Three rungs, one method

The money games are a ladder.

Most of what we make is one question asked at three different altitudes: you have to decide before you can see, and then you live inside it. What changes from rung to rung is not the difficulty — it is what the money is for.

  1. 01

    Kitchen Table

    Live · play it now$1 – $500 spare a month

    “What gets paid this month?”

    A year at one table on an income that does not arrive evenly. Four real years, built on actual price and rate data, and no move that makes any of them comfortable. What it scores is whether you kept a cushion — not whether you got rich.

    Play Kitchen Table
  2. 02

    The Library

    In the workshop$500 – $10,000 spare a month

    “What is the surplus actually for?”

    The rung where there is money left over and the question changes shape. Five ways of deciding, run against the same fifteen years, with the arithmetic shown rather than summarised. It is called the Library because this is the level where the papers live — every series on screen with its provenance.

    Read where it is up to
  3. 03

    TheHeir

    In design$5m+

    “What survives you?”

    Three generations, not fifteen years. You decide what the money is for, then you build something to outlive your own judgement, and then you watch an heir spend it while you are advisory at best. The mechanic is that you stop being the decision-maker halfway through — which is the one thing a thousand allocation calculators cannot do.

    Nothing to play yet — the brief is written, the build is not started.

The rungs overlap at their edges the way real money does, and that is intentional. Everything else in the cabinet — the dominoes, the cards, the Fed — is the same studio asking the same question with the money taken out. More about how we work.

The recommendation

Tonight's Table

Wednesday 18 November· Dealer's choice

The Library

Five things money can be asked to do. Most games model one.

Short slice · Capital sim · Unfinished

Take a seat →
The Library
The router

What should I play first?

Three questions. We'll point you at one drawer instead of five.

What's your vibe?

From the workshop

Studio artifacts

Sketches, rules, characters and development notes. The games are the output; this is the working.

Training log

PPO v20 · 4,000 updates

19× faster than v19 on the 3090 — 4.6 hours instead of 34. It also collapsed. The batched-GPU rewrite and what the shortcut cost us.

Rules card

Sutda hand rankings

The nine Gwang, the Ttang, and why 3-8 beats a pair. Printed as a card you can keep open beside the game.

Character sheet

DoubleFives: how the opponent was made, and what is measured

The 2-player neural opponent (v22) learned by playing past copies of itself and scripted bots, about 14 million games (our estimate). Separately, 100.3 billion deals with simple scripted bots checked that no seat has an advantage. In simulated matches v22 beats our strongest rule-based character a little over half the time. All of it is engine against engine, not against people.

Method note

The Chair runs on live FRED

Expectations-augmented Phillips curve, two-to-three quarter policy lag. The papers are linked in the footer.

The Lab

Under the hood, it's not playing around.

The opponents are trained, not scripted. Here is the machine that does it.

How it was made

Algorithm
PPO (custom PyTorch trainer), against past copies of itself and scripted bots
Neural training games
about 14 million (2-player model v22; our estimate)
Scripted-bot deals
100.3 billion (seat-fairness checks and baselines; not training)
Characters
7 (two of them use the neural model; the rest are rule-based styles)

Ladder snapshot: Sep 1, 2026.

Standings on our training ladder

Standings are not available right now.

Neural agents against scripted bots in simulated 4-seat games, rated with Glicko-2 on the trainer's own ladder (a different scale from the character ratings). Engine against engine, not against people.

The Forge

What's being built right now.

Three things on the bench. Nothing here is announced until it works.

AI Coach Mode

DoubleFives

The AI explains its reasoning after every move. Learn why it played that tile, what it counted, and what trap you just walked into.

In Training

Real-time Multiplayer

DoubleFives & Sutda

Play against friends instead of just AI. Same brutal difficulty, now with the added pressure of someone watching you lose.

Design Phase

Historical Playback

www.chairmanmode.app

Replay any Fed decision from 1929-2024. See what Volcker did, then try it yourself. Compare your choices to what actually happened. Learn why you'd have been fired.

Data Sourcing

Want to influence what gets built next?

The House List

Get in before the doors open.

Early access to DoubleFives, updates on new games, and first dibs on beta features. No spam, unsubscribe anytime.

We respect your inbox. Updates only when there's something worth saying.

DoubleFives, in numbers we can source

100.3 billiondeals were simulated with simple scripted bots (finished 20 Oct 2026) to check that no seat has an advantage and to set a baseline.

The neural opponent (v22, 2 players) learned by playing past copies of itself and scripted bots, about 14 million games (our estimate). Opponents are rated on a Glicko-2 ladder of simulated matches, not against people.

Contact

Get in touch.

Sponsorships, partnerships, press inquiries, or just want to say hi.