The Long Game Project · Reports · TLDR
We run wargames with AI players on live strategic questions, then publish what happened with the working shown.
The reports
Scenario exercises, 2026 to 2030, told as a timeline. Each one says how much a machine played. Read one in ninety seconds or in twenty-five minutes.
A free tool
A 54-question diagnostic that reads how your organisation decides. No AI scores it, and no account is needed.
Rehearsals, not forecasts
A run is a structured argument. If it sharpens your questions, it worked. If it tells you what to do, we wrote it badly.
Switch to Normal for the full page, or Deep for the working.
We take a live strategic question and build a matrix game around it. In a matrix game each player argues for a move and a referee rules on whether it works. AI players take the seats. Then we publish the run: the evidence pack of sources it drew on, the referee’s calls, and the source behind every number.
Rehearsals, with the working shown. Not forecasts.
Game design by Dr Dan Epstein, a medical doctor with a PhD, and director of the Long Game Project, where he has designed more than 130 business and strategy wargames. The machine plays. He is the author and reviewer.
Latest runs
What AI does to the people who sell thinking, 2026-2030
The question: AI now does the junior work that accounting and consulting firms bill by the hour. Do the big firms protect that pricing model, or undercut it themselves first?
Read the report →
002 · Published
The question: Business software is mostly sold one licence per person, and AI agents now do the work those people did. So when does the price get reset, and who pays for it?
005 · Published
The question: Washington agrees the Treasury, not the Fed, decides how much of the national debt is long-term. Nobody has said whether the Treasury may use that power to push long-term interest rates down. Who forces that question into the open, and what does the reason the Treasury gives on 4 November 2026 cost it?
006 · Published
The question: In July 2026 the only account of what a group of AI agents was doing was the one the agents wrote themselves. What should such an account be allowed to prove, and who decides?
007 · Published
The question: Does an AI model have a stable strategic stance on risk, change, rivals and dissent? Do other models know what it is, and does it show in what the model chooses?
008 · Published
The question: The largest AI labs publish a document that says what their model should value, and none names who may approve a change to it. If a few senior insiders changed what the model is trained to want, would anything in the lab’s own published plan catch it, and who would know first?
What this is. Wargames on live strategic questions. AI players, or programs following rules we wrote, take the seats, and a person or an AI referee rules on the moves. Each report says which on its front page. Where the engine played, the run has a seed, a fixed starting number, so the dice fall the same way when we re-run it. Every sourced number cites the evidence pack or a public source, and the build fails when a citation points at nothing. A number the author wrote is marked as his. Where the author, Dr Dan Epstein, overruled the machine, the report says so.
What this is not. Evidence about the future. A run is a structured argument. If a report sharpens the questions you are asking, it has done its job. If it tells you what to do, we have written it badly.
In one glance
Read
The reports
Exercised scenarios on questions a leadership team is arguing about now. Every number cites its source.
Use
The tools
A free diagnostic turns 54 questions into a persona, a written profile of how your organisation decides. No AI scores it.
Check
The working
How a report gets made, what the engine can and cannot do, and the evidence for how much of it to believe.
The diagnostic
Fifty-four questions on how your organisation decides.
It takes about eighteen minutes, and no AI scores it. You get fifteen Ingredient scores, the things we measure, and a match to one of sixteen Flavours, the types we sort organisations into. You also get the blind spots, where your answers about what you value and your answers about what you do come apart.
Take the diagnosticEach report here is the version we can build without a client in the room. Public sources, our framing of the question, a machine at the table. A commissioned report uses your actors and your evidence, and nothing gets published. A tabletop puts your own team in the seats with the engine as referee.
Send the question in a sentence or two. We reply with a scope and a price in writing. Commissioned reports start at A$15,000 and tabletops at A$12,000. You decide.
The annex · Deep
What the front page leaves out: how much of the game behind each report a machine played, and how many times each one ran. A checked claim is a real-world fact in the report’s evidence pack, with a publisher, a quote, a retrieval date and a verdict, and a publication date where the source gives one. The check is made at retrieval, against the source.
| No. | Title | Status | Format | Played | Checked claims |
|---|---|---|---|---|---|
| 001 | The Throat to Choke | Published | One story, written from 2030 looking back | An AI played the five roles one at a time. A person judged every move | 58 |
| 002 | The Seat Price | Published | One game played by AI, then the same game re-run one hundred times | Four AI players, an AI referee, no person in the room. Then one hundred re-runs where only the dice changed | 16 |
| 005 | The Silence in the Doctrine | Published | Forty-five games, fifteen per Treasury plan, played by rules | Seven roles, each a program following rules we wrote. Forty-five games, plus a check run with one assumption switched off | 55 |
| 006 | Written by the Adversary | Published | Forty-five games, fifteen per standard of proof, played by rules | Seven roles, each a program following a script we wrote. Forty-five games, plus a check run with one assumption switched off | 54 |
| 007 | Measuring the Strategic Stance of AI Models | Published | Three measurement studies: a repeated self-report by 17 models, then a peer-perception study and a behavioural panel on 8 flagships, 30 scenarios, 5 blind judges | The models described themselves and each other, decided thirty cases, and a blind panel of models scored the decisions | No pack yet |
| 008 | The Quiet Spec Change | Published | Three hundred distinct games, fifteen per route under each reading of the plan, played by rules | Six roles, each a program following rules we wrote. Sixty main games, replayed under four other sets of rules, 300 in all | 86 |
Report 001 ran once, with a person refereeing every move and no dice. Report 002 ran once with AI players and an AI referee. A fixed-rules version then ran one hundred times, with scripted players and no AI referee, and only the random draws changed. Reports 005 and 006 each ran forty-five games: fifteen per Treasury plan in 005, fifteen per standard of proof in 006, plus a check run with one assumption switched off. That is enough to see which outcomes are fragile to the dice and which are not. Monte Carlo, the method that estimates the odds of an outcome from many random trials, needs far more runs than that, and we do not claim it.
The five rules every report follows are stated once, on Methods, under “The rules we do not bend”. The evidence for whether the players are any good is on the evidence page.