Interactive companion to the paper
Agreements Have a ClockCompute Growth and the Verification of AI Restraint
Rival powers have agreed to limit nuclear arms, phase out ozone-depleting chemicals and cut greenhouse gases, though each agreement was hard to reach. This paper asks what an agreement to pause frontier AI needs. Compute keeps growing during a pause, so a secret advance reaches the frozen frontier sooner and the response window shrinks: agreements have a clock.
Each step below follows a section of the paper. The figures run the paper’s model in your browser and open at the paper’s settings, where every number equals the one the paper prints. Move a slider to see what drives a result; “Paper’s settings” puts it back. The longer searches of Steps 4, 5, 10, 13 and 14 run when you reach them. All numbers come from an illustrative calibration and illustrative national values; they are not estimates of real-world risk or of any country’s position. The paper treats its 75 percent full-speed catastrophe probability as the upper end of a range that runs down to 25 percent (Step 10).
Contents
- Five findings at a glance
- One country’s restraint
- One frontier
- A follower and diffusion
- Catch-up time
- Overtaking
- Restraint by both
- Gains from mutual restraint
- The response window
- The game’s payoffs
- Deterrence and participation
- The clock
- The agreement clock
- Matched safety and the scenario’s risks
- When rivals are more alike
- The lead and domestic rules
- Widening the lead
- Instruments with a racing follower
- Closure and foreign safety
- Limits
- What the model leaves out
At a glance · Paper, Introduction
Five findings, and the one mechanism behind them
One concrete proposal for a pause agreement is Plan A of the AI Futures Project. The United States and China agree in 2029 to pause frontier development from 2035 to 2040, with declared compute, counted chips and audited data centers. The paper’s result is a test that Plan A, and any agreement like it, must pass. A mutual pause is worth having, but it holds only while each country can still answer a secret advance by the other. The diagram below shows why that time runs out as compute grows. The fifteen steps that follow build the argument piece by piece; here is where they lead.
A way to picture it: two race cars, one ahead, joined by an elastic cord. The follower has its own engine, and the cord pulls it harder the farther behind it falls. If the leader parks, the pull fades as the follower closes in, so the catch-up takes longer than it would if the follower kept its old speed. A pause is a red flag that stops both cars, but the mechanics keep tuning the engines because compute keeps growing. So the longer the flag stays out, the stronger the engines. A follower that restarts without telling the other then catches the parked leader sooner, and the other has less time to notice and answer. That shrinking time is the clock.
- A country that restrains alone, by pausing or slowing its capability research, can be caught up. A follower advances by its own research and by learning from the leader, and there is less to learn as it closes in, so it slows toward the pace of its own research alone. A halted leader therefore stays ahead longer than its raw lead: in the calibration a follower six months behind needs 12 months, not six. The room is still finite, and a leader whose speed is below the follower’s own research speed is overtaken unless it finishes first. Steps 2–4
- A pause by both countries pays each of them. Both countries stop capability research and keep doing safety research, so the gap stays fixed and neither concedes progress to an advancing rival. This assumes both comply and the first country to finish ends the danger for both, the so-called first-arrival closure rule. A 12-month mutual pause then lowers the global catastrophe probability from 80 percent, when both race, to 27 percent. Under the game’s payoff assumptions, the gains each country receives can pay for verification. Of course, each country still has an incentive to advance on its own. Step 5
- An agreement must meet the following conditions. First, monitoring must find a secret advance while there is still time to respond; the shorter the response window, the more monitoring deterrence takes. Second, the monitoring cost must not exceed what either country gains from cooperating, and restarting after a timely finding must be worth its extra cost and risk. Cooperation is supportable only when both hold. Deterrence does not need near-certain detection: at the start of the pause, a 32 percent chance of a timely finding already makes a secret advance pay no more than cooperating. Mutual racing is also an equilibrium. Steps 6–8
- Agreements have a clock. Compute keeps growing during a compliant pause, so a secret advance reaches the frozen frontier sooner and the response window shrinks even though the capability gap has not changed. The monitoring needed rises until it costs more than the gains from cooperating. In the paper’s example with a fixed leader window, cooperation is supportable through month 18.9. At the baseline hazard level and with the scenario’s own risks, a race early in the pause is too dangerous for the threat to restart to be credible, so the agreement also has an opening date and is supportable only from month 6.1 to 10.4. At 25 percent there is no opening date, but the window is far shorter. Cheaper monitoring, a faster response and verified matched safety research move the end later. Steps 9–11
- Domestic pacing rules can raise global risk. With the follower racing, a monthly risk ceiling and an R&D-model lag lower the leader’s own catastrophe probability but raise the global one, because both slow development late in the takeoff period, when the follower can pull alongside at high capability. Under the baseline closure rule the ceiling raises the global probability from 80 to 97 percent. Cutting the follower’s own research or diffusion lengthens catch-up time instead. Steps 12–14
A compliant pause: both countries stop advancing capability and keep doing safety research, so the capability gap stays fixed
What cooperation is worth
- Gain
- each country’s gain over a race: lower catastrophe risk and development costs, with its strategic position kept
- Pays for
- monitoring up to the participation bound, 0.05556 findings per month at the baseline cost
- Over time
- held fixed in the example; matched safety research raises it
What deterrence costs
- Each month
- more compute, so a secret advance reaches the frozen frontier sooner
- Window
- the follower’s catch-up time, which sets the response window, falls from 12 months at the start to 5.74 after 24 months
- Needs
- monitoring equal to the breakout temptation divided by the window, so it rises as the window shrinks
Cooperation is supportable while deterrence costs no more than cooperation is worth
The two bounds meet at a critical window of 6.89 months. In the paper’s example the follower’s window falls below it at month 18.9, before the capability gap closes.
How to read the diagram. At each review date a pause holds only if two things are true: monitoring can deter a secret advance, and each country gains more from cooperating than the monitoring costs it. The left box is what cooperation is worth, which caps the monitoring both countries will pay for; in the example it does not change with the date. The right box is what deterrence costs. Compute growth during the pause lets a secret advance reach the frozen frontier sooner, so the response window shrinks and deterrence needs more monitoring. When the monitoring needed costs more than either country will pay, the agreement’s clock has run out. Steps 6 to 9 compute each piece.
The implication is that an agreement’s terms must keep the response window open or raise what cooperation is worth: cheaper monitoring, a faster response, verified matched safety research, a detection standard that tightens over time, or an earlier scheduled restart. The clock is a property of the physical model; the date depends on the inputs.
Terms in plain words
- Diffusion
- progress the follower, the country behind, gains from the leader’s work; it fades as the gap closes
- Raw lead and catch-up time
- the gap in months at the follower’s pace before a halt, and the time the follower actually needs to reach a halted leader
- Compliant pause
- both countries stop advancing capability and keep doing safety research, so the capability gap stays fixed
- Secret advance
- restarting capability research while appearing to comply; a breakout if the rival does not answer it in time
- Response window
- the time a country has to answer a rival’s secret advance; for an advance by the follower, its catch-up time less the time needed to restart
- Monitoring
- inspections that produce findings of a secret advance at some rate per month; each country pays for it up front
- Review date and supportable
- at a public review date each country decides whether to keep complying; cooperation is supportable there if neither gains from a secret advance and both prefer to take part
- Critical window
- the shortest response window at which the monitoring needed to deter costs no more than cooperation is worth
- Restart
- after a timely finding, the compliant country starts capability development again, which restores a contested race
- Matched safety research
- equal safety research by both countries during the pause, verified along with the pause
Part 1One country’s restraint: the frontier, the follower and catch-up
Step 1 · Paper, Section 2
One frontier: compute builds capability, and hazard builds up on the way
The model follows one AI frontier through a takeoff period. It starts at the Automated Coder, a model that automates machine learning engineering, and ends at top-expert-dominating AI (TED-AI), a model that surpasses human scientists in every domain. Capability is an index from 0 at the first milestone to 1 at the second. A lab splits its compute among customers, capability research and safety research. Installed compute doubles every six months, and research speeds up as models improve.
At full speed a lab puts half of its compute into capability research and 5 percent into safety, and the takeoff period takes 18 months, as in the worked example of the AI Futures Project. It reaches the halfway point after about 15 months, so the second half of the index takes about 3 months.
The catastrophe, a catastrophic failure or a loss of control over AI, is treated as a hazard: the chance per month that it happens now, given that it has not happened yet. State hazard accrues per month at a given capability; step hazard accrues per unit of capability gained. Safety knowledge lowers both at every later level. Unpaced development carries a catastrophe probability of 75 percent, three quarters of it in the last 3 months. A year of safety research at the full-speed share adds about 1.03 units of knowledge, which multiplies later hazard by about 0.36.
A pause stops capability research but not safety research. The figure pauses the frontier at the start and moves its research share into safety. Compute keeps growing during the pause, so the takeoff period that follows runs faster.
Where the baseline numbers come from
| Number | Value | Source |
|---|---|---|
| Full-speed takeoff period, Automated Coder to TED-AI | 18 months | AI Futures Project (2026), worked example: 1.5 years. The paper uses it as an illustrative target. |
| Catastrophe probability at full speed | 75% (range 25 to 75%) | AI Futures Project (2026), worked example; the same 75% for almost no safety effort in Greenblatt (2025). The paper also reports 25%, the lower case of its labs companion, the probability Samotsvety Forecasting (2022) give for misaligned AI takeover by 2100; surveys of AI researchers report medians of 5 to 10% (Grace et al. 2025). An illustrative target. |
| Catastrophe probability under a safety-case regime | 40% (10% at 25%) | AI Futures Project (2026). Treating it as a 12-month pause at the start, which fixes the safety yield, is the paper’s assumption. When the full-speed probability changes, the regime’s probability moves so that it cuts cumulative hazard in the same proportion. |
| Follower’s lag; catch-up time after the leader halts | 6 months; 12 months | AI Futures Project (2026): a raw lead of 4–8 months and about a year to catch up, which the paper adopts as uncertain inputs. |
| Growth of installed compute | doubles every 6 months | Assumed. For comparison, training compute has doubled about every six months since the early 2010s (Sevilla et al. 2022). |
| Full-speed compute shares; research speed at TED-AI | 50% research, 5% safety; 10 times the Automated Coder | Assumed. |
| National values, branch catastrophe probabilities, monitoring cost | see Steps 7 and 8 | Illustrative inputs of the paper, not estimates of any country’s position. |
Capability over time
Catastrophe probability over time
Left panel, capability. Capability, from 0 at the Automated Coder to 1 at TED-AI, against months. The gray dashed curve is full speed; the blue curve pauses for the months on the slider and then restarts at full speed. The flat stretch is the pause. The climb after it is steeper than at full speed, because compute grew in the meantime.
Right panel, catastrophe probability. The cumulative probability of catastrophe on the same two paths. At full speed it rises slowly at first and fast near TED-AI, ending at 75 percent. During the pause step hazard stops, while state hazard continues at the frozen capability and falls as knowledge grows.
The implication is that slowing down has a benefit and a cost. Each capability level is reached with more knowledge, but more months are spent at each level. In this calibration the benefit is larger. The rest of the page adds a rival.
The math
Step 2 · Paper, Section 3.1
The follower advances by its own research and by diffusion from the leader
Unilateral restraint exposes a country to an advancing rival. Even if its own labs follow a common rule, a government fears falling behind an unpaced rival. The rival faces the same fear, and that shared fear is what makes a mutual agreement worth having. The paper models the rival as a follower frontier with the same takeoff technology. It starts six months behind the leader, below the Automated Coder milestone, and runs at full speed. Its progress has two parts:
Own research grows with the follower’s research capacity and its own capability. Diffusion closes a share δ of the capability gap each month, so it is fast while the gap is wide and fades as the gap closes. The calibration matches the six-month lag and a catch-up time of about a year, the time the follower would need to reach a leader that halts. It gives δ = 10.6 percent a month, and diffusion accounts for 81 percent of the follower’s initial progress. This share is implied by the model, not measured.
The follower’s safety knowledge grows at half the leader’s full-speed rate, and its hazard counts from the Automated Coder on. So its frontier is an independent source of global catastrophe.
Capability of the leader and the follower
The follower’s speed during the leader’s pause
Left panel, capability paths. Solid lines are the leader and dashed lines the follower. Blue is full speed; orange is a leader pause of the length on the slider, with safety research. At full speed the gap stays wide. During the pause the follower closes in, slowing as it gets near, and the leader pulls away again when it restarts.
Right panel, the follower’s speed. On the pause path, the green curve is the follower’s own research and the blue curve its diffusion, in capability per month. Diffusion dominates at first and shrinks as the gap closes; own research grows steadily with compute. After the leader restarts, the gap widens again and diffusion rises with it.
The implication is that a paused leader stays ahead longer than its raw lead suggests, because diffusion fades as the follower approaches. A pause longer than 12.0 months lets the follower overtake.
The math
Step 3 · Paper, Section 3.2, Proposition 1
Catch-up takes longer than the raw lead
Diffusion is fast while the gap is wide and fades as the gap closes, so a halted leader stays ahead longer than the raw lead suggests. To see how, take a constant own research speed a and a gap G0 when the leader halts. The derivation takes three steps.
Step 1, the gap after a halt. The gap shrinks at the follower’s speed, a + δG. Own research closes a fixed amount each month; diffusion closes a fixed share of what is left.
Step 2, the catch-up date. This linear law makes the gap decay toward −a/δ at rate δ. It reaches zero after ln(1 + u)/δ months, where u = δG0/a compares diffusion with own research at the halt.
Step 3, the raw lead. A follower keeping pace covers the gap at speed a + δG0, so the raw lead is G0/(a + δG0) months. Dividing the catch-up time by the raw lead gives a function of u alone (Proposition 1):
The ratio rises from 1 without bound as u grows. So a catch-up time much longer than the raw lead implies a large diffusion ratio. The paper uses the 4–8-month raw lead and the roughly 12-month catch-up time proposed by the AI Futures Project, as uncertain scenario inputs. Raw leads of 4, 6 and 8 months give u ≈ 15.8, 3.9 and 1.4, so diffusion accounts for 94, 80 and 58 percent of the follower’s progress.
The gap after a halt
Catch-up time over raw lead
Left panel, the gap after a halt. The gap as a share of its size at the halt, against months, for the raw lead on the slider and a 12-month catch-up time. The gray dashed line is a follower that keeps its speed at the halt; it closes the gap at the raw lead. The blue curve is the actual follower. It starts at the same speed, but diffusion fades as the gap closes, so it needs the full 12 months.
Right panel, the ratio. Catch-up time over raw lead, (1 + u) ln(1 + u)/u, against the diffusion ratio u on a log scale. The gray dots are the three proposed raw leads; the orange dot is the slider’s.
The implication is that a catch-up time three times the raw lead needs diffusion to supply about 94 percent of the follower’s progress. The accelerating calibration of Step 2 gives 81 percent; it differs from these inversions because own progress changes during catch-up.
The math
Step 4 · Paper, Section 3.3, Proposition 2
A leader at a lower speed is overtaken; a faster one can still lose catch-up time
A leader that moves at a lower speed instead of halting keeps its lead only if that speed is high enough. Let it advance at a steady pace ν, measured against the follower’s own research speed a; a halt is ν = 0, the case of Step 3. The gap then changes at ν − a − δG: the leader adds ν, own research takes away a, and diffusion takes away δG. So the gap moves toward the level G* = (ν − a)/δ at which these forces balance.
If ν ≥ a, that level is not negative and the gap never reaches zero. If ν < a, it is negative, and the gap crosses zero in finite time (Proposition 2). Either way a shrinking gap shortens the catch-up time, T*(G) = ln(1 + δG/a)/δ, even before any overtaking.
In the accelerating model the same logic gives a floor on the leader’s research share. At equal capability the follower’s own research runs at 0.131 of the leader’s full-speed pace, so a constant share keeps the lead on a continuing horizon when it is at least 0.5 × 0.131² ≈ 0.86 percent. Keeping the lead only until the leader reaches TED-AI needs 0.85 percent; at that share the follower draws level at month 46.9, just as the leader completes its takeoff period. At 0.5 percent research the follower overtakes after about 20 months.
A leader at a lower speed: the gap
Accelerating model: a constant research share
Left panel, constant speed. The gap, as a share of its starting size, against months, for a leader that advances at the speed ν on the first slider, in multiples of the follower’s own research speed a (the follower is the 6-month case of Step 3). The dashed line is the resting level G*. Below ν = a the resting level is negative and the curve crosses zero: the follower overtakes. Above it the gap settles at a positive level, but below its start, so catch-up time still falls.
Right panel, the accelerating model. Capability of the leader (solid) and the follower (dashed) when the leader holds the research share on the second slider, with 25 percent safety. The dot marks overtaking when it happens.
The implication is that diffusion creates room for restraint but does not remove the constraint. The leader must keep enough speed relative to the follower’s own progress, and even a slowdown that keeps the lead shortens later catch-up time.
The math
Part 2Restraint by both: gains, the response window and the game
Step 5 · Paper, Section 4.1
Mutual restraint creates gains for both countries
Reciprocal restraint can benefit two rivals even when neither is willing to restrain itself alone. Under compliance, both countries stop capability research while continuing their safety programs. Their capability gap stays fixed, so both build safety knowledge without conceding progress to an advancing rival.
The paper uses two closure rules, which differ in what happens to the follower when the leader finishes. Under first-arrival closure, its baseline, the dangerous transition ends when the first frontier reaches TED-AI. From then on neither country adds catastrophe risk, so the follower’s unfinished takeoff period is not counted, for example because a defensive intervention follows the first arrival. This is an assumption of the paper, not a result. Under independent closure, each frontier carries hazard until its own arrival, so the follower’s takeoff period keeps adding risk after the leader has finished.
Both rules count only catastrophe risk. Neither says who gains or loses when the first frontier arrives. The countries’ gains, the value of surviving and of strategic control, enter only through the game of Steps 7 and 8.
If compliance holds, longer mutual pauses remove more risk under first-arrival closure. A 12-month mutual pause with safety research lowers the global catastrophe probability to 27 percent, against 80 percent when both countries race at full speed and 29 percent under the same unilateral pause. The probability falls to 7 percent at 24 months, 2.3 percent at 36 and 1.5 percent at 48, while a unilateral pause longer than 12.0 months lets the follower overtake. Closure matters a great deal. Under independent closure the 24-month mutual pause gives 79 percent, because the follower’s takeoff period still runs to completion. So choosing a duration needs the costs of delay as well as risk.
The game of Steps 7 and 8 asks whether each country’s national gain can support an agreement. Under its payoff assumptions, cooperation lowers catastrophe risk and development costs while keeping each country’s share of strategic control. These assumed gains differ from the simulated risk reductions in the figure, and Step 10 shows how a simulated cut enters the game. The gains can pay for verification, but an agreement also needs terms that deter a secret advance, and how much of this risk reduction is achievable depends on how long the response window stays open.
Global catastrophe probability against pause length
The panel. Global catastrophe probability against the length of a pause with safety research that starts at the Automated Coder. Blue is a compliant mutual pause and orange a unilateral leader pause, both under first-arrival closure; dashed blue is the mutual pause under independent closure. The gray vertical line is the pause length on the slider. The orange curve stops at 12.0 months: a longer unilateral pause lets the follower overtake.
The implication is that compliance is worth a great deal, because the mutual curve keeps falling well past the point where a unilateral pause must end. How much it is worth depends heavily on the closure rule, which the paper states as an assumption.
The math
Step 6 · Paper, Section 4.2, Lemma 1 (Appendix D)
The response window shrinks while the capability gap stays fixed
Suppose both countries halt capability research while continuing safety research, so the capability gap stays fixed under compliance. If the follower secretly advances, the leader stays paused until its alarm response finishes or the agreed pause ends. Let D be the detection delay, ℓ the implementation delay, and R the remaining pause. The leader’s exposure, the shorter of its alarm response time and the remaining pause, is min{D + ℓ, R} months.
Until the leader restarts, the gap shrinks along the halt path of Step 3 and reaches zero at the follower’s catch-up time. So the leader keeps a positive lead when it restarts before the catch-up time, and keeps it afterwards if it then moves at least as fast as the follower’s own research (Lemma 1 in the paper’s Appendix D). This is a physical result; it says nothing yet about whether either country prefers to comply.
Compute keeps growing during a compliant pause, so a fixed capability gap does not preserve a fixed catch-up time. A secret program launched later starts from the same frozen states with more research capacity, so it reaches the frozen frontier sooner. The calibrated catch-up time falls from 12 months at the start to 8.61 months after 12 months of compliance and to 5.74 months after 24 months. Less the implementation delay, this catch-up time is the follower’s response window of Step 9, the time the leader has to act on a secret advance.
A fixed response standard fails as catch-up time falls. A secret advance at month 12 of a 24-month pause permits catch-up after 8.61 months. Detection ten months later meets the one-year standard that suits the start, but comes after overtaking. So a longer agreement needs a date-dependent response standard, an earlier scheduled restart, or other safeguards.
Catch-up time by the date of a secret advance
One secret advance
Left panel, the clock. The blue curve is the follower’s catch-up time after a secret advance at each date of a compliant 24-month pause, against that date. It falls because compute keeps growing while both frontiers stay frozen. The orange dashed line is the leader’s alarm response time on the slider and the gray dotted line the remaining pause. The leader is still ahead when it restarts on the dates where the lower of the two lies below the blue curve.
Right panel, one secret advance. Capability after an advance at the date on the slider. The flat line is the leader, frozen at the Automated Coder; the dashed curve is the follower, starting from its compliant position. The blue line marks when the follower reaches the frozen frontier, the orange line when the leader restarts.
The implication is that a visible capability gap is a poor measure of the response window. The same gap leaves a catch-up time of 12 months at the start of the pause and 5.74 months after two years.
The math
Step 7 · Paper, Section 4.3, Proposition 3
What each country gets, and why restarting is credible
The agreement game covers one review block, the round of decisions at one public review date. The leader proposes a monitoring intensity, the follower accepts or refuses, and each government then chooses whether to pause or secretly advance. Each country has four branch payoffs: cooperative value C, breakout T (an advance the rival does not answer), contested race R, and waiting value S while the rival advances. The game keeps the prisoners’ dilemma ordering T > C > R > S. Each country prefers mutual restraint to a mutual race, but without monitoring advancing is dominant, so without an agreement both advance.
The branch payoffs come from a few primitives: Y, the value of survival apart from strategic control; Z, the additional value of control, shared equally in a contested race (χ = 1/2); and costs, c for active development and d for pausing. Each payoff is surviving value times one minus that branch’s catastrophe probability, less its cost. The probability is p0 when no country advances, p1 when one does and p2 when both do:
Cooperation keeps the contested shares of control while lowering catastrophe risk and saving development costs after paying for delay. The gain from cooperation, C − R, limits what each country will pay for verification.
The example sets Y = 1, Z = 4, costs c = 0.8 and d = 0.2, and probabilities 0.1, 0.2 and 0.4. So C = (1 − 0.1)(1 + 2) − 0.2 = 2.5, R = (1 − 0.4)(1 + 2) − 0.8 = 1.0, and the same formulas give T = 3.2 and S = 0.6. The gain from cooperation before monitoring is 0.3 × 3 + 0.6 = 1.5 for each country, of which lower catastrophe risk supplies 0.9 and the saved development cost 0.6. These national values are chosen inputs, not estimates.
After a timely finding the monitoring cost is sunk, so restarting gives R and waiting gives S. Restarting is credible when the surviving strategic stake exceeds the extra development cost plus the extra risk to non-strategic value (Proposition 3, Credible Response):
In the example the left side is 1.2 and the right side 0.8, so restarting pays. A costly response is carried out because its strategic benefit exceeds its cost. Credibility is the R > S part of the payoff ordering. If the inequality fails, the threat to restart cannot support cooperation.
Branch payoffs
Credible response: what restarting recovers and what it costs
Top panel, branch payoffs. The four payoffs at the sliders’ values; the gray text gives the paper’s. The cooperative value C sits between the breakout payoff T and the contested race R, so each country prefers mutual restraint to a mutual race but is tempted to advance alone.
Bottom panel, credible response. The upper bar is the strategic value that restarting recovers, (1 − p2)χZ. The lower bar is what restarting costs: the extra development cost c − d (orange) plus the extra expected loss of non-strategic value, (p2 − p1)Y (gray). Restarting is credible while the upper bar is longer.
The implication is that a finding deters only if the payoffs make the response worth carrying out. A lower strategic value or a riskier race can make the threat to restart empty, and Step 10 shows a case where the scenario’s own risks do.
The math
Step 8 · Paper, Section 4.3, Proposition 4
Monitoring must deter a breakout, and both countries must be willing to pay for it
Cooperation is supportable when the monitoring needed to deter a breakout costs no more than either country gains from cooperating. Monitoring delivers truthful public findings at rate m per month. A secret advance is caught within the response window w with probability 1 − e−mw. Caught, the country gets the contested payoff R; uncaught, the breakout payoff T. Three steps give the condition for cooperation.
Step 1, the gamble: if the rival complies, a secret advance pays R or T with those probabilities. Step 2, deterrence: compliance must be worth at least this gamble, which needs m ≥ ξ/w, where ξ = ln[(T − R)/(C − R)] is the breakout temptation, how tempting a breakout is relative to cooperation. Step 3, participation: each country pays ηm for monitoring and joins only if cooperation net of that cost beats the race, so m ≤ (C − R)/η. Both countries must pass both tests (Proposition 4):
The lower bound prices deterrence, because a shorter window needs more findings per month to prevent a profitable head start. The upper bound prices consent, because verification cannot cost more than either country’s gain from cooperation. Among protocols that support cooperation, both countries prefer the least costly deterrent intensity, mmin.
In the example ξ = ln(2.2/1.5) = 0.383 and the monitoring cost is η = 27 per unit, so mmax = 1.5/27 ≈ 0.05556 findings per month. At the start of the pause the follower’s window is 12 months, so deterrence needs m = 0.383/12 ≈ 0.0319, and cooperation is supportable. Mutual racing remains an equilibrium too, so the result shows that cooperation can be supported, not that it will be chosen.
Deterrence does not need near-certain detection. At the threshold a secret advance is caught within the window with probability 1 − e−ξ ≈ 32 percent, and the mean time to a finding is about 31 months. Deterrence is an incentive test, and at the threshold a breakout still succeeds with probability 68 percent. Holding that probability below a tolerated level β needs m ≥ ln(1/β)/w, which is more than deterrence needs whenever β is below 68 percent.
The two bounds meet at the critical window ξ/mmax ≈ 6.89 months. A shorter window needs more monitoring than both countries will pay for.
Payoffs against monitoring intensity
Monitoring needed against the response window
Left panel, deterrence and participation. Payoffs against the monitoring intensity m, in findings per month. The blue curve is the expected payoff of a secret advance, which falls as findings get likelier; the green line is cooperation C. Deterrence starts where blue falls below green, at mmin. The orange line is cooperation net of its monitoring cost, C − ηm, and the gray line the race R; participation ends where orange falls below gray, at mmax. The shaded interval supports cooperation.
Right panel, the window. The monitoring needed to deter, ξ/w, against the response window w (blue), beside the most both countries will pay (orange dashed). The two meet at the critical window; the blue dot is the slider’s window.
The implication is that faster detection can help and still be too costly to secure voluntary entry. Once the window is shorter than about 6.89 months, no monitoring level both deters a breakout and keeps both countries in.
The math
Part 3The clock: how long cooperation lasts, and what extends it
Step 9 · Paper, Section 4.4, Corollary 1, Figure 1, Tables 2 and 3
Agreements have a clock
The clock follows from tying the game to the physical response window. For a secret follower advance at pause date d, the window is the catch-up time of Step 6 less the implementation delay ℓ, the time needed to restart: wB(d) = max{Tcatch(d) − ℓ, 0}. The model of the follower gives no window for a secret advance by the leader, so the leader’s window is set separately, at 12 months in the example. When it binds, it decides whether cooperation is supportable even if the follower’s window is long. The proposed pause lasts 36 months.
The window assumes that being overtaken at the frozen frontier costs the strategic position, and the physical model does not supply that loss by itself. In the calibration, a follower that advances from the start of the pause while the leader waits the full 36 months overtakes the frozen frontier, yet the leader still reaches TED-AI first, at month 39.6. Once ahead, the follower loses diffusion, and its own research runs at 0.131 of the leader’s pace at the same capability and date. So the game describes stakes in which a temporary capability lead is valuable in itself, for example because it can be used before either side reaches TED-AI. If decisiveness meant arriving first, a follower breakout would not be decisive in this calibration.
Three steps give the corollary. The follower’s window must be at least the critical window ξ/mmax. So the catch-up time must be at least ℓ plus the critical window. During a compliant pause the catch-up time falls toward zero, so that floor fails at late enough dates, even though the visible capability gap has not narrowed (Corollary 1). The corollary holds national values, monitoring costs, the implementation delay and the leader’s window fixed.
In the example, with fixed national values, monitoring costs and a leader window of 12 months, the deterrence and participation bounds meet at a critical window of 6.89 months. The follower’s catch-up time passes that value at month 18.9, so voluntary cooperation is supportable in review blocks up to that month. At month 18 the least costly deterrent intensity is 0.0539 and net cooperation pays 1.044, above the outside payoff R = 1. At month 24 the required intensity is 0.0667 and net cooperation pays only 0.699. Deterrence is still possible there, but both countries refuse its cost in this example.
The clock comes from the physics; the date comes from chosen inputs. One or two months of implementation delay move the last date to 14.8 and 10.9. Halving the monitoring cost keeps cooperation supportable through month 31.8. After that date the remaining pause is shorter than the catch-up time, so the scheduled restart comes first and the game no longer applies. Doubling the cost rules cooperation out from the start, because the required monitoring then exceeds what either country will pay even before the window shrinks. The leader’s own window fails once η exceeds 47.0.
The paper’s online supplement shows how the date depends on the inputs. Varied one at a time, compute growth and the monitoring cost move it most. With compute doubling every 3 months instead of 6, the last supportable review falls from month 18.9 to 10.9, while the length of the takeoff period hardly moves it.
Section 4.5 of the paper makes the countries more alike, in compute and in capability, and also varies the stakes and the leader’s window. Step 11 takes these cases one at a time (Table 4).
When to sign is a question the paper does not compute: its clock counts months from the start of a pause, and its examples all start the pause at the same point of the takeoff period. A check with the paper’s model, not reported in the paper, starts the pause later instead, after both countries have raced. The window at signing is only a little shorter, because racing first also widens the follower’s gap, which offsets most of the extra compute. The dates do not all move one way either: a later pause makes an unanswered breakout far more dangerous, so the threat to restart becomes credible sooner. The stronger reason to sign early is what the pause protects. A later agreement comes after more of the takeoff period has been raced, and pausing at a higher capability carries more risk while both countries wait, so it buys less safety. Signed much later, the paper’s assumed ranking of outcomes starts to fail, so those results are rough.
The clock: catch-up time falls
Monitoring needed against monitoring affordable
Table 3 at the sliders’ settings: response windows and cooperation in the solved example
Left panel, the clock. The blue curve is the follower’s catch-up time after a secret advance at each pause date. The orange dashed line is the catch-up time that the deterrence and participation bounds need, ℓ + ξ/mmax, at the monitoring cost and delay on the sliders; the thin gray lines give the paper’s other two costs, 13.5 and 54. The gray dotted line is the remaining 36-month pause. Cooperation is supportable on the shaded dates, where the blue curve lies above the orange line and the game applies.
Right panel, monitoring. The intensity needed to deter, mmin, against the most both countries will pay, mmax. Needed monitoring rises as the window shrinks, and the last supportable date is where it crosses mmax.
The implication is that an agreement’s duration cannot be read off a capability lead. It depends on what monitoring costs, how fast a response can be implemented, and how much cooperation is worth to each side. For Plan A, whose pause is to last five years, they decide whether the clock runs out first.
The math
Step 10 · Paper, Sections 4.6 and 4.7, and the supplement’s Tables S2 and S3
Matched safety research lengthens the clock, and at the baseline hazard level the scenario’s own risks add an opening date
An agreement can lengthen its own clock by committing both countries to the same safety research and verifying that commitment with the pause. Under compliance the follower’s safety knowledge then matches the leader’s. The term binds only while both comply, so only the cooperative branch becomes safer and the catch-up path, the physical clock, is unchanged.
With matched safety research, a compliant 36-month mutual pause lowers the global catastrophe probability from 2.3 to 1.6 percent under first-arrival closure, a cut of 32 percent. The cut is 22 percent at 12 months and 39 percent at 24. In the game this lowers the cooperative-branch risk p0, which raises C, lowers the breakout temptation ξ and raises the most both countries will pay for monitoring. The critical window shortens, so cooperation stays supportable at later review dates, when catch-up time is shorter.
As an illustration, apply the simulated 32 percent cut to the game’s chosen p0 = 0.1, which gives p0 = 0.068. The critical window falls from 6.89 to 5.44 months, and cooperation stays supportable through month 25.5 instead of month 18.9. Each percentage point cut in p0 adds 2.0–2.1 months. The gain survives while the monitoring cost per unit stays below 34.2, 27 percent above the baseline of 27.
At the baseline hazard level, the scenario’s own risks add an opening date to the closing date. The game’s branch risks of 10, 20 and 40 percent are milder than the scenario’s. Along the model’s own paths a race from the start carries 80 percent, cooperation through the 36-month pause 2.3 percent, and an unanswered follower breakout 14 percent. The paper re-solves the game at each review date with these risks and keeps the national values, costs, control shares and the leader’s 12-month window of the solved example (Section 4.7 and the online supplement’s Table S2).
At the start a race is so dangerous that the leader would rather keep waiting than answer a breakout, so the threat to restart fails the test of Step 7 and no monitoring can deter. Safety research during the pause makes a race survivable. The threat becomes credible once a race carries less than about 50 percent, which takes 6.1 months, and the clock of Step 9 then applies. The table gives the review dates at which cooperation is supportable, at three monitoring costs.
So a pause must first make a retaliatory race survivable before monitoring can deter a breakout. The solved example has no opening date, because its perceived stakes make a race survivable from the start. These dates come from chosen inputs.
At a full-speed probability of 25 percent, the lower end of the paper’s range, a race from the start carries 28 percent, below the 48 percent at which restarting beats waiting, so there is no opening date. But cooperation is worth less while monitoring costs the same. The largest supportable monitoring cost at signing is 27.5 per unit, barely above 27, so at the baseline cost the agreement is supportable only through month 0.1, and at half the cost through month 5.3 (the online supplement’s Table S3). Below the paper’s range the dates close altogether: at the baseline cost, cooperation is supportable at signing only if the full-speed probability is at least 24.6 percent, and at half the cost only if it is at least 11 percent.
Review dates at which cooperation is supportable, by monitoring cost
| Monitoring cost per unit | Solved example: branch risks 10, 20, 40% (months) | The scenario’s own risks (months) |
|---|---|---|
| 13.5, half the baseline | 0 to 31.8 | 6.1 to 14.6 |
| 27, the baseline | 0 to 18.9 | 6.1 to 10.4 |
| 54, double the baseline | never | 6.1 to 6.5 |
Compliant mutual pauses: own safety against matched safety
The clock with a safer cooperative branch
Left panel, the risk cut. Global catastrophe probability of a compliant mutual pause against its length, with the follower’s own safety program (blue) and with safety research matched to the leader’s (green), under first-arrival closure. The labels give the cut at 12, 24, 36 and 48 months.
Right panel, the clock. The follower’s catch-up time after a secret advance (blue) against the catch-up time that cooperation needs, at the paper’s p0 = 0.1 (gray dashed) and at the p0 and monitoring cost on the sliders (orange dashed). Each line meets the blue curve at its last supportable date.
The implication is that an agreement that makes cooperation safer also makes cooperation last longer, as long as verifying the safety program does not cost too much. If the inspections that verify a pause also cover the safety program, the added cost can be small.
The math
Step 11 · Paper, Section 4.5, Table 4, Appendix B
When the rivals are more alike, the clock still runs and the date moves
So far the follower has less compute and trails by six months. Section 4.5 of the paper asks what changes when the two countries are more alike. “More alike” can mean closer compute or equal capability, and both shorten the time a country has to respond. Two assumptions of the main example bear on the same question and are relaxed too: stakes that favor the leader raise the window that cooperation needs, and the leader’s own window can shrink.
Take closer compute first, keeping the six-month lag. As diffusion’s share of the follower’s initial progress falls from 81 to 61 and 41 percent, its catch-up time at the start falls from 12.0 to 9.1 and 7.6 months. The last supportable review date falls from month 18.9 to month 8.0 and month 2.5. A follower with no diffusion catches up in 6.0 months, less than the critical window of 6.89 months, so cooperation is never supportable.
Now take two countries equal in capability as well as compute. The paper’s window, the time to reach a frozen rival, is then zero, because a secret advance puts the deviator ahead at once. Between equals a natural measure is the time a secret full-speed program needs to open a decisive lead of Δ months. If a lead of 12 months is decisive, cooperation is supportable until month 12.4 instead of month 18.9. Matching the baseline date takes Δ = 15.5 months, and a lead of six months is never enough.
Stakes matter too. The game gives each country half of control in a contested race. If the leader would win 55 percent, the follower’s breakout temptation rises, the critical window grows from 6.89 to 8.94 months, and the last supportable date falls from month 18.9 to month 10.8. A larger share for the follower helps: if the leader would win 45 percent, the date is month 25.7.
The main example also fixes the leader’s window at 12 months. If the leader’s window shrinks as compute grows, as the follower’s does, it binds first and the last supportable review moves to month 12.4, the same date as for equal countries with Δ = 12. The baseline date holds if a leader breakout must cover at least 15.5 months of full-speed progress.
In every case with a positive window, the window that binds is shorter at month 24 than at the start, so the clock survives. What moves is the date. A lead buys time to respond, because the follower must first erase it, so rivals that are more alike need faster detection and response to keep cooperation supportable. The monitoring needed to deter is inversely proportional to the window: a follower with the leader’s compute, still six months behind, catches up in 4.4 months at the start of the pause, so deterring it takes about 2.7 times as much monitoring as in the main example. Cheaper monitoring or a larger gain from cooperation, such as the matched safety of Step 10, can offset a shorter window.
Selected rows of the paper’s Table 4
| Case | Window at the start (months) | Critical window (months) | Last supportable review (month) |
|---|---|---|---|
| Baseline: six-month lag, diffusion share 81% | 12.0 | 6.89 | 18.9 |
| Closer compute, diffusion share 61% | 9.1 | 6.89 | 8.0 |
| Closer compute, diffusion share 41% | 7.6 | 6.89 | 2.5 |
| Closer compute, no diffusion | 6.0 | 6.89 | never |
| Equal capability, a lead of 12 months is decisive | 12.0 | 6.89 | 12.4 |
| Equal capability, a lead of 18 months is decisive | 18.0 | 6.89 | 22.8 |
| Leader’s share of control 55% | 12.0 | 8.94 | 10.8 |
| Leader’s share of control 45% | 12.0 | 5.39 | 25.7 |
Part 4Widening the lead and ranking domestic rules
Step 12 · Paper, Section 5.1, Proposition 5, Table 5
Widening the lead: own research and diffusion
The response window is the follower’s catch-up time, so it helps to know what sets it. In the constant-speed model of Step 3 two forces matter, and they enter differently. Own research enters only through u, so a unit rise in ln a shortens the catch-up time by the raw lead, in months. A unit rise in ln δ shortens it by T* less the raw lead, because the higher ratio u gives back part of the gain (Proposition 5):
With a 12-month catch-up time and a 6-month raw lead, the two have the same size: a 10 percent change in own research and a 10 percent change in diffusion each move the catch-up time by about 0.6 months.
The full model gives the same picture. A slower follower, in own research or in diffusion, means a longer catch-up time, so the unilateral pause can run longer and the response window stays open, which lowers the global catastrophe probability.
Step 13 · Paper, Section 5.2, Table 6, Figure 2
A domestic ranking of instruments need not survive foreign exposure
Domestic pacing rules can raise global risk, because they let the follower pull alongside. The paper evaluates the pacing instruments of its companion papers when the follower races in parallel. A compute floor fixes the research and safety shares for the whole takeoff period and gives the rest to customers. A safety floor raises the safety share at the customers’ expense. A risk ceiling slows development just enough to hold the hazard at the ceiling and moves the freed compute into safety. An R&D-model lag makes research use the model from a fixed number of months earlier. A pause with all compute to customers sets research and safety to zero for its duration; a pause with safety moves the research share into safety instead.
Global probability splits into the part caused by each frontier. The cause-specific incidence of a frontier weights its hazard by the chance that neither frontier has failed yet, so the two parts add up to the global probability.
On the leader’s own frontier, with the follower ignored, the risk ceiling and the R&D-model lag lower the catastrophe probability from 75 to 65 and 68 percent. With the follower racing they raise the global probability instead, from 80 percent to 97 and 87 percent. They slow development late in the takeoff period, when capability is high, so the follower’s unpaced frontier pulls alongside the leader. The follower’s part of global probability rises from 13 to 47 and 36 percent.
Compute floors spread delay over the takeoff period. The best tested floor, 5 percent research with 25 percent safety, gives 41 percent. Choosing a common research share to maximize discounted global survival while keeping the lead selects 2.5 percent research and gives 57 percent. A coordinated domestic industry that ignored the rival would choose 0.06 percent research, but that share lets the follower overtake after about 13.6 months.
Global catastrophe probability by instrument
Compute floors: global probability against the research share
Table 6: catastrophe probability with the follower racing
Top panel, instruments. Global catastrophe probability for each instrument when the follower races. Under first-arrival closure each bar splits into the part caused by the leader’s frontier (blue) and the follower’s (green); under independent closure it shows the global probability alone. The settings are the equal-lead ones of the companion papers.
Bottom panel, compute floors. Global probability (black) and its two parts against the research share of a compute floor with 25 percent safety, on a log scale, under first-arrival closure. Lower shares cut the leader’s part but raise the follower’s. The dotted line is the lead boundary, below which the follower overtakes.
The implication is that an instrument’s domestic ranking need not survive foreign exposure. Instruments that slow development late let the follower pull alongside at high capability; floors that spread delay do better. These are scenario comparisons, not rankings of instruments in general.
The math
Step 14 · Paper, Section 5.3, Table 7
When foreign risk runs to completion, the follower’s safety matters more than pacing at home
Closure matters because the leader’s arrival may not end the follower’s danger. Under independent closure the follower’s takeoff period runs to completion, and global probability stays high: 99 percent without pacing, and 90 percent with the equal-delay floor of 7.7 percent research.
Safety on the follower’s side then does more of the work. Equal own safety for the follower lowers the floor outcome to 51 percent, and adopting half of the leader’s safety progress lowers it to 35 percent. So international safety diffusion is an important channel, though not indispensable in every specification; an agreement can also write it into its terms (Step 10). Closure is especially consequential for agreements, as the compliant 24-month mutual pause of Step 5 shows.
Global catastrophe probability under independent closure
Table 7: global catastrophe probability when the follower’s takeoff period runs to completion
The panel. Global catastrophe probability under independent closure, for the full-speed race, the longest leader pause with safety research and the equal-delay compute floor, at the follower safety settings on the sliders. The gray text gives the paper’s baseline, a follower with half the leader’s safety yield that adopts none of the leader’s progress.
The implication is that once foreign risk runs to completion, pacing at home buys much less, and the follower’s own safety knowledge matters more. The table gives the paper’s three columns.
The math
Part 5What the model leaves out
Step 15 · Paper, Section 6
What the model leaves out
The results rest on assumed national payoffs, a specified follower path, a stipulated closure rule and an illustrative calibration. Four limits qualify them.
- The follower’s research and safety program are specified paths, not an equilibrium. The game solves one review block for stipulated national values and costs, with terminal risk lotteries, committed programs and truthful public findings. The physical model supplies one response window, the follower’s; the leader’s window is assumed. The contested share of control is an input, and governments are risk neutral over the terminal lotteries. The game does not solve repeated renewals, renegotiation, or a joint equilibrium of labs and governments.
- Compute is an exogenous growing stock. Pacing can change investment incentives, and endogenous investment can change catch-up time during a treaty.
- Safety learning does not scale with total compute, and first-arrival closure assumes that the first arrival ends the other frontier’s risk. Independent closure often gives much higher probabilities.
- The calibration uses uncertain illustrative targets and model-implied diffusion parameters, not identified causal estimates.
These omissions cut both ways. Repeated renewal could make agreement easier, because a longer shadow of the future supports cooperation, and better verification technology would help. False findings, private monitoring and endogenous investment could make it harder. How labs within one country race, and why a common rule slows them, is the subject of the companion paper Prizes, Shared Hazard, and the Speed of Frontier AI.