TycoonGlobal

Design

Simulating a Thousand Companies for Ninety Days

A deterministic Monte Carlo validator that found five behavioural bugs, two tests that could not detect failure, and one scale law that invalidates every small test run.

Design8 min readPublished 27 Oct 2026Updated 15 Nov 2026#design#simulation#engineering

You cannot playtest an economy with a hundred people and learn anything about how it behaves with a hundred thousand. So the economy gets its own simulator: 1,000 agents, 90 simulated days, 2,160 hourly ticks, deterministic from a fixed seed, roughly 14 seconds of wall clock.

It has been more valuable for what it revealed about our tests than about our economy.

Six gates

The validator asserts six properties. An early run of the current set:

GateTargetMeasured
(a) weekly money-supply change at equilibrium< 2%31.13%fail
(b) faucet/sink ratio0.93—1.070.948pass
(c) median newcomer day-7 net worth growth> 0+54.97%pass
(d) megacorp share of raw inputs bought> 65%86.74%pass
(e) money supply retained0.5x—2.0x0.080fail
(f) median agent wage coverat least 4h0.0fail

Three of the four original gates passed. The economy they certified had lost 92% of its money supply, and the median company could not pay one hour of wages.

Two tests that could not fail

Gate (a) could not see deflation. It reduced over a weekly series seeded at 0, so no negative week could ever become the maximum. A world collapsing by 92% reported 0.00%. It is now symmetric, and measured at equilibrium rather than across the growth phase.

Gate (b) is a ratio. A faucet and a sink that both shrink toward nothing still divide to roughly one. Over the final 30 days it read 0.948 — healthy-looking — while reading 0.471 across the whole run.

Gates (e) and (f) were added because no ratio can detect this alone. Gate (f) asks the only question a player would ask: can the median company make payroll?

The general lesson generalises past this project. A metric that cannot move in the direction of failure is not a test. Both broken gates were plausible, well-intentioned and had been green for weeks.

Five behavioural bugs

Every fix turned out to be about the simulated operator, not a game constant.

Retail demand was invented, and 10x to 80x too small. The simulator used hand-picked constants (LAPTOP: 0.15) while the game derives demand from population, income-driven per-capita consumption and seasonality (LAPTOP: 12.0). It was validating an economy the game does not have. It now calls the real demand model directly.

Wages were charged at zero utilisation. An input-starved plant paid a full shift to produce nothing — deadweight loss with no counterparty, and the largest single driver of the deflation. Plants now idle when unfed and pay pro-rata on partial runs.

Agents reinvested to zero cash. The investment routine spent down to nothing every tick. Agents now hold 24 hours of wage bill as working capital before expanding.

Agents produced at nameplate regardless of demand, flooding intermediates until prices pinned at the 0.25x floor. Production now plans against a smoothed sales rate plus the agent's own downstream consumption — counting only external sales starves a vertically integrated chain, because a mill feeding its own bakery records no sales and throttles to zero. New lines start optimistic at capacity and decay toward realised demand, because starting pessimistic deadlocks the chain: every agent's input is another agent's throttled output.

Agents expanded while loss-making. A firm with negative smoothed net income no longer invests.

Results: weekly change 31.13% to 0.16%, retention 0.080 to 0.919, median wage cover 0.0h to 72.3h.

The scale law

The most important single finding, and the most inconvenient.

Retail demand scales as N^0.85 while agent count scales linearly. Per-agent earnings therefore fall as the world grows. The same economy inflates at small scale and deflates at large: 200 agents produced +109% weekly growth; 1,000 agents produced -92% over the run.

Any scaled-down run validates a different economy. A fast green test on 200 agents would have been actively misleading. The test mix now matches the real configuration, which costs wall-clock and is worth it.

Fixing the underlying sensitivity means revisiting the demand exponent or the wage curve. It does not mean relaxing the gates.

What the run does validate

The anti-whale brake works, emergently. Megacorps reached a median 111 total building levels — past the Tier-0 break-even landmark of 100 — and bought 91.3% of their Tier-0/Tier-1 inputs rather than making them. Nothing told them to. The agent compares its own unit cost at its own overhead against the market price and declines to build extraction it cannot run profitably.

Newcomers thrive. The median casual player grew net worth +97% over seven days at 0% overhead — the comparative wage advantage the design intends.

The ledger identity holds exactly. Across 2,160 ticks and ~2.9 million transactions, drift is $0. Asserted as a test.

The economy converges. Weekly change runs -15.0%, -3.7%, -2.9%, -2.5%, -3.2%, -3.6%, -3.1%, -3.7%, -3.3%, -3.0%, -2.8%, -1.3% — a front-loaded contraction as starting cash converts into buildings and inventory, then a monotone approach to equilibrium. No runaway inflation, no collapse.

Known limits

The simulator abstracts what does not affect aggregate faucet/sink behaviour. It uses a single global clearing price per SKU rather than the per-city limit order book, and it does not model freight, B2B contracts, quality tiers, biome bonuses, the macro cycle, or loans and receivership.

Order-book microstructure does not change money-supply behaviour. The four missing sinks do — R&D patents, loan interest, Harberger lease taxes and freight fuel, budgeted together at roughly 20—30% of total drain. Their absence is why measured wage share sits at 86% against a 25—35% design budget, and it is why the wage constant is not being retuned to chase that number.

Two gates remain outside band on the final-30-day window: 4.51% weekly change and 0.905 faucet/sink. That gap is open work, stated here rather than rounded away.