How to write rules for a thing that rewrites itself.
Law is the slowest thing civilization builds on purpose. Frontier AI is the fastest. In 2026, governments and labs alike began publishing frameworks for governing the exponential. Most of them will fail for the same reason: they regulate a snapshot of a moving object.
In brief
- You cannot regulate a compounding technology with a fixed rulebook; by the time the ink dries, the thing described no longer exists.
- The way out is to regulate the ratio, not the artifact: bind the speed of change to the speed of correction and keep R = L / H below 1.
- Each of the book's five gears translates directly into a legislative instrument, and the hardest clause of all, the pause, is empty without verification.
Policy has a clock speed too
The two clocks of the book are usually read as an engineering story: systems change faster than their operators can check. But the purest example is not a trading system or a reactor. It is legislation. A bill is drafted against the technology of one year, debated against the next, enforced against a third. Committees, comment periods, court challenges: the machinery is not broken, it is doing exactly what it was built for, in a world that no longer waits for it.
So the honest question for AI policy is not "what should the rules say?" It is "how can a slow rulebook stay true about a fast object?" Any framework that cannot answer that question is a photograph of a river.
Regulate the ratio, not the snapshot
The book's answer is to stop legislating descriptions and start legislating a relationship. Danger, in the equilibrium law, is how fast a thing changes divided by how fast we can correct it. A rulebook that wants to stay true does not freeze the numerator. It binds the two speeds together, so that whenever capability compounds, correction is forced to compound with it.
Concretely, that means rules written as feedback loops rather than fixed lines: capability thresholds that trigger obligations automatically instead of waiting for a new bill; reporting cadences that tighten when the pace of progress rises; review authorities empowered to act on evidence between legislative sessions. The test of every clause is the same one the book applies to every system: does this keep R = L / H below 1 as the denominator's job gets harder?
Five gears, five instruments
Read through the book, the anatomy of a workable frontier-AI framework is not mysterious. Each gear names an instrument a legislature already knows how to build.
Governance
An authority that can say stop before harm, not litigate after it. Post-hoc liability is an autopsy; the gear demands a brake.
Equity
The upside must be broadly shared, or the public that bears the risk will eventually vote to smash the machinery rather than steer it.
Aligned incentives
Penalties that scale with the violator's size, so caution is cheaper than recklessness for the largest actors, not only the small ones.
Resilience
Detection, containment, and recovery funded in peacetime. No crisis has ever waited for a response capability to be stood up.
Steering
Sunset clauses, evidence reviews, thresholds that move: the rulebook itself must remain steerable, or it becomes the next unsteerable system.
The clause nobody can write alone
Every serious framework eventually arrives at the same sentence: in an emergency, development can be paused. And there the drafting stops, because a pause clause has a defect no single government can fix. A pause you cannot verify is a pause your rivals will not believe, and a pause your rivals will not believe is one nobody can afford to honor. The clause reads as safety and functions as surrender.
This is why the book insists that the deepest problem in AI governance is not speed limits but verification: infrastructure that lets rivals confirm each other's restraint without exposing each other's secrets. Build that, and the pause clause becomes real everywhere at once. Skip it, and every framework on earth shares the same dead letter. The book's proposal for that infrastructure, the Verifiable Compute Commons, is the subject of its own page.
The moment to hold the map
None of this is an argument against the frameworks now appearing; their existence is the good news. After years of the question belonging to labs and philosophers, 2026 made "how do we keep control of fast technology" front-page policy in multiple capitals. The argument is about what to demand of them. Ask of every proposal, whoever publishes it: does it bind correction to capability, or does it photograph the river? Does its pause clause come with a way to check? Is the rulebook itself steerable?
Those are not partisan questions. They are the control ratio, asked politely. The reader who carries R = L / H into the hearing room owns the debate, because every good framework turns out to be an attempt to answer it, and every doomed one turns out to be an attempt to avoid it.
Further reading: for one example of the 2026 wave of frontier-AI governance proposals, see Anthropic's public policy materials. This essay is independent commentary; no affiliation or endorsement implied.
Keep reading
The brake rivals can trust
The verification infrastructure that makes every pause clause on earth mean something. Six layers, five principles, one page.
Previous essay · 6 minWhen AI builds itself, the bottleneck becomes you
Why the reviewer is now the slow part of the loop, and what an honest response looks like.