The Long Game Project · The Engine · Changelog · TLDR

The engine began in 2022 as a form that sent text to an AI and got a document back. It now plays whole games, with referees, dice, evidence and a board. Seven stages got it there.

2022 to 2025A form sent text to an AI, and a document came back.
March to April 2026A questionnaire briefed AI players who argued in rounds.
May 2026A referee, dice and a log made games repeatable and forkable.
June 2026Many worlds per game, and a conversation to design it.
July to August 2026Checked evidence went in, and the players were tested.
September 2026Games played again on a fixed calendar and checked against what happens.
October 2026Swappable rules, a board, talks and fog.

The newest stage adds a board that holds several domains at once, such as sea, cyber, opinion and markets. Players can talk, work through fog, run short of supplies and play against a day clock.

Switch to Normal for the timeline and the board, or Deep for how each stage was checked.

The Long Game Project · The Engine · Changelog

How the engine changed

The first version was a form that sent text to an AI and got a document back. The current one plays whole games, with referees, dice, evidence and a board. This page follows the change from 2022 to October 2026, stage by stage. It shows what each stage did and the idea behind it. It leaves out the prompts, settings and code.

2022 to 2025

One form, one text AI, one document

Our first engine was not an engine. A visitor filled in a form. An automation on Make.com sent the answers to OpenAI's GPT-3 with a fixed prompt, and the reply landed in a Google Doc. A second form turned one idea into a full facilitation guide, with injects (events dropped in mid-game), handouts and a debrief.

Added

DALL-E drew a picture for each scenario.
A row in a database for each request, so we could see what people asked for.

What it could not do yetIt wrote one document in one pass. Nobody played it and nothing checked its facts. An experienced wargame designer reviewed three of its scenarios. The verdict was more scenario than game: a description of a situation, with no one given a job to do.

How we knewOnly by review. The designer’s critique was the one outside test, and it set what we built next.

March to April 2026

A questionnaire briefed AI players, and they argued in rounds

In March 2026 the project became Strategy Soup. A 45-question diagnostic turned an organisation into a profile. The profile briefed an AI player to act the way that organisation tends to act. We put several players in one room, ran rounds of argument, and had an observer score the room and an analyst write the note.

Added

The diagnostic, and a library of ready-made profiles.
Persona chat: talk to the AI playing your organisation and see its reasoning.
Rounds of play, streamed to the screen as they happened.

Retired

The picture step and the one-shot guide.

What it could not do yetPlayers argued, but nothing settled who won. The note summarised a conversation and ruled on nothing. No run could be repeated, so no result could be checked.

How we knewBy reading transcripts. With no replay, a surprising run could not be reproduced.

May 2026

A referee, dice and a log made every game repeatable

We rebuilt the engine around a log of what happened. A player claims a move and argues its odds. Rivals argue back. A referee sets the target and writes its reasons. A die that the engine rolls itself settles the move. Everything goes on the log, and the log is the game: the state at any turn is the log replayed up to that turn.

The first test was a game we had already run by hand on a spreadsheet. The engine reproduced each team’s result to within a few points. The rebuild was merged, reverted the same day and merged again the next.

Added

Replay: the same log and seed give the same game, roll for roll.
Fork: copy the log to one turn, change one thing and play on.
Four referee modes, from fully automatic to a person ruling with the AI assisting.
A console to watch a run, step through it and compare two.

Changed

Players and the referee can each be an AI or a set of written rules, so a game can run on any mix.

What it could not do yetIt played one game at a time. One run is one path, and we could not yet say how fragile that path was.

How we knewThe spreadsheet game, replayed. Then the same log and seed, replayed, gave the same result each time.

How one beat works

June 2026

Many worlds, and a conversation to design the game

Two additions. First, the engine plays one game many times, each with a different seed, a starting number that fixes the dice. It reports the spread of endings and not one story. Second, a Design Room: a guided conversation that turns a question into a game, in stages, from the question itself to the list of moves.

Added

Ensembles: one game run across many seeds, with the spread of endings shown.
Outcome worlds: endings grouped by how they came about, with the chain of decisions behind each group.
A gate in the run console, where a person approves a ruling before it counts.
A safety and correctness review of live runs, before anyone outside the team used them.

What it could not do yetPlayers still played from their written briefs alone. If a brief got the world wrong, the players played on the error.

How we knewReplays from the same seed had to match, and did. The review covered the engine’s arithmetic and its safeguards, and what it found was fixed.

July to August 2026

Checked evidence went in, and we tested the players

Before a game starts, a research pass builds an evidence pack: real-world claims, each checked against its source and given a verdict. Players and scenario numbers cite the pack, and the pack is published with the report.

In August we ran five studies of whether the AI players act like the people they stand for. The players proved distinct from each other, and they chose differently when briefed differently. We could not show they were accurate. One result we had announced did not survive review, and we took it back.

Added

Evidence packs, with a verdict on every claim.
Injects now reach the players and can cite the pack. Before July they were logged and never shown to the players.
Studies whose rules were written before each one ran.

Changed

The question moved from “does swapping a persona change the outcome?” to “does it change behaviour?”, because a review found the first test faulty.

What it could not do yetA pack holds only what public sources say. A player cannot know more about a market than the documents behind it.

How we knewEvery claim in a pack carries a verdict. The studies are written up on the evidence page, with the failures next to the passes.

The evidence page

September 2026

Games that keep going, played from the record

A perpetual game does not end with a report. It is played again on a fixed public calendar and checked against what really happens. Each claim is written down with the date it resolves, before the answer is known. The first cycle is played again after each Federal Reserve meeting, from the public record.

Researched organisations now play from their own public record, and not from a written character. We also had the engine’s code reviewed from scratch, and fixed the faults that review found.

Added

A claims ledger: every claim is dated when written and again when it resolves.
Players briefed from public records.
Ensembles of forty-five games in published reports.

Retired

Persona chat came off every public page on 18 September.

What it could not do yetThe first cycle is still open, so no claim has been scored. One cycle would not be a track record anyway.

How we knewEach claim is fixed in writing before the answer is known, so the check cannot be moved afterwards. The code review read the engine with fresh eyes.

Perpetual games

October 2026

Rules became parts you can swap, and a board arrived

In October the engine started taking its rules as parts, so one engine can run different kinds of game. Players now act on a board.

Added

Rule sets as parts: a race between groups where sharing can help a rival, stocks and flows a reader can play with, and talks between players.
Talks: promises, threats, side deals and leaks, with a private notebook for each player.
A board: places, who holds them, supplies that run down, and a clock that counts days.
Fog: a player sees only what its role could see.
Dice odds that update as evidence arrives.
Reader maps: step through a decision and see where each choice leads.
Paired experiments: change one condition, keep the seeds and compare.

What it could not do yetMost of this is new. Some of it has not yet run in a published report, so treat it as built and not yet proven.

How we knewMost new parts stay off unless a game switches them on, so a game written before October is not changed by them. Each part had a code review from a fresh start.

What the board can hold and do Economy flow

A form sends text to an AI. The AI returns a document.

The board, as capabilities

What the board can hold and do

A domain is one arena with its own map and its own speed. A game stacks the domains it needs, such as sea, air, cyber, public opinion, institutions or markets, and effects pass from one domain to another. This is what the engine can do now. Not every game uses every part.

On the board

Places and assets. Each domain is a map of places. Players own assets and put them on the map.
Control. Who holds a place follows from what each side has there.
Orders. Players give orders that move assets, tie them up, and act on places and on other domains.
Effects across domains. A hit in one domain can land in another after a delay.

Time and supply

Time. A game can count real days, so journeys, repairs and resupply take time.
Supply. Assets use up supplies and need resupply.
Pressure. Rivals can threaten places, and a game can end when pressure crosses a stated line.

Seeing and talking

Fog. Each player sees its own picture of the board. It can be late, partial or wrong, and some assets stay hidden.
Talks. Players can make promises, threats and side deals, and can leak.
A plain brief each turn. An AI player is told what it holds, what is contested and what is coming.
Odds that follow evidence. The odds on a move can update as evidence arrives.

For a report, the board records who held each place on each turn, and what each player believed against what was true.

How we knowThe board was built to leave games that do not use it unchanged, and it was checked against a large set of older games when it went in. A test checks that a hidden asset does not change what a player who cannot see it is told.

Where this leaves the engine, 23 October 2026

The engine runs games with referees, dice, evidence and a board. It still cannot predict, and it still cannot know what its evidence pack leaves out. The engine page lists what it can and cannot do today, and the evidence page holds the tests of the players.