it supports.

Critical transitions in ecosystems. Scheffer et al. (2001) showed that some ecosystems respond to gradually changing conditions not gradually but with abrupt shifts between alternative stable states, and that these shifts exhibit hysteresis: once the system has tipped, restoring the driving conditions to their pre-tip value does not restore the original state. A shallow lake that has shifted from clear to turbid does not clear again when nutrient loading is reduced to the level at which it turned; it must be reduced much further. The relevance is narrow and specific. It establishes that a system can absorb escalating pressure with little visible change and then respond to a small increment with a large, discontinuous, and partly irreversible shift. Whether a given competitive system has this property is an empirical question, not something the analogy settles. But the possibility is what the two drivers cannot rule out, and it is what makes "we'll slow down when it gets dangerous" unreliable as a plan.

Cascading failure in coupled systems. The 2003 North American blackout, documented in the U.S.–Canada Power System Outage Task Force report (2004), began with a small number of line faults in Ohio and propagated across the interconnected grid within minutes, disconnecting roughly 50 million customers. Perrow (1984) had argued that systems combining interactive complexity, in which combinations of failures are difficult to foresee or understand, with tight coupling, which leaves little slack or time for intervention, produce accidents that no single component failure explains. Buldyrev et al. (2010) showed formally that networks that depend on one another can fail catastrophically at a failure level that either network alone would have survived. The mechanism for this essay: when a race has caused many things to depend on its continuation (Section 5.2), a failure in one participant or one dependency does not remain local.

Financial instability and leverage. Minsky's financial instability hypothesis (1992) holds that periods of stability breed the conditions for instability, because sustained good outcomes lead participants to accept financing structures that depend on those outcomes continuing. Leverage rises during the calm; the system becomes dependent on refinancing; a modest disturbance that would have been absorbed at lower leverage instead forces asset sales, which lower prices, which force further sales. Brunnermeier (2009) describes the loss and margin spirals of 2007–08 in exactly these terms. Kindleberger (1978) traced the same displacement-euphoria-distress-panic sequence across three centuries of episodes. The relevance is the dependence of stability on continuation: the escalated state is stable only while it escalates, and the correction, when it comes, is abrupt because the position that looked sustainable at rising prices is insolvent at falling ones.

The three literatures do not describe one mechanism. Hysteresis, cascade, and repricing are different. What they share is the property the argument needs: each describes a way in which an escalated system can stop without deceleration having been anyone's decision, and each shows that the stopping can be discontinuous rather than smooth.

6.3 What the thesis does and does not claim

The claim to be tested was: unsustainability guarantees only that the trajectory cannot continue indefinitely, not that restraint will arrive before the limit. That claim survives. Nothing in the mechanisms of Sections 3 and 4 refers to the limit, so nothing in them causes participants to stop before it. The limit is reached by default unless something intervenes.

Two things the thesis does not claim. It does not claim that failure is inevitable; Section 5.3 listed the stabilizers, and coordination (Section 8) is a further one. And it does not claim to predict what fails or when. "Something eventually prevents continuation" is a structural statement. "The tire will fail at 142 mph in eleven minutes" is a prediction, and the essay makes none. Conflating the two is how this argument becomes apocalypse rhetoric, and the conflation is unnecessary: the structural statement is alarming enough on its own, because it says the correction is coming and the participants have not chosen its form.

7. The later the brake, the harsher the correction

Why should a delayed correction be worse than an early one, rather than merely later? The car supplies the first answer literally, and the systems literature supplies the rest.

Kinetic energy scales with the square of speed. A car at 120 mph carries four times the kinetic energy of the same car at 60 mph, not twice. Stopping distance under constant braking force scales the same way. Coordinated braking at 60 is a mundane manoeuvre. At 120 it requires four times the distance and dissipates four times the energy during braking. Under repeated or sustained heavy braking, that heat can contribute to brake fade as components overheat and braking effectiveness falls. At 120 with faded brakes, the problem has changed a third time. It is no longer a question of whether the drivers agree to slow but whether the equipment that would implement the agreement still works.

The systems version has the same shape and four components.

Less redundancy remains. The buffers consumed during escalation (Section 5.1) were the system's capacity to absorb a disturbance without failing. A correction is a disturbance. A system that has converted its margins into speed meets the correction with nothing in reserve.

Participants are more committed. Sunk investments, leverage, and adaptation to the escalated state (Section 5.2) mean that each participant has more at stake in continuation and less capacity to exit gracefully. Minsky's leverage mechanism is the financial instance: the position that would have been unwound easily at low leverage forces a fire sale at high leverage.

Dependencies have adapted. Other systems now assume the escalated state. A correction in the race becomes a shock to everything built on it, which is the cascade mechanism of Section 6.2.

Small disturbances can have large consequences. In systems with thresholds, moving close to a boundary means that a disturbance that would have been absorbed earlier can now cross it. This is one operational meaning of lost margin: less room remains between ordinary variation and a state change. Scheffer's hysteresis adds that, in systems with that property, crossing the boundary may not be reversible by a proportionate reduction in pressure.

Resilience engineering has a term for the desirable alternative. A system degrades gracefully when it loses function progressively and visibly as stress increases, giving operators time and information to respond. It fails brittlely when it maintains apparent performance up to a boundary and then loses function abruptly. Escalation that consumes margin converts the first kind of system into the second, not by changing its components but by moving its operating point to a region where its behavior is no longer graceful.

The practical conclusion is uncomfortable in a specific way. The moment at which coordination is often cheapest is the moment at which the danger is least visible, the margins are largest, and the participants have the least reason to pay for restraint. By the time the danger is undeniable, coordination can be harder because the game may have drifted toward a Chicken-like region (Section 3.2), the participants may be more committed, and fewer buffers remain to absorb the correction. The information and the incentive can arrive in the wrong order.

8. Changing the game

The preceding sections could be read as fatalist. They should not be. Coordination problems are solved regularly, and the record of how they are solved is instructive, because successful solutions share a property that unsuccessful proposals lack.

The property is this: a working solution changes the payoff structure or the confidence structure. It does not ask participants to behave better within the existing one. Applied as a test, this discards most proposals immediately. If the entire content of a proposal is "everyone should voluntarily be more responsible," the structural problem has not been engaged. The driver has already told us that responsibility is not the variable in play.

What follows sorts mechanisms by what they change.

8.1 Mechanisms that change confidence

These address the assurance problem, where participants would restrain if they trusted others to.

Communication and focal points. Schelling (1960) showed that people can coordinate on salient alternatives even without communication, because a mutually recognizable focal point gives expectations somewhere to converge. Communication can further reduce strategic uncertainty when it is available. A speed limit sign is a focal point: it gives both drivers a number to converge on that neither had to propose.

Verification. Restraint that can be observed can be reciprocated. Arms-control agreements from the 1970s onward were built around verification provisions precisely because unverifiable restraint is indistinguishable from weakness (Section 4). The 1987 Intermediate-Range Nuclear Forces Treaty included on-site inspection for that reason. In the two-car case, a shared speedometer visible to both drivers changes the game by making each driver's restraint legible to the other.

Repeated interaction and reputation. Axelrod (1984), building on Axelrod and Hamilton (1981), showed that in repeated games with an indefinite horizon, conditional cooperation strategies can be stable, because defection today is punished by non-cooperation tomorrow. This works only when the shadow of the future is long enough and identities are persistent. It fails when the horizon is short, when actors can exit and re-enter under new names, or when the payoff from a single defection exceeds the discounted value of future cooperation.

Transparency. Making the state of the system visible to participants and outsiders shifts baselines back toward reality (Section 4) and makes drift harder to normalize.

8.2 Mechanisms that change payoffs

These address the dilemma, where participants would escalate even if others restrained.

Enforceable limits. A rule with a penalty changes the (Fast, Slow) cell directly: pulling ahead of a restrained competitor now costs something. Speed limits, capital requirements, emissions caps, and salary caps all work this way. The Washington Naval Treaty of 1922 imposed tonnage ratios on the major navies' capital ships; Japan gave notice of termination in December 1934, and the treaty ceased to have effect after 31 December 1936, which is itself instructive about enforceability: a limit holds only while the parties continue to accept it.

Quotas and property rights. Where the escalation is over a depletable common resource, converting open access into allocated, tradable shares changes each participant's incentive from "take it before others do" to "preserve the value of my share." Individual transferable quotas in fisheries are the standard example. Ostrom (1990) documented communities that sustained common resources for generations through locally designed rules with monitoring and graduated sanctions, without either privatization or central control. Her findings are the empirical rebuttal to the view that commons dilemmas are unsolvable, and also a caution: the successful cases had specific structural features that not every commons possesses.

Compensation for restraint. Restraint can be paid for. The 1990 London Amendment to the Montreal Protocol on ozone-depleting substances established a financial mechanism, including the Multilateral Fund, to meet the agreed incremental costs of compliance for developing countries. The treaty describes this as cost-sharing; in game-theoretic terms it is a side payment. Paying participants for whom restraint is most costly is often cheaper than tolerating their continued escalation, and it can convert a dilemma into an assurance game by removing the cost of the (Slow, Fast) cell.

Changing the metric. If actors are evaluated on relative position, the race is structural. If they are evaluated on absolute outcomes, or on a different variable altogether, the relative-position incentive disappears. This is the most fundamental intervention and the hardest, because whoever sets the metric is usually not the participants, and often benefits from the race.

8.3 Mechanisms that impose constraints

These do not change what participants want or believe. They make certain outcomes unavailable.

Circuit breakers and rate limits. After the October 1987 crash, the Presidential Task Force on Market Mechanisms (the Brady Commission, 1988) recommended coordinated trading halts, and U.S. equity markets subsequently adopted market-wide circuit breakers that halt trading when an index falls by set percentages. The design principle is precise: a precommitted, automatic pause that no participant can override and that therefore does not require any participant to be the first to stop. It converts a forced correction into a deliberate one by scheduling the deliberation in advance.

Automatic stabilizers. Mechanisms that dampen escalation without anyone deciding to, such as margin requirements that rise with volatility, or progressive costs that increase with speed.

Precommitted thresholds. A rule adopted in advance, when margins are wide and the game is still an assurance game, that binds behavior later, when margins are narrow and the game may have drifted toward Chicken. The value is temporal: it moves the decision to the moment when it is cheap (Section 7) and enforces it at the moment when it would otherwise be impossible.

The unifying observation is that none of these asks the driver to be a better person. Each changes what the driver can see, what the driver can expect, what the driver is rewarded for, or what the driver is able to do. The driver's objection, "I can't slow down or they'll be ahead," is answered not by disputing it but by making it false.

9. The races we are already running

The framework is only useful if real systems survive mapping onto it without distortion. The discipline adopted here is to state, for each case, what corresponds to each element of the model. Cases that cannot fill every row have been left out. The rows are:

1. the cars (participants);
2. being ahead (the relative-position variable);
3. acceleration (the escalating behavior);
4. the cost of unilateral restraint;
5. the buffer consumed;
6. the limiting mechanism if uncoordinated;
7. the available coordination mechanism.

9.1 Naval and nuclear arms races

Cars: states. Ahead: military superiority relative to a specific rival. Acceleration: procurement and deployment. Cost of unilateral restraint: perceived vulnerability, and domestic political penalty for appearing weak. Buffer consumed: fiscal capacity, and, in the nuclear case, decision time and crisis stability. Limiting mechanism: fiscal exhaustion, war, or one party's internal failure. Coordination: treaties with verification (Washington 1922; INF 1987), hotlines, focal-point doctrines.

This is the canonical case around which much of the security-dilemma literature developed. Its distinctive feature is that the game can be mixed: some parties may face an assurance problem, others a dilemma, and the same party's incentives may change with circumstances. Verification can make restraint legible enough to support assurance where the underlying incentives permit it.

9.2 Financial leverage

Cars: financial institutions. Ahead: relative returns. Acceleration: rising leverage and maturity mismatch. Cost of unilateral restraint: lower returns, loss of clients and capital to competitors who continue. Buffer consumed: capital and liquidity cushions; collectively, the system's tolerance for a shock. Limiting mechanism: repricing, funding runs, forced deleveraging, institutional failure. Coordination: capital and liquidity requirements, margin rules, circuit breakers, lender-of-last-resort backstops.

The best-known articulation of the cost of unilateral restraint came from Citigroup's chief executive in July 2007, who told the Financial Times that as long as the music was playing, one had to get up and dance. The statement was widely treated as recklessness. In the same interview he said that liquidity would eventually reverse; he was explaining why the firm was participating anyway. In the framework here it is the driver's answer: an unusually explicit description of the competitive cost of exiting while peers continued. Minsky's hypothesis predicts both the escalation and the abruptness of its end, and Brunnermeier's account of 2007–08 documents the cascade mechanics.

9.3 Fisheries

Cars: fishing operators. Ahead: share of the catch. Acceleration: fleet capacity, effort, technology. Cost of unilateral restraint: the fish one operator leaves are caught by another. Buffer consumed: the stock's reproductive capacity. Limiting mechanism: stock collapse, then moratorium. Coordination: quotas, transferable rights, seasonal closures, monitoring.

The Northwest Atlantic cod fishery is the standard cautionary case; Canada imposed a moratorium in 1992 after the stock collapsed. Gordon (1954) predicted the mechanism theoretically decades earlier. Pauly's shifting baselines were observed in exactly this domain. What distinguishes the fisheries case is that the buffer being consumed is also the prize, so the limiting mechanism is not incidental to the race but internal to it, and the forced correction, when it came, eliminated commercial access to the prize the fleet had been competing for.

9.4 Working hours

Cars: employees, or firms competing on responsiveness. Ahead: relative visible effort or output. Acceleration: longer hours. Cost of unilateral restraint: being perceived as less committed; losing promotion or clients. Buffer consumed: health, attention, error tolerance, retention. Limiting mechanism: burnout, attrition, error rates, or regulatory limits. Coordination: hours regulation, norms enforced by the employer, changing the metric from hours to outcomes.

Landers, Rebitzer, and Taylor (1996) studied this in law firms and found evidence consistent with Akerlof's rat-race model: billable hours functioned as a signal in promotion decisions, and associates preferred lower hours when the reduction applied to their peers as well, which is the positional preference the model predicts. This case is included because it demonstrates the framework operating entirely inside a single organization, at human scale, with no exotic dynamics.

9.5 Research publication

Cars: researchers and labs. Ahead: publication count and prestige. Acceleration: faster, less rigorous, more numerous publications. Cost of unilateral restraint: fewer publications, weaker position in hiring and funding. Buffer consumed: reliability of the literature; public and institutional trust in it. Limiting mechanism: replication failure, loss of credibility, funder intervention. Coordination: changed evaluation criteria, preregistration norms, replication funding.

Smaldino and McElreath's (2016) model is the relevant formal result, and it illustrates the selection mechanism of Section 4: no researcher need prefer bad methods for bad methods to spread, if bad methods are what the incentive structure rewards.

9.6 Doping and sports technology

Cars: athletes or teams. Ahead: rank. Acceleration: pharmacological or technological enhancement. Cost of unilateral restraint: losing to enhanced competitors. Buffer consumed: athlete health; the legitimacy of the competition itself. Limiting mechanism: health crisis, scandal, loss of audience. Coordination: enforced rules with testing; equipment standards.

This case earns inclusion for one reason: it is the clearest example of a buffer being legitimacy. The competition depends on spectators believing the ranking reflects something real. Escalation consumes that belief. Doping and equipment rules are enforced-limit mechanisms (Section 8.2), and they demonstrate that the coordination problem is solvable when a governing body exists with authority over all participants.

9.7 AI development

This case is included as one instance of the framework and is confined to structure. It makes no claims about specific firms, capabilities, timelines, or outcomes, and the mapping below should be read as a hypothesis about payoff structure, not as an assertion about the current state of the field.

Cars: developers of frontier systems. Ahead: capability or market position relative to rivals. Acceleration: faster development and deployment. Cost of unilateral restraint: ceding position to a competitor who does not restrain, with the further concern that the competitor may be less careful. Buffer consumed: the time available to understand and evaluate systems before they are deployed, and the margin for error in that evaluation. Limiting mechanism, if uncoordinated: an incident, a regulatory or political reaction, a failure of a participant, or a repricing of expectations. Coordination: shared evaluation standards, verifiable commitments, regulatory thresholds, precommitted conditions under which participants agree to pause.

The structural question, stripped of prediction, is the one this essay has been asking throughout: if multiple actors believe that restraint imposes a strategic cost unless competitors also restrain, what would make restraint compatible with competitive survival? The framework's answer is Section 8's. Restraint becomes more compatible with survival when it is verifiable (so that assurance is possible), when it is enforced or compensated (so that a dilemma's payoffs change), or when it is precommitted (so that the decision can be made before escalation narrows the available options or pushes the interaction toward a Chicken-like region). Whether any of these is achievable in this domain is an open question. The framework does not settle it. It says only what a solution would have to do.

9.8 What the cases share, and what they do not

Across the seven cases, the participants differ, the buffers differ, and the limiting mechanisms differ. The framework admits them for the same two reasons: each contains a material cost to unilateral restraint, and each has a successful or proposed coordination mechanism that works by changing confidence, payoffs, or constraints rather than character.

The cases also suggest an institutional difference. Where a governing body has authority over all participants (sport, domestic financial regulation), common constraints can be imposed directly, though enforcement remains imperfect. Where no such authority exists (interstate competition, many global commons), coordination depends more heavily on verification, reciprocity, and continued mutual interest. The difference matters because one setting allows the payoff structure to be changed by rule, while the other generally requires agreement among the players themselves.

10. Conclusion: the brake arrives either way

Return to the car.

The alarming feature of the scene is not the speed. Speed is a number, and numbers can be changed. The alarming feature is that neither driver believes speed is a variable they are free to choose. Each experiences the accelerator as fixed by the other car. What they experience as their choice is position, and position is a variable that can be held constant at any speed, including speeds that cannot be sustained. The two drivers are, in the only sense that matters to them, standing still. The road is what is moving.

This is what the game-theoretic analysis establishes and what the systems analysis then follows to its end. Individually intelligible choices, made under a relative-position payoff, can produce a trajectory that neither participant wants and neither believes they can individually alter. The trajectory consumes margin. Margin, once consumed, can change the system: less redundancy, more commitment, more dependency, and, in thresholded systems, less room for disturbance. The strategic interaction can change with it. If the incentive structure contains no mechanism that ties present choices to the approaching limit, the limit can be approached by default.

The limit is real. Tires, fuel, road geometry, mechanical tolerance, reaction time. In systems language: capital, capacity, legitimacy, attention, error tolerance, the reproductive rate of a fish stock, the time it takes to understand what has been built. These are absolute constraints, and they are indifferent to relative position. They do not care which car is ahead. When the participants have declined to choose a speed, the constraint chooses one for them, and the participants discover that the variable they thought was fixed was the only one that was ever really theirs.

Coordination, in the precise sense this essay has used, is the set of arrangements that return that variable to the participants' control before the constraint takes it from them. It is not goodwill. It is verification, enforcement, compensation, precommitment, and the redesign of what counts as winning. Its function is to make the driver's true statement, "I cannot slow down unless they do," into a statement that is no longer true.

The two sentences the essay set out to test can now be stated in the form the analysis supports.

Before the limit, slowing down is a coordination problem. It is difficult and not automatic, but there are well-understood mechanisms that have solved related coordination problems before.

After the limit, slowing down is not a problem of any kind. It is not a decision. It is something that has already happened, on terms nobody chose.

Provenance

Abstract

A passenger asks a driver to slow down; the driver replies that they cannot, because the adjacent car would then be ahead. The essay takes this answer seriously as a correct response to a relative-position payoff and follows it through game theory, safety science, resilience engineering, ecology, and finance. It distinguishes the assurance (Stag Hunt), dilemma, and Chicken structures a competitive race can occupy, and proposes a conditional model in which escalation can move a race between them by consuming margin and raising the value of continuation. It gives "unsustainable" an operational meaning (drawing down a finite resource faster than it replenishes), distinguishes deliberate from forced correction, and explains why delayed correction is harsher rather than merely later. Surveying coordination mechanisms that have worked, it finds that all of them change confidence, payoffs, or constraints rather than requesting virtue. Seven domains are mapped explicitly onto the model. The conclusion: before the limit, slowing down is a coordination problem; after it, slowing down is no longer a decision.

Sources (35)

Status note. The drafting session had no web access. Independent verification passes were subsequently run against primary texts where accessible, official records, publisher material, and authoritative scholarly sources. The "Verified" column records whether the specific use in this essay was checked; the fact-check table records qualifications forced by that review.

Fact-check table (35)

Statuses: VERIFIED (the claim was checked against primary, official, publisher, or authoritative scholarly material), QUALIFIED (the claim was narrowed or explicitly bounded because the stronger version was not established).

Editorial note: original synthesis

Established terminology used as such: Stag Hunt / assurance game; Prisoner's Dilemma; Chicken / Hawk-Dove; security dilemma; arms race; collective action; tragedy of the commons; positional goods; tournament theory; rat race; focal point; shadow of the future; drift into failure; normalization of deviance; shifting baseline syndrome; graceful extensibility; alternative stable states and hysteresis; interactive complexity and tight coupling; cascading failure; financial instability hypothesis; circuit breaker.

Adapted from established work: Rasmussen's boundary model is applied to competitive escalation with the rival as the gradient; Rasmussen's own framing of risk management in a competitive environment makes this an application rather than a distortion. Minsky's hypothesis is applied to escalation generally; he wrote about financial cycles. Scheffer's hysteresis is invoked for the possibility of discontinuous, partly irreversible correction, not as a claim that competitive systems generally have this property. Jervis supports the premises about verification and offense–defense distinguishability; the formulation "assurance collapses toward a dilemma" is the essay's.

Analogies used heuristically: the car throughout; the commons as a model for safety margin (§3.3 notes where the analogy fails); kinetic-energy scaling as a model for why late correction is harsher (§7 is explicit that the systems version has different mechanics); "side payment" for the Montreal Protocol's cost-sharing mechanism.

Coined or synthesized in this essay:

• The conditional dynamic model in which escalation can move a race between game structures (assurance → dilemma → Chicken-like region) by consuming margin and raising the value of continuation through lock-in, irreversible commitment, and dependence (§3.2). The individual games are established. The trajectory between them is this essay's proposal, and it is a conditional model, not a theorem: whether and when the ordering changes depends on how payoffs vary with escalation in the particular race.
• Deliberate correction vs. forced correction (§6.1), proposed as a cross-domain framing of which graceful/brittle, soft/hard landing, and gradual/catastrophic shift are domain instances.
• The operational definition of "unsustainable" (§5), with the explicit separation of "cannot continue indefinitely" from "will stop soon."
• The observation that information and incentive arrive in the wrong order (§7).
• The seven-row mapping discipline for applying the model to domains (§9).
• The "competitive escalation trap" formulation from the original brief was tested and not adopted as a named term; its conditions are expressed in existing vocabulary.

Where the literature does not support as strong a claim as the prose initially suggested: Draft 1's §3.1 said dilemmas do not yield to reputation; repeated dilemmas do (Axelrod & Hamilton), and the text now says so. Draft 1's §3.2 asserted a definite transition into Chicken at a definite speed and attributed the assurance→dilemma shift to sunk costs; both were overclaims relative to the payoff logic and are now stated conditionally with the correct mechanisms. Draft 1's Perrow sentence compressed interactive complexity and tight coupling into one mechanism. Draft 1's fisheries sentence said the correction eliminated the prize; it eliminated commercial access to it. The brief's opening formulation implied that unsustainable races end in failure; the essay narrows this (§5.3, §6.3) to "unless a stabilizer or coordination intervenes." Domains cut for incomplete mapping: cybersecurity, advertising, social media, price competition, environmental extraction.