The case library

Same story, across a century.

Whenever the clock of capability outran the clock of control, the same failure repeated. Here is the pattern, with the numbers attached, read through the control ratio R = L / H.

Control lost · R > 1

Knight Capital

US equities · August 1, 2012

A stale test routine woke up on one server that a deployment had missed. At the market open, the firm's systems began firing millions of unintended orders at machine speed. Humans needed most of an hour to find and stop what the software did in milliseconds.

The book's reading: intervention latency was measured in tens of minutes; the system's oversight validity horizon in seconds. The outcome was decided before anyone decided anything.

~$440M

lost before the halt

45 min

from open to shutdown

1 of 8

servers missed by the deploy

Losses compounding at machine speed while humans searched for the cause.

Control lost · R > 1

The Flash Crash

US markets · May 6, 2010

Interacting automated strategies chased each other into a spiral. The Dow plunged roughly a thousand points, close to nine percent, and about a trillion dollars of market value briefly vanished. Most of the fall and the rebound happened inside about thirty-six minutes. No human hand was fast enough to be on the wheel.

The book's reading: a steering failure. Everyone could watch; no one could turn. The circuit breakers that now pause trading were installed only afterward, brakes bolted on after the crash they were meant to prevent.

~9%

intraday plunge in the Dow

~36 min

for the dive and rebound

~$1T

in value briefly erased

The V that a trillion dollars drew in thirty-six minutes.

Control lost · R > 1

Chernobyl

Nuclear power · April 26, 1986

At 1:23 am, during a test run with safety systems disabled, reactor four's power surged roughly a hundredfold in seconds. The operators' understanding of the machine was minutes behind the machine itself. The wider world learned of the disaster only after radiation tripped alarms at a Swedish plant more than a thousand kilometers away, two days later.

The book's reading: governance and resilience both red on an irreversible system, the exact combination the Scorecard treats as a veto. Stop, no matter what the schedule says.

~100×

power surge in seconds

2 days

before the world was told

30 km

exclusion zone, still in force

A curve no operator could chase: seconds up, decades down.

Control held · R < 1

The Montreal Protocol

The ozone layer · 1987

Faced with a planetary threat that moved slowly, humanity moved faster. The treaty became the first in UN history ratified by every member state, and it was built to tighten: amendment after amendment ratcheted the limits down as the science firmed up. Nearly ninety-nine percent of ozone-depleting substances have been phased out, and the layer is on track to heal this century.

The book's reading: the denominator won. Detection, decision, and correction all outpaced the harm. Proof that the gap is a choice, not a law of nature.

198

parties, every UN state

~99%

of ozone-depleting substances phased out

tightening amendments since 1987

The curve that bent: emissions turned within a decade of signature.

Control held · R < 1

Commercial aviation

A standing discipline · 1950s to today

Aviation got faster, bigger, and busier, and safer, all at once. Not by slowing down, but by building the world's most relentless correction loop: black boxes, checklists, blameless investigation of every incident, findings published and turned into procedure. Since the start of the jet age the fatal-accident rate has fallen by well over ninety percent while traffic multiplied many times over.

The book's reading: velocity with vigilance, institutionalized. Every gear turning, so the oversight validity horizon keeps growing alongside capability.

>90%

drop in fatal-accident rate since the jet age

every

incident investigated and published

1 in millions

flights ends in a fatal accident today

Correction compounding: the only industry where speed and safety rose together for seventy years.

R approaching 1

Recursive AI

Frontier AI · now

Every case above involved a product accelerating. This one is different in kind: the thing accelerating is invention itself. By mid-2026, most new code at a frontier lab was being written by its own model, engineers were shipping several times more than two years earlier, and the honest ones were saying out loud that human review was becoming the bottleneck.

The book's reading: "review is the bottleneck" is the plain-English sentence for L overtaking H. The sharpest instance of the pattern yet, and the reason the book exists.

Read the essay: When AI builds itself →

80%+

of one lab's merged code, AI-written

~4 mo

doubling time of unsupervised task length

R → 1

review becoming the bottleneck

The gap, live: generation compounds while verification walks.

Geopolitical reading · illustrative composite · not a legal finding

Correction attempted · authority split

Frontier AI Call, September 2026

UNGA week · Stubb-Støre Call for Control of Frontier AI Models

Twenty-two leaders from twenty countries, plus the European Commission, asked for human direction and control, mandatory pre-deployment testing, shared serious-incident reporting, and exploration of an international verification institution. The United States rejected a global control architecture without naming the Call. China stayed off the list and already had WAICO. France supported, and the United Kingdom welcomed, without signing.

The book's reading: Effective Equilibrium colliding with geopolitics. Correction was attempted at UNGA scale. Authority to stop is still fragmented. When states endorse without signing, Governance is not green. Harbor already taught the lab version: a low ratio is not permission if a gear is still a blocker.

Governance

Yellow

Split signature / support / rejection; missing national republications.

Equity

Yellow

Eval capacity will not fall evenly; access is asked for, not shown.

Aligned incentives

Yellow

Welcome without signature leaves race incentives intact.

Resilience

Unknown

No shared international breaker proven. Unknown, not a milder yellow.

Steering

Yellow

Threshold convening is still explore, not operational.

When states endorse without signing, Governance is not green.

Read the Field Note → Open CALL-2026 → Primary sources →

22 / 20

leaders / countries + EC

Eligible*

with conditions, illustrative

R ?

L and H not jointly measured

Four camps, one Call. Permission: fragmented. Global-layer R: unknown.

The book's flagship proposal

The Verifiable Compute Commons: the brake rivals can trust

One case in this library is still being written. The book's answer to recursive AI is not a plea to slow down; it is infrastructure that makes slowing down verifiable, so a defector is seen, not merely asked to stop. It gets its own page.

Explore the Commons →

What the pattern teaches

Pause, race, or equilibrium

Three camps read this same history and draw three different conclusions.

EA · "pause"

Slow down to stay safe

Reads the disasters and concludes we must halt. Takes tail risk seriously, but forfeits the upside, and history shows pauses are hard to coordinate and easy to defect from.

e/acc · "race"

Speed up to win

Reads the recoveries and concludes progress fixes itself. Captures the upside, but assumes correction keeps pace with capability. Knight, the Flash Crash, and Chernobyl say otherwise.

The book · velocity with vigilance

Effective Equilibrium

Reads both columns. Montreal and aviation prove the gap can be beaten, not by stopping, but by making correction faster than change. Keep R below 1 and keep going.

Score your own project →