A framework for simulating whole worlds — real or invented.
A federation of AI personas plays a scenario out turn by turn — one owns the fighting, one the supply lines, one the politics, one the chain of command. An orchestrator who isn't allowed to predict stitches them together. A contrarian who has to disagree attacks the consensus every turn.
Large language models are extraordinary engines of plausible invention. Ask one for facts and that same talent shows up as hallucination. Point it at building a world instead, and invention becomes the whole point. This is a framework that harnesses that generativity — and disciplines it, so the world it produces holds together instead of drifting into nonsense.
The trick is to refuse to let one model improvise the whole story. The situation is broken into domains — the kinetic fight, logistics, infrastructure, the information war, the politics, the chain of command, the slow integrity of institutions — and each domain is handed to its own specialist persona, modeled on a real school of thought. Nobody sees the whole board. The orchestrator works the seams between them and keeps the record. The same machinery runs on a scenario drawn from the real world or one made up from scratch.
How it works, in four moves
A shared way to make sense of the world.
Everything that matters is split into domains. Each domain is a persona named after the thinker whose tradition it draws on — Clausewitz for war, Wohlstetter for warning and surprise, Arendt for institutions. The domains are the sense-making framework; the personas are those domains turned into agents.
Specialists who can't see each other's hand.
Each persona gets only its own slice of the world. They never talk directly. The point of all that separation is that any single analyst, working alone, pattern-matches too fast. The federation is built to make pattern-matching expensive.
A referee forbidden from predicting.
It resolves what happens where domains collide — a strike that hits the grid, a cyber effect that cascades into logistics — and writes down a legible record. It produces ranges and likelihoods, never a single confident forecast.
An agent required to disagree.
Every turn, one persona — codename CASS — gets the draft consensus and a note on which way it's leaning, then argues the opposite. Her objections go into the record alongside the consensus. Disagreement is part of the output, not a problem to be resolved away.
Generative where you want texture; constrained where you want trust. Every estimate is a range, every persona declares its own blind spots, and every turn leaves an audit trail you can walk back. See how a turn actually works, and meet the full roster →
From here down: one run, in costume
Everything below is a single run of that framework, dressed as a declassified intelligence file from the Northrim Group, Division of Federated Assessment — an institution that does not exist. The fiction is deliberate: the costume lets the run read like the kind of document it's modeling. Now that you know it's a costume, the stamps and classification markings are part of the fun.
What you're looking at
One run of the framework, pointed at a near-future geopolitical crisis. The scenario it was handed:
A Persian Gulf maritime crisis with a direct Iran–Israel exchange, under a stipulated compromise of the US executive. Eleven turns, from June 15, 2026 through January 31, 2029.
None of it is a prediction. The starting conditions — the closed Strait of Hormuz, the compromised president, two separate sets of kompromat held by parties who don't share goals but share an interest in keeping him in office — were chosen inputs. The federation's only job was to play them out and see where they led. It got somewhere unexpected: by the close, the question was no longer whether the system would rupture, but what to call the fracture that had hardened into a permanent feature. On January 20, 2029, two presidents took the oath of office.
In its own voice, the file introduces itself like this —
Three things to know before you read a turn
Estimates come as ranges, not points.
Format [low, best guess, high]. The bar below is the federation's read at turn 7 on whether the president stays in office through January 2029. Best guess: about 55%. Defensible range: 40 to 72.
KARL, ROBERTA, CASS — these are AI agents.
Each is a language-model persona running one domain, named for the analytic tradition it draws on — KARL after Clausewitz, ROBERTA after Wohlstetter, CASS the contrarian after Cassandra. No humans were consulted inside the run. Meet the roster →
The scenario is the setup, not a claim.
The compromised president, the kompromat, the closed strait — chosen inputs, not statements about the real world. Dates are simulated time. The interesting part is what those inputs cause across eleven turns.
Three ways to explore the run
By turn
Eleven turn summaries in chronological order. Each opens with a plain-English read of what happened, with the federation's actual end-of-turn briefing in a collapsible underneath. Start here if you want the story as it unfolded.
By thread
Eight curated throughlines: the coalition, the Compact's slow secession, the nuclear ladder, the financial cascade. Each tracks a single question across every turn it touches. Start here if you have a specific question in mind.
By framework
The machine itself: the eight specialists in full, how a turn is run, why every number is a range, and the failure modes the federation discovered about itself. Start here if you care more about the method than this particular world.
The eleven turns
- T0 Jun 15, 2026 The board, set
- T1 Jul 20, 2026 First exchange
- T2 Aug 24, 2026 Things break loudly
- T3 Sep 28, 2026 The kompromat lands publicly
- T4 Nov 10, 2026 Midterm — opposition wins
- T5 Mar 31, 2027 Impeachment #4 fails
- T6 Sep 30, 2027 The wounded predator
- T7 Mar 1, 2028 Chronic crisis as the new normal
- T8 Aug 1, 2028 Primary cycle + DNC
- T9 Dec 1, 2028 The general election
- T10 Jan 31, 2029 Two inaugurations
Click any turn to open its briefing. Dates are simulated time. The interval shrinks at the end — turn 10 covers just two months, because the contrarian convinced the federation at turn 9 that the actual cliff event was five weeks past the standard interval boundary.
Selected threads
The federation produces a briefing per turn. The threads are something it doesn't produce on its own: each one tracks a single question — does the Compact secede? does Iran use the nuke? does the cascade tip over? — across every turn it touches.
The Coalition of the Compromised
The president isn't the only one compromised.
The Compact, quietly seceding
The Governors' Compact started in turn 1 as 10–18 states issuing non-cooperation orders.
The nuclear ladder
Iran started the simulation as a threshold state — capable in principle, untested.
The federation arguing with itself
The most unusual feature of this simulation isn't the scenario — it's that the federation treats its own framing as something to be questioned, every turn.
Standing dissent from CASS
From turn 3 forward, the contrarian made the same argument across the series: the federation's "muddle through" scenario looks structurally like the Compromise of 1877 — the federal government keeps its title, the resisting states keep the actual power, and the cost falls on people who can't relocate. The federation logged it every turn and never modeled the cost. By turn 10 the structural break she'd been describing had become visible: the Compact's parallel financial settlement system carried more than 50% of its commerce, two parallel inaugurations happened on the same day, and the federation had stopped pretending the formal-secession question still mattered.