One claim, from the day it is written
Says what it thinks.
How often it was right? Nobody checked.
Read the record. Decide how much to trust the next one.
What a cycle can show
A standing brief, re-played on the dates that matter
the long end of the debt market
a new power
the question does not close →
The archive: every closed cycle, frozen
Replay every fixture in the archive under it.
The rule can change.
The change stops. Explain the flip in writing. Correct the earlier report.
A change that quietly rewrites history: the failure we guard against.
A defect: anything a cycle shows the game got wrong
A prediction of what it will fix.
A condition under which we will undo it.
A named person changes it.
Earlier cycles that must replay with no published finding flipped, before a rule can change
cycle
The game gets harder to change as it gets older.
What every rule change carries
A handfulof rule changes a year
A human choice. The plan is the forecast.
Games we wrote
Say, nine of twelve
Not a probability.
The record page marks each outcome
From a game played by AI models, one seat each.
Marked: a line the games played by rules alone do not also support
Kept on the games played by rules alone. Only the dice change.
The two kept apart on purpose
Marked claim by claim, on its record.
A dated record instead of an opinion.
A perpetual game writes each claim down with the date it was written and the date it resolves, before the answer is known.
A check on the mechanism, not just the call.
Getting the answer right through the wrong mechanism is luck, and a perpetual game is built to notice the difference.
A question that keeps coming back.
Some questions do not close.
What it is useful for
A dated record instead of an opinion.
A report played once can say what it thinks. It cannot say how often it was right, because nobody checked. A perpetual game writes each claim down with the date it was written and the date it resolves, before the answer is known. After a few cycles you can read the record and decide for yourself how much to trust the next one.
A check on the mechanism, not just the call.
The most useful thing a cycle can tell us is that something happened which no player in the game could have done. We call that a surprise. A surprise means the game was missing a move, or a player, and it means that whether or not the call came out right. Getting the answer right through the wrong mechanism is luck, and a perpetual game is built to notice the difference.
A question that keeps coming back.
Some questions do not close. What a central bank does about the long end of the debt market is one. What a regulator does with a new power is another. If your organisation has a question like that, a perpetual game gives it a standing brief that is re-played on the dates that matter and read against the record each time.
How it improves itself, and why slowly
Every closed cycle becomes a frozen, replayable fixture: the same game, the same dice, and what the report said. The fixtures together are the archive. Before we change any rule of the game, we will replay every fixture in the archive under the new rule. If a finding we published in an earlier cycle flips, the change will stop until we can explain the flip in writing and correct the earlier report. A change that quietly rewrites history is the failure we are guarding against, not the goal.
A defect is anything a cycle shows the game got wrong: a surprise, a move that fired for the wrong reason, a player that never did anything. A defect has to be seen twice, in two different cycles, before it can move a rule. One surprise is noise. Two is a pattern. And every change is written up before it is made, with a prediction of what it will fix and a condition under which we will undo it. Programs play the game. A named person changes it, and there are few of those.
The arithmetic does the rest. Before a rule can change in the fourth cycle, the three earlier cycles must replay with no published finding flipped. By the twelfth cycle, eleven must. The game gets harder to change as it gets older, which is what a record is for.
That makes improvement slow. A handful of rule changes a year, each with a written reason and a replayed archive behind it. We think that is the right speed for anything that goes out with our name on it. (It is also the speed at which we can read everything we wrote.)
What it does not claim
A perpetual game is not a forecast, and its numbers are not probabilities. When a report says an outcome happened in, say, nine of twelve games, that is a count over games we wrote, under a plan we chose for one player. Change the plan and the count changes. The plan is the forecast, and the plan is a human choice. The record page marks which outcomes follow from the plan and which come from the play.
The story you read in a perpetual report is written from a game played by AI models, one seat each. The record is kept on the games played by rules alone, where only the dice change. We keep the two apart on purpose, and the report marks any line in the story that the games played by rules alone do not also support.
Whether the game tracks the world is answered claim by claim on report 005’s claim ledger, first marked on 4 November 2026.