Case studies
The acceleration paradox is the oldest pattern in technology, not a one-off glitch. Whenever the clock of capability outruns the clock of control, when correction takes longer than verification stays valid (R = L / H ≥ 1), the same failure repeats. Recursive AI is its sharpest current instance, not the only one. Frontier labs now say so too: Anthropic warns that AI is beginning to improve itself faster than humans can review it, and the answer it reaches for, verifiable coordination, is a denominator move this book maps in full.
EA vs e/acc vs Effective Equilibrium
| EA: “pause” | e/acc: “race” | Effective Equilibrium: “velocity with vigilance” | |
|---|---|---|---|
| Core move | Slow or halt to avoid catastrophic risk. | Accelerate; progress solves its own problems. | Go as fast as your ability to correct allows. Keep R < 1. |
| Which clock it watches | The clock of capability (treats it as the threat). | The clock of capability (treats it as the prize). | The gap between both clocks: that is where danger lives. |
| Lever it pulls | Brake the numerator (how fast things change). | Ignore the denominator (how fast we can correct). | Grow the denominator with the five gears. Brake only when you must. |
| Strength | Takes tail risk seriously. | Captures the upside of progress. | Keeps the upside and the brakes: a portable test instead of a posture. |
| Failure mode | Forfeits benefits; a pause is hard to coordinate and easy to defect from. | Assumes correction keeps pace. It rarely does. | Demands honest measurement. Vanity dashboards defeat it. |
The case library
Equities market making · August 1, 2012
A faulty software deployment began firing erratic orders into the market at machine speed. By the time humans understood what was happening and pulled the plug, the damage was done: the loss that nearly ended the firm accrued in roughly forty-five minutes.
Two clocks: capability acted in milliseconds; control needed tens of minutes. L ≫ H, so the outcome was decided before anyone could decide anything.
US equities · May 6, 2010
Interacting automated strategies drove a sudden, deep plunge and partial recovery within minutes. No single actor intended it; the system moved faster than any participant could interpret or arrest in real time.
Why it fits: the steering gear failed: you could watch, but not turn the wheel fast enough. Circuit breakers were the resilience patch added afterward.
Nuclear power · April 26, 1986
A test run outside safe parameters, with safety systems disabled, pushed the reactor into a state that escalated faster than operators could comprehend or counter. Understanding arrived far too late to correct.
Governance + resilience, both red: on an irreversible, high-stakes system, that combination is the book's veto condition: stop, no matter the schedule.
Ozone layer · 1987
Faced with a global, slow-moving threat, governments detected the problem, agreed on binding limits, and adjusted them as the science evolved. Correction kept pace with, and then outran, the harm.
The denominator won: governance and steering were strong enough that humanity could notice, decide, and act before the gap became unrecoverable.
A standing discipline of correction
Aviation grew safer as it grew more capable, not by slowing down, but by building a relentless feedback loop: investigate every incident, publish the findings, change the procedures, repeat. Correction compounds.
The doctrine in practice: velocity with vigilance. Each of the five gears is institutionalized, so H keeps growing alongside capability.
The book's flagship proposal · AI governance
Neutral, open infrastructure that makes a coordinated slowdown credible. Labs verify each other's compliance (FLOP ceilings, training provenance) through hardware attestation, without exposing models or data. The point is not enforcement: a defector is seen, not blocked. It is a worked answer to the verification problem that frontier labs themselves say a real slowdown would require.
Why it matters: a denominator move that buys back oversight by making a credible pause switchable in peacetime, not just a brake on speed. See the live project ›
The lesson is not that speed is bad. It is that speed is only safe when correction can keep up. That is the one test the book gives you, and the one the Scorecard makes routine.