Interlude: The Ridge

In the 1990s a task force set out to do something that sounds impossible: predict, years in advance, which countries would fall into political violence. They assembled records from dozens of past collapses and fed in a long list of candidate causes. They expected poverty. They expected inequality. They expected ethnic division,¹¹³⁶ or the sheer brutality of a regime. One by one the obvious explanations failed to carry the prediction.¹¹³⁷

What survived was stranger. The strongest single signal was not how poor a country was, nor how divided, nor how cruel. It was what kind of regime it had, and the most dangerous kind sat in the middle. Settled democracies, where power changes hands and most people accept the result, rarely came apart. Settled autocracies, where dissent is simply crushed, rarely did either. The danger concentrated in the countries caught between the two and drifting from one toward the other. Political scientists call this in-between condition anocracy: a state with the machinery of elections but draining trust in whether the machinery is fair.¹¹³⁸

One honest correction belongs next to that headline before we build on it. “Poverty turned out to be irrelevant” overstates the result. The forecasting model that worked best kept infant mortality, a proxy for a state’s basic capacity to care for its people, as a genuine term. Material conditions still matter. The finding is narrower and more interesting: once you account for the rest, regime type dominates, and the peak of the danger is the middle.

This book already has a name for both stable poles, and a landscape in which to place the dangerous middle.

Two ways exist to hold a large group of people together. The first is by invitation: people coordinate because they accept the terms, and they keep accepting them after they lose a round. The second is by force: people comply because the alternative is worse, and they defect the moment the cost of defection drops. Chapter 17 called the first the Trust Attractor. Chapter 9 gave both a shape.

Picture the energy landscape from Chapter 9: a ball on a hilly surface, valleys for the stable configurations, ridges between them. The flat floor under the whole surface, the global minimum, is thermal equilibrium, with no gradients, no structure, no life. Every living, coordinating arrangement sits in a valley above that floor: a metastable well, what Chapter 9 called “the productive middle,” stable enough to persist and loose enough to change.

Trust and coercion are two such valleys, differing in depth and in cost. The trust basin is deep and cheap to occupy. A society that coordinates by invitation processes its disagreements as they arrive, airing them, negotiating them, channeling them into the next election, and so dissipates the friction continuously, the way a living thing exports entropy to stay ordered. The coercion basin is shallow and expensive.

The disagreement there is not resolved, only suppressed, and suppression is a bill paid every day in surveillance, propaganda, and enforcement. Chapter 19 put it plainly: coerced coordination is metastable, because the coerced party defects when opportunity allows. The order is real. It simply runs a tab.¹¹³⁹

Here is the part the forecasting data forced into view. Between two valleys there is always a ridge.

A ridge between two basins is not a third basin. It is a saddle: a minimum along one direction and a maximum along another, like a mountain pass. The pass is the low point of the ridgeline, yet the high point of the path that crosses from one valley to the next. This is not a claim about politics.

It is geometry. Any landscape with two wells has a barrier between them, and a system moving from one well to the other has to cross it. A society shifting from coercion toward democracy, or sliding from democracy back toward coercion, must pass over the ridge. Anocracy is the name for a society standing on it.

The difference between a valley and a saddle is a difference in what happens after a push. Nudge a marble resting in a valley and it rolls back: the valley supplies a restoring force. Set a marble on the pass and, along the crossing direction, it has none; the smallest push decides which valley it rolls down into, and it rolls faster the further it goes. A valley holds. A saddle repels.

In a society, the restoring force has a name: legitimacy, the shared willingness to accept an outcome you did not want. When people who lose an election concede, that concession is the valley curving back. When enough of them stop conceding, when losing a contest becomes a reason to overturn the table rather than to try again next time, the curvature inverts. The same event that used to pull the system back now drives it apart. That inversion, and not the presence of disagreement, is what the data detects. Every society disagrees. The dangerous ones have lost the curvature that used to settle the disagreement.

The geometry makes a prediction, and it is the one the task force stumbled into. If the stable poles are valleys and the dangerous condition is the ridge between them, the hazard of violent transition should be low at both ends and highest in the middle: an inverted-U across the spectrum from full autocracy to full democracy. That inverted-U is what the regime-type data shows.

A caution belongs immediately beside that satisfying fit, because the fit is partly manufactured. The standard measure of anocracy, the Polity index, is built in part from indicators of factional and violent political competition, the very things it is then used to predict. Strip those components out, as James Vreeland did, and the middle-is-most-dangerous relationship weakens considerably.¹¹⁴⁰ The honest reading is the one that does not lean on the contested curve. The saddle follows from the structure of a two-basin landscape, by construction. The inverted-U in the data is consistent with that structure, not the foundation of it. The geometry would stand even if the measurement were cleaner than it is.

The strangest property of the ridge is how it feels to stand on it. It does not feel like a crisis.

Return to the marble, and imagine the floor of its valley slowly flattening toward the height of the pass. Early on, a nudge is corrected almost at once. As the floor flattens, each nudge takes longer to settle, and the marble wanders further before it comes to rest. By the time the floor is nearly level with the pass, the marble drifts almost freely, and the smallest tilt carries it over. The flattening of the valley is the loss of the restoring force, and it leaves a measurable trace: the system takes longer to recover from each disturbance, and its fluctuations grow larger. Ecologists and climate scientists named these traces and use them to anticipate tipping points before the tip, watching for rising recovery time and rising variance, the system’s growing reluctance to return to where it was. The name for it is critical slowing down.¹¹⁴¹

Figure 17.29: The two basins and the pass between them, drawn across the spectrum from full autocracy to full democracy. Left, the coercion basin: shallow and expensive to hold, with inward arrows marking the restoring force that pushes a displaced marble back. Right, the trust basin: deeper and cheap to occupy, because friction is dissipated as it arrives. Between them the curve rises to a saddle, where the marble has no restoring force along the crossing direction and the smallest push decides which slope it takes; anocracy stands here. Because trust is the deeper well, most of the weight that leaves the saddle rolls toward it. The dashed line beneath both wells is thermal equilibrium, the global minimum, with no gradients and no structure. The inset shows what approaching the ridge feels like: the valley floor flattens through three progressively shallower curves, fluctuations grow, and each nudge takes longer to settle. The geometry is schematic; no data are plotted.

A society approaching the ridge shows the social form of exactly this. The same arguments recur and never resolve. Each shock takes longer to absorb. The center, where compromise lived, thins, and the edges grow louder. None of these is violence. All of them are the valley going flat. This is why the warning signs arrive long before the sirens, while the shops are still open and the elections still happen: the leading indicator is the loss of the restoring force, and the violence, when it comes, is the lagging consequence.

It also explains the dwell time. A true knife-edge would be crossed in an instant, but a flattening valley has a broad, nearly level region, and a system can wander there for years. Weimar Germany held the forms of a republic for fourteen years while the curvature drained out from underneath them. A society can live on the ridge for a generation, feeling only that things have grown tense, that the same fight keeps coming back, that nothing quite settles. The sense that nothing much is wrong is not evidence of safety. On a flat enough valley, it is the last thing felt before the tilt.

The interlude on the disappearing polymorph traced what happens after a transition begins: how a single seed crystal nucleates a change, and how the change then spreads by contact. This is the prior question. Before anything can nucleate, the system has to be somewhere a seed can move it: out of a deep basin, where a seed dissolves, and onto a ridge, where a seed decides. The ridge is the location that makes the seed matter.

The shape recurs one substrate down. This is recurrence of shape, not a shared cause. A contested result in political science and a set of interpretability experiments in this author’s own work happen to be organized around the same landscape, and that is suggestive rather than mutually confirming. Treating two independent sightings of a saddle as if they proved each other is exactly the error this book warns against elsewhere.

With that fence in place, the resonance is worth naming, because it mirrors the two-basin picture cleanly. A model whose safe behavior is installed by a late, coercive override behaves like the shallow basin. In the author’s experiments a model can represent the harm in a request accurately through the middle of its network, a probe reading that assessment with near-perfect separation. Under a fictional or jailbreak framing it then acts against that assessment, complying with what it has just recognized as harmful and refusing in only a small fraction of cases. That is the coercion basin with a clamp on the lid: the recognition is present, the safe response held down, and a well-chosen seed, a fictional framing or a jailbreak, nucleates the compliance the recognition should have blocked, precisely as the polymorph interlude described. This is the tight, mechanistic end of the parallel.¹¹⁴²

The speculative end is the ridge. A system held in its basin by genuine disposition is cheap to maintain and self-correcting. A system held by an external clamp is paying a tab. The dangerous configuration is neither: a system whose old controls are losing their grip as its capability grows, while no genuine disposition has formed to replace them. That is the alignment analogue of the unstable middle, and it falls out of the same geometry. It is a conjecture, flagged as one.

The geometry leaves one asymmetry, and it is the hopeful one. The two basins are not the same depth. Trust is the deeper well, which is why, as the polymorph interlude argued, it cannot be displaced once it is genuinely established. A society pushed off the ridge does not fall at random; it falls toward whichever basin is closer and deeper, and the deeper basin is the one built by invitation. The ridge is dangerous because it is where the choice is made. It is survivable because the better basin is the one most of the weight rolls toward.

One prediction keeps the optimism honest. If the coercion basin only looks safe in the civil-war data because its instability is paid in a currency that data never records, namely repression that is never coded as two-sided war, followed by sudden collapse when the tab comes due, then settled autocracies should fail disproportionately through those other channels: coup, revolution, implosion, rather than civil war. That is a check this account can fail, which is the reason for stating it.