Everybody Knows

About this model

For this game, the first playable should be brutally small:

3–5 players. 6 rounds. 12 event nodes. 5 asymmetric factions: The Registry, The Broadcast, The Curators, The Skeptics, The Synthesists. Two public clocks: Consensus and Divergence. One danger clock: Heat. Three player resources: Evidence, Influence, and Credibility. One private card per player: an Anchor, meaning the specific reality-fragment they want preserved or rewritten.

Each round should have this shape:

Reveal an anomaly card. It hits one or two event nodes and raises pressure.

Players each take two actions. The first version only needs five actions: Investigate, Seed, Stabilize, Publish, and Suppress.

Then comes the Consensus phase. Players secretly commit Influence, reveal, argue briefly, and resolve contested nodes. This is where the table politics lives.

Then score, check Heat, and move to the next round.

The key is that every action must have a visible conversion rate. Something like:

Investigate: gain 2 Evidence from a node. Seed: spend 2 Evidence to add 2 Divergence to a node. Stabilize: spend 2 Influence to remove 2 Divergence or lock 1 Consensus. Publish: spend 1 Evidence and 1 Credibility to score 1 and push a node toward public Consensus. Suppress: spend 2 Credibility to cancel one other player’s unresolved effect, but gain 1 Heat.

That gives you the first balance spine:

1 action ≈ 2 raw resource 2 raw resources ≈ 1 board-state shift 1 board-state shift ≈ 1 scoring opportunity suppression/counterplay must cost tempo or add Heat

That is not final math. It is a calibration rig. You want numbers simple enough that wrongness is visible.

The tuning loop should be this:

First, make the game terminate reliably. In early tests, ignore faction balance. Ask only: does it end in 6 rounds, does Heat matter, and does the board state change enough? If games end with nothing resolved, actions are too weak. If games collapse by round 3, anomaly pressure or Heat is too strong.

Second, make all five basic actions worth taking. If Investigate is always correct, resources are too scarce or conversion is too clean. If Suppress is never used, counterplay is too expensive. If Seed dominates, Divergence is too easy to create or too hard to unwind. If Publish dominates, scoring is too direct.

Third, make the factions crooked but not broken. Each faction should cheat one part of the economy:

The Registry gets cheaper Stabilize. The Broadcast gets stronger Publish. The Curators get better Anchor protection. The Skeptics get cheaper Suppress with less Heat. The Synthesists get better conversion between Consensus and Divergence.

Do not give each faction an entire subsystem at first. One broken lever per faction. Later, add texture.

Fourth, simulate. Start with dumb bots, not clever ones. Random legal agents. Greedy score agents. Heat-avoidant agents. Saboteur agents. You are not trying to prove the game is good. You are trying to find stupid exploits quickly.

Track these metrics:

Faction win rate. For five factions, rough target is 20% each, but early anything between 12–28% is fine.

Game-end timing. Most games should reach round 6. Some early collapses are fine, but not many.

Action pick rate. No action should be below about 8–10% usage after players understand the game. No action should sit above 40–50% unless it is the core action.

Runaway index. The leader after round 3 should not win more than about 55–65% of the time.

Comeback paths. A behind player needs at least one high-risk way to matter late.

Heat pressure. Heat should force compromise, not merely punish fun.

Fifth, use human playtests only after the machine tests stop finding obvious garbage. Human tests are expensive. Use them for table feel, kingmaking, negotiation, confusion, spite, and delight. Do not burn humans discovering that “Publish scores too many points.”

A reasonable tool stack is:

Machinations for visual stock-flow models and Monte Carlo sanity checks. Google Sheets or Excel for card/action tables and cost curves. Python for the actual simulator. LLM self-play later, once the rules are formal enough. There is current research using multi-agent LLM self-play plus Bayesian optimization to tune asymmetric games, but I would treat that as a power tool, not a foundation.

My strong read: the first version of Everyone Knows should not be balanced for elegance. It should be balanced for motion.

Every round, something should become publicly harder to deny. Every player should have one lever that feels unfair. Every unfair lever should create a liability somewhere else. The game should feel like trying to repair a map while people argue over whether the map was ever true.

That is the playable core. Balance comes after that, but not much after.

"Everybody Knows" is a public Machinations diagram shared in the Community by Alexander Templeton, with 28 components.

It was created in June 2026 and last edited in June 2026.

Open the diagram in the Machinations app and press Play to watch the resources move, adjust its values to test how the system responds, or remix it to make an editable copy in My diagrams.