Mosaic Lab · Cross-study report

The prism block: one game, four table sizes

complete

4,000,000 complete games at five, four, three and two players — every table anchored by the Prism.

Studies 1–4 played one million games each at a shrinking table. Every empty chair becomes an imaginary player: dealt a goal and a hand like anyone else, but never acting and never winning — goals that sit in the rotation as pure information. So these four corpora ask one question four ways: what does removing real minds from the same game do to it?

Two kinds of answer showed up. Some numbers refuse to move — they are properties of the rules, not the table. Everything else slides monotonically with player count, in one direction, across all four studies.

The four studies

4,000,000
games in this comparison
1st & last
honest tops, liar bottoms — at every size
46.1% → 55.3%
stealth wins, five players → duel
4.6% → 1.5%
first-mover edge over fair share

What table size cannot touch

Normalize every persona's win rate by its table's fair share (1/5, 1/4, 1/3, 1/2) and the four studies collapse onto one ladder. Order fixed by the five-player standings — a crossing would be visible, not resorted away.

Personas

Persona win rate as a multiple of fair share, by table size

1.00× = fair shareThe Honest OneThe Honest One, 5 players: 1.139× fair share1.14×The Honest One, 4 players: 1.151× fair share1.15×The Honest One, 3 players: 1.164× fair share1.16×The Honest One, 2 players: 1.151× fair share1.15×The HunterThe Hunter, 5 players: 1.060× fair share1.06×The Hunter, 4 players: 1.069× fair share1.07×The Hunter, 3 players: 1.067× fair share1.07×The Hunter, 2 players: 1.053× fair share1.05×The HoarderThe Hoarder, 5 players: 1.026× fair share1.03×The Hoarder, 4 players: 1.028× fair share1.03×The Hoarder, 3 players: 1.014× fair share1.01×The Hoarder, 2 players: 0.994× fair share0.99×The SpenderThe Spender, 5 players: 1.012× fair share1.01×The Spender, 4 players: 1.015× fair share1.01×The Spender, 3 players: 1.023× fair share1.02×The Spender, 2 players: 1.027× fair share1.03×The OpportunistThe Opportunist, 5 players: 0.972× fair share0.97×The Opportunist, 4 players: 0.966× fair share0.97×The Opportunist, 3 players: 0.967× fair share0.97×The Opportunist, 2 players: 0.972× fair share0.97×The LiarThe Liar, 5 players: 0.844× fair share0.84×The Liar, 4 players: 0.824× fair share0.82×The Liar, 3 players: 0.815× fair share0.81×The Liar, 2 players: 0.837× fair share0.84×
5 players4 players3 players2 players
Six personas × four table sizes, n=4,000,000 games. Dashed line: exactly fair.

FindingThe hierarchy is structural. The Honest One wins 1.14× / 1.15× / 1.16× / 1.15× its fair share at 5 / 4 / 3 / 2 players — never below 1.14×. The Liar never climbs above 0.84×. Straightforward play beats sustained bluffing by roughly the same margin whether the bluff has four audiences or one — consistent with the pre-registered expectation that bot-vs-bot deception pays nobody, because there are no human minds to deceive.

The one crossingThe Hoarder’s patience stops paying in the duel. At five and four players the Hoarder edges the Spender (1.03× vs 1.01× at five); by three players they have swapped, and at two the Hoarder slips to 0.99× — the only persona whose relative standing moves with table size. Hoarding tempo is a bet that the table stays busy long enough to cash it.

FairnessSeats and goal colours stay flat everywhere: no seat deviates more than 1.6% of fair share in any study, no colour more than 60.2%. The structural-fairness claim of the single-study reports survives the size sweep intact.

Turn order at every size

Win rate by real-player turn order — k bars for k players, imaginary seats hold goals, never turns. The edge is real at every size, and largest where the most opponents wait behind you.

Turn order

5 players

20% = fairMoves 1stMoves 1st: 20.9%20.9%Moves 2ndMoves 2nd: 20.6%20.6%Moves 3rdMoves 3rd: 20.0%20.0%Moves 4thMoves 4th: 19.3%19.3%Moves 5thMoves 5th: 19.2%19.2%

Turn order

4 players

25% = fairMoves 1stMoves 1st: 25.8%25.8%Moves 2ndMoves 2nd: 25.3%25.3%Moves 3rdMoves 3rd: 24.6%24.6%Moves 4thMoves 4th: 24.3%24.3%

Turn order

3 players

33% = fairMoves 1stMoves 1st: 34.1%34.1%Moves 2ndMoves 2nd: 33.3%33.3%Moves 3rdMoves 3rd: 32.6%32.6%

Turn order

2 players

50% = fairMoves 1stMoves 1st: 50.8%50.8%Moves 2ndMoves 2nd: 49.2%49.2%

The three-player quirk

Non-monotoneOne number breaks the pattern: shared victories peak at three players — 0.51% / 0.59% / 0.66% / 0.60% across 5 / 4 / 3 / 2 players, with the maximum at 3. Every shared victory in the block is an exhaustion tie, and the two-imaginary-seat table is where equal-length chains most often survive to the last tile together.

The full table

Every compared metric, all four studies. Per-study detail, methodology and chart provenance live in the individual reports.

Appendix

Metric5 players4 players3 players2 players
Games1,000,0001,000,0001,000,0001,000,000
Mean moves per game22.321.420.318.7
Median moves19181716
Chain endings99.3%99.0%98.6%98.1%
Exhaustion endings0.7%1.0%1.4%1.9%
Shared victories0.51%0.59%0.66%0.60%
Wins ended by an action23.1%24.5%26.6%28.7%
Actions, share of all moves27.2%26.8%26.2%25.4%
Stealth wins (goal unproven)46.1%48.9%51.7%55.3%
Avg opponents who knew the winner’s goal0.910.850.790.70
Goal changes per game5.545.234.834.33
Wins with an acquired goal46.7%45.5%44.4%43.2%
First mover vs fair share1.05×1.03×1.02×1.02×
Avg board area (cells)24.824.123.121.7
Avg tiles left in deck at end23.724.726.027.8

← Mosaic Lab