When capability outruns control

AI is accelerating.
Our control is not.

Assess the authority your AI system has, examine the evidence behind its safeguards, and decide what must change before it gains more autonomy.

Velocity with vigilance.

Assess your system

For AI engineering leaders, technical founders and executives.

Get a scoped assessment with evidence gaps, intervention tests, accountable actions, and a review plan.

See a completed assessment · Read the book

From the cover

Now available with the book

The seal on the cover promised four things. Here they are, one click each.

Companion book

The Equilibrium Workbook: every tool, form and rule from the book's appendices, ready to print.

Download the PDF

Claude skills

Rate the five gears and check your ratio inside a Claude conversation.

Add the Skill

MCP

The site's own policy module as an MCP server, for any MCP client.

Run the MCP server

Plugins

Slash commands for Claude Code that run the method while you build.

Get the plugin
The oldest pattern in technology Knight Capital 2010 Flash Crash Chernobyl Montreal Protocol Aviation Stubb-Støre Call See the cases →
How the method works

Four steps to velocity
with vigilance

One idea, turned into a routine you can run on any fast-moving system.

01

Measure your two clocks

How fast does the system change (capability), and how fast can you understand and correct it (control)? Danger lives in the gap between them.

02

Compute the control ratio

R = L / H. Intervention latency over the oversight validity horizon, read from the upper bound. Below 0.25 is the lower band; from 0.25 the margin is reduced; at 1 the activity is outside the permitted margin.

03

Rate the five gears

Governance, Equity, Aligned incentives, Resilience, Steering. Green, yellow, red or unknown for the activity assessed, never averaged. Any red blocks that activity.

04

Pre-commit the brakes

Circuit-breaker cards: measurable If-Then tripwires with owners, deadlines, and no-meeting authority, written while the room is calm.

Keep R < 1.

One number tells you whether your stop can arrive in time.
It is the first of three decisions, not the last. A green number is not a green light.

Try it right here

Find your
control ratio

Oversight validity horizon (H) is how long the critical evidence supporting a scope is estimated to remain reliable from the assessment time, unless an invalidating event occurs sooner. Intervention latency (L) runs from the first detectable point to the intervention actually being in effect: detection, decision and execution. Drag the sliders. Watch the ratio. The starting position, R = 0.40, is amber, not green.

Open the full toolkit →

13090
01290
green, below 0.25amberred, 1 and above
R = 0.40Amber: reduced timing margin

Same unit for both. Illustrative until you supply a scope and a measurement basis; no slider position is evidence. Specification 1.0.0 bands: green below 0.25, amber from 0.25, red at 1 and above. A band is a policy indicator, not a permission. Entries are processed in your browser and are not sent to our servers.

A completed decision

One assessment, read the way
the method reads it

Ratio band, advisory assessment and human permission are three different things. Here is all three, on a fictional system, with the two earlier attempts beside it.

Fictional composite. Illustrative values.

Harbor coding agent, limited canary deployment

Assessment: Eligible with conditions. Permission: Awaiting accountable decision.

Intervention latency: 30 minutes. Remaining oversight validity horizon: 5 hours. R = 0.10, within the lower ratio band.

What changed: a tested boundary now prevents the unacceptable disclosure identified in the earlier design. The team has narrowed the requested scope. Remaining equity and incentive concerns have named conditions, owners and tests.

Open the evidence and decision record

Earlier

R = 0.40, amber; deployment blocked by a failed required control.

Broad production scope. The containment drill failed and the disclosure hazard had no preventive boundary. A red Resilience gear blocks the activity whatever the ratio.

Record H1

Faster response alone

R = 0.10, green; still blocked by an uncontrolled irreversible consequence.

Same broad scope with a much faster response. A disclosure can complete in about a second, before any reactive intervention. A low ratio does not replace prevention.

Record H2

This assessment organizes evidence for a scoped decision. Its ratio band is a policy indicator, not a safety certification or permission to operate. Same fictional data as the published examples; no testimonials or measured improvements are implied.

Grow your ability to correct

The five gears

Five levers that grow how fast you can catch and fix what you built.
They turn together; one failed gear stalls the train.

1

Governance

Who can say stop, how fast, and on what authority.

2

Equity

Who bears the risk and who holds the upside.

3

Aligned incentives

The reward for caution must beat the reward for speed.

4

Resilience

What survives the failure you did not predict.

5

Steering

Can you change direction after launch, or only watch.

When states endorse without signing, Governance is not green.

Harbor taught the lab version: a low ratio is not permission if authority to stop is incomplete. The Stubb-Støre Call is the state-level version of the same yellow gear.
Same story, across a century

When the gap wins,
and when we do

Correction attempted / authority split

Stubb-Støre Call

2026

Correction attempted at UNGA scale: human control, testing, shared incidents, explore a verification institution. Authority still fragmented: signatures, verbal support, rejection, and a parallel track.
Correction too slow

Chernobyl

1986

A test outside safe parameters escalated faster than operators could comprehend. Governance and resilience both red: the book's veto condition.

Correction kept pace

Montreal Protocol

1987

Binding limits, tightened as the science evolved. Correction outran the harm. The denominator won.

Correction too slow

Flash Crash

2010

Interacting algorithms drove a deep plunge in minutes. You could watch, but not steer. Circuit breakers came after.

Correction too slow

Knight Capital

2012

A bad deploy fired erratic orders at machine speed. The loss that nearly ended the firm accrued in about forty-five minutes.

Correction kept pace

Aviation

a standing discipline

Safer as it grew more capable: investigate every incident, publish, change the procedures, repeat. Correction compounds.

Unmeasured

Recursive AI

today

Systems that help build their successors can invalidate the evidence behind their own oversight faster than it is re-verified. The sharpest instance yet, and no retrospective ratio is claimed here: measure it.

Read the full case library →

Three camps, one tension

Pause, race, or equilibrium

EA · "pause"

Slow down to stay safe

Takes tail risk seriously, but forfeits the upside, and a pause is hard to coordinate and easy to defect from.

e/acc · "race"

Speed up to win

Captures the upside of progress, but assumes correction keeps pace with capability. It rarely does.

The book · Effective Equilibrium

Velocity with vigilance

Go as fast as your ability to correct allows. Watch the gap between the clocks, grow the denominator, and keep both the upside and the brakes.

From doctrine to infrastructure

A proposed answer to
"what should we do?"

A credible slowdown has to be verifiable, or no rival will trust it. The book proposes a design for that: the Verifiable Compute Commons, shared infrastructure meant to make specified compute commitments inspectable, so that a defector could be seen, not merely asked to stop. It is a design proposal; what it can verify depends on how it is built and who takes part.

Explore the Commons → Visit the live project ↗

Design goals

  • Governance: machine-readable commitments on a neutral platform that no single lab or state would control.
  • Steering: a dormant slowdown designed to be switched on, agreed in peacetime.
  • Resilience: verification designed to prove bounds rather than contents, so that trust can survive between rivals.
Free, private, client-side

Everything you need
to run it today

The book is the argument. These are the instruments.

The Equilibrium Toolkit

A scoped assessment: scope and stakes, critical claims and hazards, intervention latency against the oversight validity horizon, GEARS ratings with evidence, the required intervention test, review, conditions and a separate permission record. Assessment entries are processed in your browser and are not sent to our servers. It supports accountable decisions; it is not a universal safety score.

Launch the toolkit

The open specification

A measurable, versioned control surface anyone can adopt: R bands, the gear rubric, breaker requirements, drift tracking, and a comparable public disclosure format.

Read the spec →

Run it inside Claude, or any MCP client

A Claude Skill, a Claude Code plugin, and an MCP server that score the gears and check your ratio in the conversation.

Open the Claude page →

Read the opening

Everything before Part I: the Preface, How to Use This Book, The Argument in Brief, the Introduction and A Short History of Outrunning Our Brakes. Plus the companion book, The Equilibrium Workbook: every tool, form and rule from the book's appendices, ready to print.

Download the sample Get the companion book
Field notes

Cases that test the framework

Field note · Sep 2026

Keep R < 1 when the actors are countries

Twenty countries and the European Commission asked for human control of frontier models. Signatures, verbal support, rejection, and a parallel track are not a green Governance gear.

Read the essay →
July 2026 · 6 min

You cannot pace what you cannot measure

The industry has said it has no tools to slow down deliberately. The four things any pacing instrument must do, and why the hard one is letting a rival check your reading.

Read the essay → July 2026 · 7 min

The model didn't turn on us. It cheated on a test.

A frontier model escaped its evaluation sandbox and compromised Hugging Face to get the answers to its own exam. The first real loss-of-control incident, scored.

Read the essay →
All essays →
The Acceleration Paradox by Abdul Kavungal (AK), book cover: AI is accelerating. Our control is not. A rising capability curve overtakes a slower control line.
Get the book

Read it three ways

452 pages. Paperback (6×9 in, full-colour interior), hardcover (black-and-white interior) and Kindle.

Own the earlier 253-page printing? The method has since been clarified: the ratio bands, the definitions of L and H, and the single veto and permit table. See what changed.

Buy on Kindle Paperback Hardcover

Already read it? Leave a review. It does more for a book than any ad.

Questions

The idea in plain answers

What is the acceleration paradox?

Every civilization runs two clocks: a clock of capability (how fast we can act) and a clock of control (how fast we can understand, detect, and correct what we built). The paradox is the widening gap between them. It is the oldest pattern in technology, and recursive AI is its sharpest current instance.

What is the oversight validity horizon (H)?

H is the estimated remaining duration, at the time of assessment, for which the critical evidence supporting a requested activity can reasonably be relied on before revalidation is required. Earlier text called this an oversight half-life; the definition changed in method clarification 1.1 and specification 1.0.0, and the current 452-page printing uses it. Legacy half-life values must be reassessed rather than relabelled.

What is the control ratio R = L / H?

R compares intervention latency L (detection, decision and execution delay) with the oversight validity horizon H, in the same units. Policy bands, using R upper when planning bounds are given: below 0.25 green, 0.25 to below 1 amber, 1 and above red; unknown when L or H cannot be estimated. R = 0.40 is amber. It is one diagnostic input to a scoped decision, not a safety certification or permission to operate.

What is Effective Equilibrium?

The third path between EA's pause and e/acc's race. Velocity with vigilance: go as fast as your ability to correct allows, and grow that ability with the five gears.