How a report gets made

1. Commission - a person picks the question

A report starts as a signed brief with a person’s name on it. The brief states the strategic question in the words of the team that would have to make the call: the decision they would lose sleep over. It names a candidate thesis, the answer we expect the game to test, and the game format, the kind of game, that fits the question. Format follows question, never the reverse.

The brief also names the kill criteria, the results that would mean we do not publish. No run has ended there yet.

2. Play - the engine runs the game

The engine, the software that runs the game, is an event-sourced matrix-game simulator. In a matrix game, each player argues for a move and a referee rules on whether it works. Event-sourced means every move, ruling and roll goes into a log the moment it happens, so the whole game can be replayed.

Each actor is an AI agent playing a briefed persona, a written profile of how one organisation decides. On its turn an actor states what it is trying to do and its odds. Every rival argues against the move. An arbiter, the engine’s referee and an AI itself, turns the exchange into a ruling with a difficulty rating, the number the die has to beat, and itemised reasons. Then the dice decide.

Three rules hold in every engine run:

• The dice belong to the engine, never the AI. The model writes the rulings. A seeded random number generator decides the outcomes, and the seed is a starting number that makes the rolls repeatable. Same seed, same rolls, every time. The rulings come from a model, so they are not repeatable in the same way. A machine-played report, one where an AI arbiter ruled, is replayed from its log, which holds every ruling as it was made. After a fork the game continues with rule-based players and a rule-based referee, and their moves go on the log as new events. AI players and the AI referee do not replay a branch yet.
• Everything is on the record. Every claim, objection, ruling and roll lands in an event log that only ever grows. We can replay the run move for move, and fork it: same game, different dice. Same game, one shock, an outside event, removed. Same game, one decision changed at a chosen beat, a single turn.
• Human overrides are events too. If the operator, the person running the engine, overrules the arbiter, the original ruling stays in the log and the override sits beside it. Where the author went beyond what the run showed, the report says so in its own text.

What the engine can and cannot do is on the engine page. The mechanics of one beat, step by step, are in the engine explainer.

3. Present - a person writes the argument

The engine hands over a findings bundle: the event log, auto-drafted findings with citations, branch comparisons, and visualisation data. Dr Dan Epstein reads all of it against the brief, picks the five to eight findings that carry the argument, and writes the report. We treat the machine’s draft prose as quarry. We cut from it. Nothing the engine wrote ships unedited.

We built it this way for a plain reason. The engine can make sure every move, ruling and roll is on the record. It cannot pick the thesis, the argument the report makes.

The rules we do not bend

• Every claim with a number in it cites the evidence pack, the table of checked real-world facts built before the game, or the public source it came from. A linter, a script that checks the prose, enforces this, and the build fails without it.
• A demo is labelled as a demo. A machine-played run is labelled as one. A real client engagement is named, with the client’s consent, or it is not mentioned at all.
• Client-room material never appears in a public report. The engine keeps a wall between public evidence packs and client evidence packs. Crossing it is a hard error in the code.
• Every report states how much of it a machine played, above the fold. A human-refereed run, where a person judged every move, and a machine-played run, where an AI arbiter did, are different things, and we will not blur them.
• Where a machine played, the report states how many times it ran. 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. 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.

Where the rest lives