Beta test results
First beta test, 4 and 5 November 2026. The numbers are from the 38,579 runs stored by 2026-11-05 12:38 UTC. All results
The first beta test sent 40,000 tasks to volunteers: 100 scenarios, each run with 400 different random seeds. A scenario is one combination of the eleven settings the project varies, such as how much is launched, how reliably old satellites are taken out of orbit, and how often a small piece of debris disables a satellite it hits. Each run follows one possible future of low Earth orbit from October 2026 to 2226. Two volunteers' computers ran every task, and a result counted only when both got exactly the same answer.
The beta had two jobs: to check that the whole system works on volunteers' computers, and to check how we score each run. This page is a shortened version of our analysis.
What the first beta found
- We stop a run when its tracked objects (10 cm and larger) pass 2 million, about 60 times today's 35,865, and count it as a runaway cascade. In 73 of the 100 scenarios every run got there. Half of all the runs that got there did so by 2052, 5% of them by 2038 and 95% by 2079.
- Within one scenario, the year of the cascade usually varies by only a few years from run to run. The settings move it by decades.
- The launch traffic and four other settings decide most of it. Every constellation built to its filed size brings the cascade 63 years sooner than the traffic that is licensed or being built. Doubling the traffic volume brings it 66 years sooner than halving it. Leaving out the Guowang constellation's two layers near 1,150 km delays it by 70 years. If small debris disables a satellite in a quarter or more of its hits, the cascade comes 46 to 85 years sooner than if it never does: a disabled satellite can no longer steer clear of collisions. Removing old satellites at the end of their missions 95% of the time or better, instead of 90%, delays it by 32 to 61 years.
- Removing up to 50 large pieces of debris a year, the rule for how fast old satellites must come down, the amount of 1 to 10 cm debris, solar activity and rocket-body explosions made no difference we could detect with 100 scenarios.
- In 86% of the runs that reached 2 million objects, most of the debris ended up between 1,050 and 1,200 km, the altitudes of the Guowang constellation.
Effect of each setting
Each dot is how many years an option adds before the cascade (right of zero) or takes away (left), compared with the first option of the same setting, with the other ten settings held fixed. The years count how long a run stays below 2 million objects, a run that never gets there counting as the full 200. The bar is the 95% range: a bar that crosses zero is within the noise of a 100-scenario sample. Hover over a row for its numbers.
In a few scenarios, the same settings give different futures
In 12 of the 100 scenarios the runs did not all end the same way. Within a scenario every setting is the same, so the runs differ only by chance: which objects pass close to each other, which of them collide, and when. In the charts below, the bar at the top of each scenario shows how its runs ended, and underneath are the years in which its cascading runs passed 2 million objects.
To see what separates the two kinds of future, we repeated 26 runs of one of these scenarios on our own computers and kept their full year-by-year history. The simulation is exactly reproducible, so these are the same runs the volunteers computed: each repeat ended in exactly the volunteers' result.
Runs of scenario 00021230210, repeated with their full history
Tracked objects per year in each run (log scale). The dashed lines mark 2 million objects and today's count. Hover to follow one run.
Catastrophic collisions per decade, the same runs
Collisions that shatter both objects, per decade of each run (square-root scale, so the small counts of the quiet runs stay visible). A decade a run covered only in part, the first one and the one in which a run reached 2 million objects, is scaled to a full decade.
For their first few decades the runs look alike: each one grows with the launches at the same pace and has about the same number of collisions. They separate in the 2060s. In the runs that cascade, collisions multiply while launches continue, and the count climbs to 2 million. In the others the count peaks when launches stop in 2076, and atmospheric drag then brings it back below today's level. One run, seed 18, came close: it peaked at 1,980,962 objects in 2110 and still had 1,376,546 in 2226. The original test still counted it as no cascade, because its debris was shrinking during the 50 years the test looks at.
400 runs per scenario are enough to place these on the map
How many runs does a scenario need? For each of the 12 scenarios with mixed outcomes, the chart follows the share of runs that ended in a cascade as the runs come in, one seed at a time. Where no run reached 2 million objects, it follows the share with a flagged shell instead. The shaded band is the 95% range at each point, and the dashed line is the final value.
Every share settles inside its final range after between 42 and 227 runs, so 400 runs are enough to place each scenario on the map. About 1 scenario in 8 (12 of 100) had mixed outcomes. The project plan sets aside a second pass of 2,000 or more runs for scenarios like these, where a sharper answer matters, and had assumed about 1 in 10.
Why the runaway test for completed runs is being replaced
For a run that lasts the full 200 years, the original test looked at each 50 km altitude shell after launches stop in 2076. A shell was flagged as running away if its count of tracked fragments was still rising at the end, had grown by more than 20% since 2086, and held at least 50 fragments.
How often each shell was flagged in completed runs
Of the 8,613 runs that went the full 200 years, the share in which the test flagged each altitude shell.
Flagged shells in repeated runs
Tracked fragments per year in each shell the test flagged, in completed runs repeated with their full history. Shaded: the years the test looks at. Dashed: its floor of 50 fragments.
In the beta, that test and the overall amount of debris disagreed in both directions. Of the 1,905 completed runs with no shell flagged, 1,657 still ended with more tracked objects than today, 40 of them with more than a million. Those are cascades that peaked just below 2 million and were shrinking during the test's window. And of the 6,708 completed runs with a flagged shell, 3,007 ended with fewer objects than today. Most of their flags were high up, above 1,200 km, where a few hundred fragments build up slowly because almost nothing decays at that altitude. Collisions in those runs had long since stopped. A slowly growing pile of old fragments is not a cascade, so the test had to change.
What the beta changed, and why
A run now counts as a cascade in either of two ways:
- It reaches 2 million tracked objects, in any year. This was already the rule. Launches alone cannot get there: the runs that did launched a median of 225,427 objects (at most 812,799) and made a median of 2.4 million collision fragments. 93% of them got there before launches stop in 2076.
- It runs the full 200 years and catastrophic collisions keep happening after launches stop. Debris that keeps colliding with no new launches is sustaining itself, which is the runaway Donald Kessler described in 1978. We will set the cutoff, how many collisions per decade still count, from measured runs.
The year of a cascade will be the year collisions take off, not the year a run reaches 2 million objects. A slow cascade can reach that limit a century after it starts, or never. A scenario's probability of a cascade is the share of its runs that cascade by either rule, and the same rules applied to each altitude shell give the probability map the project set out to make.
Collisions after launches stop tell the cases apart
Catastrophic collisions per decade in five of the repeated runs. Where collisions die out after the launch stop, there is no cascade, whatever the old test said. Where they go on with no new launches, there is. A dash: the run had already stopped at 2 million objects.
| Run | Label under the old test | 2060s | 2070s | 2080s | 2100s | 2150s | 2210s | All after 2100 |
|---|---|---|---|---|---|---|---|---|
| 00021230210, seed 6 | Reached 2 million | 749 | 1,837 | – | – | – | – | – |
| 00021230210, seed 18 | Completed, none flagged | 129 | 385 | 1,235 | 2,274 | 1,204 | 726 | 16,039 |
| 00021230210, seed 1 | Completed, none flagged | 129 | 136 | 94 | 4 | 0 | 0 | 6 |
| 00021230210, seed 72 | Completed, a shell flagged | 148 | 139 | 76 | 4 | 0 | 0 | 5 |
| 20020231210, seed 63 | Completed, a shell flagged | 2 | 0 | 0 | 0 | 0 | 0 | 1 |
Seed 18 counted as "none flagged", yet its collisions never stopped. The flagged runs had almost no collisions after the launch stop.
The summary each computer sends back is the only record we keep of a run, so anything the analysis will need for every run has to be in it. From app version 1.01 each result also carries:
- the catastrophic collisions in each decade, for each altitude shell: the new test, the year of the cascade, and which altitudes ran away. They are kept as counts, so the rule can be checked and changed after the main run;
- the largest number of objects the run reached and when, overall and in each shell. After 150 years without launches, the count at the end can be far below the peak;
- the first 20 catastrophic collisions, and how many working satellites, dead satellites and rocket bodies each shell held when launches stopped: the conditions under which a crowded shell runs away.
We also now keep a record of every copy of every task: which kind of computer made it, and whether it agreed. That is the evidence that the results are the same on any hardware.
Version 1.01 is being tested in a second beta, which started on 5 November 2026: the 27 scenarios of the first beta that did not cascade in every run, each with the same 400 seeds, 10,800 tasks in all. The same seed gives the same run, so every result of the second beta must end in exactly the state its first-beta twin did. Its collision counts will show where to set the cutoff before the main run begins.
About these numbers
- This is a snapshot: 38,579 of the 40,000 runs had come back when we made it. The runs still out were mostly slow ones, and in the scenarios with mixed outcomes the slow runs are mostly ones that do not cascade, so for those scenarios the cascade shares here are a little high.
- The 100 scenarios are a random sample of the 302,400 in the sweep, so no two differ in only one setting. Questions such as how many removals a year would make up for less reliable disposal need the full sweep.
- In every scenario, launches follow the traffic plans until 2076 whatever happens, and no operator reacts to a worsening environment. A cascade here means what happens if nothing is done.
- 29 runs were repeated on project computers to get their full history, and each ended in exactly the volunteers' result.