Strategy Soup How a report gets made

The Long Game Project · How it works · The engine · TLDR

A question, played as a game

A question becomes a game. Each party gets a seat and a private brief, its goals and limits. An arbiter, the referee, sets the odds on each move. Seeded dice, rolls that repeat exactly from a starting number, decide. An expert writes the argument.

ClaimObjectRuleRollRecord
One beat, one turn of play: a claim, the objections, a ruling, a roll, a record. The game is this loop, run again and again.

The Long Game Project · Method · The wargame engine

What the wargame engine is, and what it is not

A question becomes a game. Each party gets a seat and a private brief, its goals and limits. An arbiter, the referee, sets the odds on every move. Seeded dice, rolls that repeat exactly from a starting number, decide. An expert reads the transcripts and writes the argument. This page makes no claim that the machinery predicts anything.

What an event carries

A question, played as a game

An exercised scenario, start to end

A questionOne a leadership team is arguing about.
Seats and private briefsA goal, win conditions, red lines, a read on everyone else.
BeatsArgue for, argue against, set the odds, roll.
The logEvery claim, argument, ruling and die face.
The argumentAn expert reads the whole record and writes it.
Where the expert overrode the machineWhere the evidence cuts against the game

Cyan: on the recordMagenta: our judgement

Before the argument starts

The frameThe room agrees on it. The money is here.something nobody wrote downa reason nobody had named
running the game→

In the open

An ending
a load-bearing assumption, nameda load-bearing assumption, named

Disagree with them by name

Where the odds on a move come from

A language model’s judgement→A number on a table→A seeded die
2030not a forecast

Scenarios, not forecasts

On every report’s front page

Archetypes, not real companies

The incumbent vendorA megavendor. Per-user licences.
The AI-native challengerPrices per outcome.
The Quiet Discountfour seats
The corporate buyerAn enterprise customer.
The integratorA global systems integrator.
Each seat’s private briefa goalwin conditionsred linesa read on everyone else

Eight half-year moves

2026
H2
2027
H1
2027
H2
2028
H1
2028
H2
2029
H1
2029
H2
2030
H1

Where the personas come from

54questions

→

15strategic dimensions

→

16Flavours. The score sits near one.

→
The seat’s private brief

A consistent character, not a generic role.

Questionnaire to brief: the step we found broken in August 2026, and fixed.The validation page
The arbiter

Plays no seat. Rules on claims.

→

Every ruling puts four things on the log

  1. The decision
  2. The reasoning path
  3. The assumptions it rests on
  4. A self-reported confidence

Consistent with its earlier rulings in the same run

rulingrulingrulingthis ruling

certainNever presented as certain

Cyan: on the log

The arbiter’s four modes

1Rules at once

and logs the reasoning

2Settles routine calls

halts for confirmation on major ones

3Proposes every ruling

a human approves it

4A human rules

the arbiter assists

more ropeless rope

Each report states its mode

One beat, five steps

A seatClaim and oddsAnd a reasoning journal.
Every other seatArgues againstIn writing.
The arbiterSets a targetStated odds, on a fixed table.
The dieRolls the diePlus the seat’s resources.
The engineAppends to the logAll of it, in one go.

The run’s event log

an earlier beatclaimobjectionsrulingdie faceoutcome
an earlier beatclaimobjectionsrulingdie faceoutcome
this beatclaimobjectionsrulingdie faceoutcome

Cyan: on the logDashed: not on it yet

A seat’s claim

The move it proposes

With the odds it thinks the move deserves.

Its reasoning journal

  • what it considered
  • what it rejected
  • how confident it is
another seatanother seatanother seat
→
The claim

Argued against, in writing. A game, not a self-graded essay.

A seat that cannot argue: the engine warns of fewer objections

Stated odds become a target on one fixed table

likely → easy targetunlikely → hard target
easyhard
target the roll margin: how decisively it landed
ClaimObjectRuleRollRecord machine-playedbeats
A beat an author writes by hand

Outside this loop. The report says so.

Cyan: on the logMagenta: the author’s work

The engine stores events, not state

claimobjectionsrulingrollclaimobjectionsrulingroll
→
The state

Rebuilt by replaying them.

A what-if

  1. Copy the log up to a point
  2. Change one thing
  3. Replay from there
1 2 3 maina what-if

One playthrough

An anecdote. The dice could have fallen the other way.

One hundred, under fixed rules

Same seats, same menu of moves, different dice.

The live runmodel seats
The fixed-rules versionseeded policy agents
The hundred tell you

The shape of the game we wrote.

They tell you nothing

About how a model behaves.

The incumbent vendor’s strategic position after one hundred worlds

5th95th about nine worlds in ten ended in here 8295100112 0 out of themarket medianThe incumbent starts on 100 Not one world 100 to zero in eight half-years: close to worst-case dice on almost every turn 5th95th about nine worlds in ten 8295100112 0 out of themarketNot oneworld medianThe incumbentstarts on 100 100 to zero in eight half-years: closeto worst-case dice on almost every turn

Cyan: what the hundred runs showed// a gap in the scale

What if the incumbent cannibalises itself early?

Baseline

Harvest first, pivot later.

same seed
Early pivot

Forced in all one hundred worlds.

57the baseline ended higher
39the early pivot won
4tied

Of 100 paired worlds. Median cost of pivoting early: one point of final position.

Cyan: what the paired runs showed

Each world has its own seed

The same world back, roll for roll

Reproducible

Repeatable arithmetic.

Not certain

It says nothing about the future.

The beatsThe dashboard seriesThe forks and the endingsThe evidence packA seams sectionA named human

In the live run: the same public play, five times

The arbiter raised the target each time

Cyan: from the run’s record

The timelinefrom the run, beat by beat
The index linesdrawn by hand with the scenario

Not measurements. Do not quote them as data.

Cyan: from the runMagenta: drawn by the author

The Case of the Disappearing Margin: three endings

An endingand its named load-bearing assumption
An endingand its named load-bearing assumption
An endingand its named load-bearing assumption

Pick one, and see what you had to believe

A sentence with a number in it[s-003]

PublisherDateVerbatim quoteRetrieval dateVerdict
An id not in the pack→ an error
No marker→ authored, and the report says so

Cyan: a citation marker

Where the machine surprised the authorWhere the author overrode itWhere the verified evidence cuts against the scenario

The Case of the Disappearing Margin: eight places the evidence pushes back

The engine

Produces a record.

→
A person

Makes the argument, the framing and the judgement calls.

Every published report’s bylineAn expert, named on the report

Cyan: the recordMagenta: our judgement

The Case of the Disappearing Marginhuman refereed
AI playAI playAI playAI playAI play
An expert refereed them
“human judgement is the durable moat”all five plays converge here
The Quiet Discountmachine played
beat 1beat 2beat 3beat 4beat 5beat 6beat 7beat 8
live run: four model seats, an AI arbiter, dice
Fixed-rules version: the full eight beats, one hundred times

“Machine played”: five beats in eight

The 2029 “dam break” is the author’s reading. No run produced it as written.

Cyan: played in the live runMagenta: the author’s work

The table

The seats are archetypes, types drawn to be recognisable, rather than real companies. Each has a goal, win conditions, red lines and a read on everyone else.

The beat

A seat argues for a move. The others argue against. The arbiter sets a target. A die settles it. All of it goes on the log.

One hundred worlds

For “The Quiet Discount” we re-ran the game one hundred times under fixed rules, with only the dice changed, so the report describes a spread of endings instead of one story. Each report states its own count.

Not a forecast

The engine knows nothing special about 2030. Its job is to make assumptions arguable.

An exercised scenario, in one paragraph

We take a question a leadership team is arguing about and turn it into a game with rules. Each party who matters gets a seat and a private brief: a goal, win conditions, red lines, and how it reads everyone else at the table. Play advances in beats: a seat argues for a move, the other seats argue against it, an arbiter sets the odds, and a die settles it. The log keeps every claim, argument, ruling and die face. Then an expert reads the whole record and writes the argument, marking where they overrode the machine and where the evidence cuts against the game.

What it is for.

Making assumptions arguable. A room usually agrees on the frame before it starts arguing, and the frame is where the money is. Running the game forces the frame into the open. The incumbent keeps taking revenue from its old price model because of something nobody wrote down, or the challenger’s pitch fails for a reason nobody had named. Our published reports name their load-bearing assumptions, the ones an ending cannot stand without, so you can disagree with them by name.

What it is not.

It is not a forecasting tool. The engine knows nothing special about 2030. The odds on a move come from a language model’s judgement, converted to a number on a table, then settled by a seeded die. Our reports say the same on their own front pages: scenarios, not forecasts. “The Quiet Discount”, one of the published reports, says in its method note, the section that explains how it was played, that we make no claim the games predict real-world outcomes.

Who sits at the table

The seats are archetypes, not real companies.

“The Quiet Discount” had four. The incumbent vendor is a megavendor whose revenue base is per-user licences, and the AI-native challenger prices per outcome. The corporate buyer is an enterprise customer, and the integrator is a global systems integrator. Each started with a declared strategic position and played eight half-year moves, from 2026 H2 to 2030 H1. They are composites drawn to be recognisable, and the report says which real behaviour each one copies.

Where the personas come from.

The diagnostic asks 54 questions, scores an organisation on 15 strategic dimensions, and places that score near one of 16 Flavours. When we design a game, we ground each seat in one of those Flavours, so it plays a consistent character instead of a generic role. That grounding goes into the seat’s private brief. The step from questionnaire to brief is the one we found broken in August 2026 and fixed. The validation page tells that story. Read it before you trust any of this.

The arbiter plays no seat.

It rules on claims. Its job is to set a target for a proposed move from the arguments for and against, and to stay consistent with its earlier rulings in the same run. Every ruling puts four things on the run’s event log: the decision, the reasoning path, the assumptions the ruling rests on, and a self-reported confidence. The engine never presents a ruling as certain.

How much rope the arbiter gets is a setting.

The arbiter runs in one of four modes. It can rule at once and log the reasoning, or settle routine calls itself and halt for confirmation on major ones. It can propose every ruling for a human to approve. Or a human can rule while the arbiter assists. Which mode an exercise ran in is a fact about that exercise, and we state it per report.

How one beat works

A beat is the smallest unit of play. Every beat runs the same five steps.

Step 1 of 5

Each seat proposes its move as a claim and states the odds it thinks the move deserves. It also records a reasoning journal: what it considered, what it rejected, and how confident it is.

Step 2 of 5

Every other seat argues against that claim, in writing. This step is what makes it a game and not a self-graded essay. If a seat cannot argue, the engine warns that the other seats’ claims will face fewer objections.

Step 3 of 5

The arbiter reads the claim, the arguments for and against, and the seat’s current position. Then it sets a target. Stated odds convert to that target on one fixed table, the same in every game. A move argued as likely becomes an easy target, and a move argued as unlikely becomes a hard one.

Step 4 of 5

The engine rolls a die against the target, and the seat’s resources can help the roll. How far the roll clears or misses the target sets how decisively the move landed. The roll, the target and every adjustment go into the record.

Step 5 of 5

The engine appends the claim, the objections, the ruling, the die face and the outcome to the run’s event log.

The same five steps, every machine-played beat, until the game ends. A beat an author writes by hand sits outside this loop, and the report says so.

The engine never stores game state directly. It stores events and rebuilds the state by replaying them. That is why a what-if is cheap: copy the log up to a point, change one thing, replay from there. It is also why you can audit a run beat by beat instead of taking a summary on trust.

One hundred re-runs turn a story into a spread

One playthrough is an anecdote. The dice could have fallen the other way, and a report built on a single run is a report built on luck. So for “The Quiet Discount” we also ran a fixed-rules version of the same game one hundred times: same seats, same menu of moves, different dice. Then we read the spread instead of the story.

The hundred runs are not the language models playing.

Seeded policy agents, players that follow a fixed policy instead of a language model, drive the fixed-rules version. The model seats that played the live run do not drive it. The hundred runs tell you about the shape of the game we wrote. They tell you nothing about how a model behaves. “The Quiet Discount” sets this out in its own method note.

A spread instead of a story.

The fixed-rules game scores each seat’s strategic position. The incumbent vendor starts on 100, and a seat that reaches zero is out of the market. Across the hundred worlds, nine in ten ended between 82 and 112, the 5th and 95th percentiles, with a median of 95. Not one world produced a collapse, meaning no run drove the incumbent vendor out of the market inside the window. The report says that is partly a property of the board: on these rules, falling from 100 to zero inside eight half-years needs close to worst-case dice on almost every turn.

We compare forks seed for seed.

To test “what if the incumbent cannibalises itself early”, we forced that move in all one hundred worlds and compared each world against its own baseline, same seed. The baseline, harvest first and pivot later, ended on a higher final position than the early pivot in 57 of 100 paired worlds. The early pivot won 39 and 4 tied. That is closer to even than a headline would make it, and the report puts the median cost of pivoting early at one point of final position. Pairing on the seed stops anyone mistaking a lucky run for a better strategy.

Fixed rules means reproducible, not certain.

There is no ordinary randomness anywhere in the resolution path. Every roll comes from a seeded generator, a random-number source that gives the same sequence for the same seed, and the event that used the roll records its face. Each world has its own seed. Re-run world number seventeen and you get the same world back, roll for roll. It is repeatable arithmetic, and it says nothing about the future.

Six things come out the other end

1 of 6

The beats.

A timeline of what happened, beat by beat: the moves tried, the ones that failed, and why the arbiter priced them the way it did. Trying the same move again is a mechanic with a cost. In the live run the challenger tried the same public play five times, and the arbiter raised the target each time.

2 of 6

The dashboard series.

The index lines beside the timeline are drawn by hand with the scenario, so they stay consistent with the runs. They are not measurements, and both the chart note and the report’s method section say so. They show the shape the author argues. Do not quote them as data.

3 of 6

The forks and the endings.

“The Case of the Disappearing Margin” ends on three endings, and each carries a named load-bearing assumption: the thing you have to believe for that ending to hold. Pick your ending and you will notice what you had to believe to pick it.

4 of 6

The evidence pack and its citation markers.

Every bracketed marker like [s-003] in a report points to a claim in an evidence pack. Each claim carries a publisher, a date, a verbatim quote, a retrieval date and a verdict. A linter, a script that checks the prose, then holds the report to it. A sentence with a number in it has to cite a run event or a pack claim. Citing an id that is not in the pack is an error, because invented provenance is worse than none. Anything without a marker is authored, and the reports say that in plain words.

5 of 6

A seams section.

Every published scenario report carries one. It says where the machine surprised the author, where the author overrode it, and where the verified evidence cuts against the scenario. On “The Case of the Disappearing Margin” that section lists eight places the evidence pushes back.

6 of 6

A named human writes the argument.

Every published scenario report carries the byline of the expert who wrote it. The engine produces a record. A person makes the argument, the framing and the judgement calls, and that person is named on the page.

How much of each report a machine played

The answer differs per report. Methods states it for every report in one table. Two examples follow.

The Case of the Disappearing Margin - human refereed.

The author’s own account: an AI generated the five actor plays and an expert refereed them. The report’s seams section then argues against its own material. One example: all five plays converge on “human judgement is the durable moat”, which is a suspiciously comfortable finding when an AI wrote the first pass.

The Quiet Discount - machine played.

No human refereed this one while it ran. Four model seats played the archetypes in a five-beat live run, each with a private brief, win conditions and red lines. An AI arbiter set a target for every claimed move and wrote the case for and against into the transcript. Dice decided. Separately, the fixed-rules version ran the full eight beats one hundred times. An expert commissioned the question, set the table, read every transcript and wrote the account.

Where the human wrote past the run.

In “The Quiet Discount”, the live run stopped at beat five. Beats six through eight are the author’s work, written on top of the hundred fixed-rules runs, and no model seat played them. So “machine played” is true of five beats in eight, and the ending is authored. The 2029 “dam break”, the beat where the built-up pressure on the secret of what each customer pays gives way, is the author’s reading of the game’s pressure mechanics. No run produced that beat as written. The report discloses this, and we repeat it here.

New in the engine: the Domains board

The board

What is true

Escort groupEscort groupSubmarine

One rival’s picture

Escort groupEscort groupnot shown

The test: two runs differ only by the hidden submarine. The rival who cannot see it gets the same words and numbers in both.

Threat, time and supply

Threat
Held by one player, denied by another: risky to use, not taken
Time
Decide once a turn. Orders run day by day, and arrive when they arrive
Supply
Fuel, munitions or cash: used up, refilled slowly

9,304

recorded worlds, each one complete run, replayed with and without the new parts. Every log and final state the same

81,000

turns on which the opinion layer matches report 011’s Australian opinion index

4

published reports whose simulations are among those runs: 002, 009, 010 and 011

The Domains board plays several contests at once, one layer for each.

A game can now run on a board with layers, such as sea and air, public opinion, markets, and the forums where votes are cast. Each layer has places. What each player holds at a place decides who controls it, who contests it and where it is open. Layers act on each other: a blockade at sea can move opinion at home. The switch shows the same board for a defence, a business and a diplomacy game.

The board does not decide moves. The AI players still argue their moves and the dice settle them, as above. A scenario uses the board only where it needs one.

A threat can deny a place without taking it, and no player sees the whole board.

Any asset can make a place risky or costly for its rivals: a shore battery over a sea lane, a patent suit over a market, sanctions over a vote. Control still goes by what each player holds, so one side can hold a strait that the other side denies.

A player sees its own assets in full, and a rival’s late, in part or not at all. A hidden asset is seen only when another player’s sensor picks it up. We tested this with two runs that differ only by one hidden submarine. The player who cannot see it gets the same words and numbers in both.

Players decide once a turn, and their orders then run day by day inside it. Assets use up stocks such as fuel, munitions or cash, and refill them slowly. A game can set an escalation line: enough clashes in a short time end the run, recorded as escalated to conflict. Players can also talk before they move, and the record shows which deals, promises and threats had something behind them.

Adding the board changed nothing in the old runs.

A recorded world is one complete run of a simulation. We replayed 9,304 of them with and without the new parts, including runs of the simulations behind reports 002, 009, 010 and 011. Every log and final state came out the same. The board’s market layer gives the same results as our existing market model on every turn, and its opinion layer gives the same values as report 011’s Australian opinion index on all 81,000 turns. These are consistency checks. They show the new code agrees with the old. They do not test the board against real events.

The board is built, tested and in the engine since 22 October 2026. No published report uses it yet. The first will be a second version of report 011. Combat stays out. More on Methods, and the engine changelog.

Whether the players are any good is a separate question

None of the above says whether an agent seated as a regulator behaves anything like a regulator. That is the question the whole exercise rests on. It has its own page, its own tests and its own failures.

In short: a reader can tell the characters apart, and their personalities change what they decide, not only how they write. Whether that makes them accurate about real behaviour is still an open question. Our first test of it was too blunt to read either way, and we withdrew the headline we had put on it. Nobody should claim accuracy in either direction, including us.

Read the validation page for what worked, what broke, and the fault we found in our own product. The four passed studies each have a full technical report behind them:

One more limit. The machine can run this table one hundred times in an afternoon. It still cannot sit your leadership team in the seats and make them defend their assumptions to each other out loud. That is the tabletop, where your team plays the seats and the engine referees, and the difference is the one between reading a fire drill and running one.

The annex · Deep

How a stated chance becomes a roll, and what an event carries.

How a stated chance becomes a roll

A seat states the odds it thinks its move deserves. The engine turns those odds into a target on a die by one fixed table, the same in every game. The arbiter can adjust the target, and every adjustment is itemised in the record. The engine rolls, and the roll, the target and each adjustment are recorded with the move. Reports refereed by AI used this standard mode. Own-difficulty mode, where the arbiter sets the target from its own reading of the move, is an experiment used only in studies P2 and P3.

What an event carries

The log is the run. Everything else in a report derives from it.

EventWhat it records
Every eventAn id, the beat (the turn) it belongs to, a timestamp for audit, and the branch (which timeline) it sits on. The original timeline is called “main”. A what-if replay gets its own label.
A claimThe seat, the claim in one declarative sentence, the odds that seat argues for it, the arguments for, and the arguments the other seats put against it.
A rulingThe decision in plain terms, the reasoning path, the assumptions the ruling rests on, and a self-reported confidence. Also the inputs weighed, which arbiter mode produced it, and whether it needed human sign-off.
A resolved checkThe target, the die face or faces, the itemised modifiers, the total, the outcome, and the margin by which it cleared or missed.
A reasoning journalWhat the agent considered, what it rejected and why, and how confident it was in the line it took.
A human overrideWho intervened, what they did, why, and which event they overrode. An override is an event like any other.

Re-run any world with its seed and you get it back, roll for roll. That is what reproducible means here, and it is all it means.