One play-through
Every event in the tree has a written test for whether it happened, a probability, and a list of other
events it depends on. A play-through walks the events in order, earliest dependencies first. If everything
an event depends on has happened, the event happens with its own probability, like a weighted coin flip;
otherwise it can't. Events that have already happened in the real world are fixed. My own decisions are
never flipped: each fork is set to what I currently plan to do. A milestone such as "human industry reaches beyond Earth and Mars" or
"I work off Earth on rotation" is reached when everything on its list happened.
Do that 20,000 times per scenario and count. Under current trends, "works off earth on rotation"
is reached in 93 of 20,000 play-throughs, so the headline number is 0.5%.
Every probability in the tree is a conditional one: the chance the event happens before the deadline,
assuming the things it depends on have happened. That is a question a person can actually answer
("if launch is cheap, how likely is factory-scale manufacturing in orbit by 2070?"), and walking the tree
in dependency order is what keeps the chain honest.
The code that does this is public:
scripts/compute.py.
Five guesses at the deadline
The deadline is how long I live, and that is not a fixed number, because longevity research keeps moving
it. How soon medicine might start repairing the body faster than it ages is a separate question, which I
track at Outrunning Ageing; the five guesses here are
mine, and nothing on that site changes them. So the whole tree is run once per scenario, from current trends to an open-ended window, and no event's
probability is allowed to fall as the window gets longer. The year dial on the home page switches between
them.
| Key | Scenario | Window closes |
|---|
baseline | Baseline, no extension beyond current trends | 2071 ± 8 |
moderate | Moderate extension | 2080 |
strong | Strong extension | 2095 |
radical | Radical extension, a 150-year life | 2136 |
open | Escape velocity, window open | no deadline |
The five are reported side by side and never combined into one number. The spread across them is itself
a finding: it shows how much of the question is about spaceflight and how much is about medicine.
Why the events are not rolled on their own
If every event were rolled independently, the tree would say the far side of Mars is all but out of reach,
even though each event on the way there is individually likely. That is what happens when fifteen fairly
likely things are all required and are multiplied as if they had nothing to do with each other. But they
do: cheap launch, steady money and a working off-Earth economy are causes shared across most of the tree,
and a world where one of them comes good is a world where the others tend to.
So before each play-through the calculation draws one number for how favourable the world turned out, and
nudges every open event's odds by that amount before rolling it. In some play-throughs everything is easier
and in some everything is harder, which is what shared causes do. The size of the nudge is the one free
setting in the method. It is set so that a typical good world roughly doubles the odds of a middling event
and a typical bad world roughly halves them, about the swing the last two decades of spaceflight have shown
between optimistic and pessimistic readings of the same facts. Every run also prints the independent-roll
result beside the real one, so anyone can see exactly how much that setting is worth.
The thing needed, not one way of getting it
When an event needs something that more than one route could supply, its dependency names the thing, not
the route. Mining the Moon at scale needs power through the two-week lunar night, so the step it depends on
is "power through the lunar night at a surface site", not "a fission reactor on the Moon". The reactor
stays in the tree as one route, with its own test and probability, and the power step happens once any
route has. A link that named only one route would multiply the child by that route's chance, when the true
chance is that of any route coming good, and the tree would always lean low.
What the numbers assume about timing
Each number is the chance an event happens before the deadline, given that what it depends on has
happened. It says nothing about when. A parent finished in 2035 leaves its child decades; one finished in
December 2070 leaves it weeks. The estimates are read as if parents finish at about the time they typically
would, in the worlds where they finish at all. Where that is wrong, it is most likely wrong in one
direction: events at the top of long chains, under the shorter deadlines, probably lean a little high. A
version that estimates when each event lands, rather than one probability per deadline, is on the list for
later.
What is reported
Each run reports, for each of the five deadlines: the chance of each of the four stages human industry could
reach (an outpost, a settlement, the slow expanse, the full expanse); the chance of each of the four ways I
could be part of it (working for the industry from Earth, flying once, working off Earth on rotation,
living off Earth); the rungs of the Contact Clause; and every event's overall chance. The headline is
working off Earth on rotation once industry reaches the slow expanse: tens of thousands of people off
Earth, at least one site beyond the Moon, transit fast enough that people go out there to work, and
something made out there sold at a profit.
Beside it sits the nearer figure the home page shows: a rotation at a station or a lunar base, whatever
industry has reached. It is counted in the same play-throughs, so the two are always from the same run, and
it is shown so the gap between them is visible.
Weighing my own decisions
Some events in the tree are my decisions rather than the world's: at the six-year mark of my current
contract, stay or leave for civilian space work. These are never rolled. Instead the whole tree is run once
per option, with that option forced and everything else as usual, and the difference between the headlines
is what the decision is worth. That is what makes this a decision tool and not only a forecast. That part
of the tree is still a sketch and is rebuilt in 2027, so the comparison is reported but not yet read.
Running the tree for somebody else
Everything above is about my odds. The Your Odds page runs the same tree for a visitor.
The five deadlines are my ages 85, 94, 109 and 150, and no deadline at all. A visitor's are the same ages
shifted to their own birth year, and the page reads a number for each event at the shifted year by drawing
a straight line between the two nearest deadlines' numbers, on the scale where probabilities move evenly.
Nothing is read below the first deadline: for a reader older than me the number holds flat at the 2071
figure, and the page says that the tree cannot see below its own first window. Past the last deadline the
reading moves toward the no-deadline number and never passes it.
The tree can say how likely the world is to offer a chance. Whether a particular person takes it depends on
their money, health, work and luck, none of which the tree knows. So the page asks the reader for that
number themselves, shows it beside the world's part, and only then multiplies the two.
Scoring the forecasts
A forecast that is never scored is an opinion. Every event has a written test so that it can be scored.
Each January, every event that resolved in the year is scored with the standard measure, the Brier score:
take the probability that was on the record for it when it resolved, subtract one if it happened and zero
if it did not, square the difference, and average across the events. Zero is perfect, 0.25 is what saying
fifty percent about everything gets, and anything worse than that is worse than saying nothing. An event is
scored only if a probability for it was on the record before it resolved; the three that had already
happened before the tree was first estimated are not. The score is published with each resolved event's
recorded probability and outcome, and with the headline's movement over the year, which is the product the
project actually delivers.
What is fixed and what moves
Fixed, and changed only at a yearly review: the question, the four stages and what each requires, the
eight factors, the five deadlines, the size of the world nudge, and the scoring rule. Moved whenever the
world moves, with a dated note on the event: its status when it resolves, and its probability when the
weekly scan or a review finds reason. Added at any time: new events, when a scan or a critic finds a gap.
Never changed once a probability has been published against it: an event's test, so that the score can
always find the forecast it is scoring. An event that needs a different test is retired and replaced.
Where the numbers come from
The probabilities are estimates, and they are the least reliable part of the method; the rest of the method
exists to correct them over time. Each was first drafted by Claude, the AI I build this project with, with a
one-line reason, and then accepted or moved by me. Before anything was published, every factor's events,
tests and numbers were put to other AI models with a fixed brief: name a missing event, challenge a test,
argue a number up or down. I decided every disagreement myself. Every number carries its date and reason, and
every change since is a dated revision on the event, so a number can always be traced to why it is what it
is. Think a number here is wrong? Here's how to tell me, and what happens next.
The Contact Clause, beside the equation
Underneath the whole project is a dream with no practical place in the math: meeting a mind that is not
ours, whether a machine that has outgrown us or something of non-human origin. Drake's equation was about
contact with other minds; the Belt Equation is about becoming the kind of people who could be out there
to make it. The Contact Clause reconnects them. It is published beside the equation and never multiplied
into it, because nothing in the equation depends on it. Its five rungs follow the same rule as every other
event: each is in only because it has a written test. They form two branches, each ending in an exchange:
one with something not made by us, one with a mind we built. Nothing a system says about itself, and
nothing anyone feels in talking to one, counts toward any rung.
The project's own methodology document, with the date each part was settled and the reasoning behind it,
is in the repository: docs/methodology.md.
The story so far
What moved the number, and by how much
Every time something happens that touches the tree, the tree is played out again and the
headline number moves. This page is the record of those moves: what happened, which link of
the chain it touched, and how far the number went.
44 entries so far, across 9 runs of the tree.
Nothing on an entry is typed in by hand except the words: every figure is worked out from the
runs saved in the public repository, so an entry cannot claim a change the tree did not
produce. Entries are added and never edited; a mistake is corrected by a later entry that
says so. The project's full changelog, down to the commit, is
in the public repository.