A matrix game: argued in words
A briefed persona, with its own goals, red lines and information.
No actor grades its own move.
A language model under its own brief. It plays no role.
One ruling
One square is one event on the log
Before the branch point: the run as it was
After it: rule-based players and a rule-based referee, logged as new events
Cyan: the compare view marks the first event where the two branches differ
Before the game runs, a research pass builds
Cyan: each claim checked against its source at retrieval
A reader of a published report gets
Where the expert who signs the report and the machine disagreed.
For every report, in one table on Methods.
A run is one path through an argument
The value sits in
What its evidence pack leaves out
A person writes it up.
With the engine as the referee.
The strongest version of this methodEach report says how many times it ran
One square is one run
A separate question
- Play actors that want different things
- Rule with written reasons
- Replay any run and fork it
- Cite checked evidence
- Show its working
Cyan: checked against the code on 18 September 2026
- Predict
- Know what its evidence pack leaves out
- Play a persona better than its brief
- Replace a room full of your own people
- Tell you how likely anything is
- Actor-play matrix gameMost published reports
- Counterfactual fork
- Market or auction game
- Red-team single actor
- Human tabletop with the engine as refereeThe newer of our two products
One beat
One beat of a game, a single turn, in order. Every beat leaves the same five marks on the log: the claim, the objections, the ruling, the die face and the outcome.
Can
Play actors, AI players, that want different things. Rule with written reasons. Replay any run, and fork it at a chosen turn. After a fork, fixed rules play on. Cite checked evidence. Show its working.
Cannot
Predict. Know what its evidence pack, the table of checked facts it plays from, leaves out. Play a persona, a written character, better than its brief. Replace a room full of your own people. Tell you how likely anything is.
Formats
Five. Actor-play matrix game, counterfactual fork, market or auction game, red-team single actor, and a human tabletop with the engine as referee. The first is the format of most published reports. The tabletop is the newer of our two products.
Five things it can do
Run matrix games with opposed actors.
A matrix game is a game argued in words, where a referee sets the odds on each move. Each actor is an AI agent playing a briefed persona, a written character with its own goals, red lines and information. The actors argue against each other’s moves, in writing. No actor grades its own move.
Rule with reasons.
The arbiter is the referee, a language model under its own brief, and it plays no role. It sets a target for every claimed move and writes the case for and against into the transcript. It states its confidence in each ruling. The evidence page shows that AI players overstate their own odds by 14 to 28 points.
Replay and fork.
Give the engine the same seed and log and it replays the same run, roll for roll, with every ruling read back as it was made. The seed is a starting number that fixes the dice. We can rerun history with different dice, remove a shock, an outside event, and watch what changes, or branch a finished game at a chosen beat, a single turn, with one setting changed. After the branch point the game continues with rule-based players and a rule-based referee, logged as new events. AI players and the AI referee do not replay a branch yet. The compare view marks the first event where the two branches differ.
Ground play in checked evidence.
Before the game runs, a research pass builds an evidence pack: a table of real-world claims, each one checked against its source at retrieval and given a verdict. Personas and scenario numbers cite it, and the pack is published with the report.
Show its working.
A reader of a published report gets the evidence pack and a seams section that lists where the expert who signs it and the machine disagreed. Methods states, for every report, how much of the game a machine played and how many times it ran.
Five things it cannot do
Predict.
A run is one path through an argument, and dice decide part of it. The value sits in the assumptions it drags into the open and the knock-on effects nobody had written down. Read a run for its ending and you have read it wrong.
Know what its evidence pack leaves out.
The research pass uses public sources. An actor cannot know more about your market than the documents behind the actor.
Play a persona better than its briefing.
The actors are language models in character. A thin brief makes a thin actor. Now and then a well-briefed one still does something no board would do. When that happens, the seams section says so. The misfires stay in the record.
Replace a room full of your own people.
The strongest version of this method puts your own people at the table, with the engine as the referee. That is the tabletop. It is a different product from the published reports, and it runs on the same engine.
Reach statistical significance about the world.
That is the bar a result clears before you can say it is more than chance. Each report says how many times it ran. Report 001 ran once, with a person refereeing. 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 ran forty-five games each. Re-running the same game with only the dice changed tells you which outcomes are fragile to the dice. It does not tell you how likely anything is.
Whether the players are any good is a separate question. The only page allowed to answer it is the evidence page.