v0.4: Data-Driven Balance — 1 Million Simulated Games, Cooking Labs, and Teammates

This release is the most ambitious update to Drug Warz AI yet. We didn't just add features — we built the tools to scientifically measure whether the game is actually fun, then used the data to fix it.

Here's what shipped:

• Cooking/Manufacturing System — buy precursors, cook drugs in labs, upgrade purity
• Teammates — hire runners, enforcers, chemists, and lookouts
• Data-Driven Balance Retuning — 1,000,000 simulated games identified and fixed three critical problems
• 304 unit tests and 3 E2E test suites with Playwright
• TypeScript strict mode across the entire codebase

The Simulation Tool

The star of this release is tools/sim.ts — a headless game simulation CLI that runs the full game engine without any UI. It was built to answer a question we couldn't answer by playtesting: is the game actually balanced?

How It Works

The simulation tool imports the real game engine functions (buyDrug, sellDrug, travel, combatAction, etc.) and runs them through 10 AI strategies, each representing a different playstyle:

Strategy	Description
random	Buys and sells randomly. The control group.
greedy	Always buys the cheapest drug, sells the most expensive. No survival instinct.
fentanyl-rush	YOLO into fentanyl. Maximum risk, maximum reward.
stay-and-grind	Never travels. Trades local market only.
banker	Deposits cash aggressively, trades conservatively.
property-mogul	Saves up for properties, builds passive income.
heat-manager	Buys every heat reduction item, manages federal attention.
dealer-network	Hires NPC dealers for bulk discounts.
optimal-trader	Buys low, sells high, tracks price history. The "smart" player.
marathon	Uses every feature: bank, properties, teammates, heat items, dealers.

Each strategy runs 50,000 games per mode (Classic and Endless), totaling 1,000,000 simulated games. The tool uses a seeded PRNG (Mulberry32) to make results reproducible — same seed, same outcomes.

CLI Flags

npm run sim                      # Default: 5,000 games/strategy
npm run sim -- --games 50000     # 1M total (50k × 10 strategies × 2 modes)
npm run sim -- --seed 123        # Custom seed for reproducibility
npm run sim -- --endless-days 500 # Extend endless mode cap

What It Measures

For each game, the tool tracks:

• Final net worth (cash + bank + inventory value - debt)
• Peak net worth reached during the game
• Death cause (combat, loan shark, government strike, timeout)
• Heat metrics (peak heat, government strikes, heat items purchased)
• Feature adoption (did the strategy use bank? properties? teammates?)
• Economy metrics (total drugs moved, profit, most traded drug)

The output is a structured markdown report with tables, rankings, and balance flags — automated warnings when metrics look unhealthy:

- greedy (classic): 100.0% death rate — near-guaranteed death
- stay-and-grind (classic): 85.1% of games hit heat ≥90 — heat may be too aggressive
- marathon (endless, 365 days): 0% reached $1B — economy ceiling may be too low

The full simulation runs in about 2 minutes on a MacBook Pro (1M games, 122.8 seconds).

What the Data Revealed

Running 1,000,000 games exposed three critical problems:

Problem 1: Combat Kills Everyone

100% of deaths were from combat. Zero from loan sharks. Zero from government strikes. Even the best strategy (heat-manager) died 65% of the time in Classic mode. In Endless mode, every strategy hit 100% death rate.

The encounter formula was dangerLevel * 0.03, meaning Detroit (danger level 9) had a 27% encounter chance per travel regardless of how carefully you played. Low-level police did 12 damage per hit — three unlucky fights and you're dead.

Problem 2: Heat Spirals Uncontrollably

62-84% of actively-trading games hit heat ≥90 (the danger zone). Even heat-manager — a strategy that buys every heat reduction item — still hit 87% heat-≥90 rate. The decay of -3 per day couldn't keep up with the +5 to +15 heat generated by each transaction.

Problem 3: Economy Ceiling at ~$20M

Zero games out of 1,000,000 reached $100M. The economy had a hard ceiling:

• Bank charged 30% on deposits and paid 0% interest
• Properties generated $200-$3,000/day (takes 167 days for a Mansion to pay itself back)
• Price spikes capped at 4x the drug's max price

The Balance Fixes

Armed with data, we made surgical changes to each system.

Combat: Fewer Encounters, Better Survivability

Before: encounterChance = dangerLevel * 0.03 After: encounterChance = dangerLevel * 0.02 * max(0.3, heat/100)

The new formula ties encounter frequency to heat. At low heat, you get 70% fewer encounters. At heat 50, encounters are halved. At heat 100, it's close to the old rate. This means careful players who manage their heat profile can trade safely, while reckless players still face consequences.

Other combat changes:

• Starting weapon upgraded from fists (5 dmg) to switchblade (12 dmg)
• Police damage reduced from 12 to 8, flee chance raised to 85%
• Street mugger damage reduced from 10 to 7, flee chance raised to 90%
• State Troopers and Gang Ambush flee chances raised from 50% to 60%

Heat: Slower Burn, Faster Recovery

Setting	Before	After
Daily heat decay	-3	-5
Fentanyl heat per trade	15	10
Cocaine heat per trade	7	5
Heroin/Meth heat per trade	8	6
Large inventory penalty	+2/day	+1/day
Property heat (≥$500K)	+3/day	+2/day
Heat item cooldown	3 days	2 days
Diminishing returns floor	30%	40%
US strike threshold	heat ≥85	heat ≥92
International strike multiplier	1.0	0.5

Economy: Bank Interest, Better Properties, Higher Ceiling

Setting	Before	After
Bank deposit fee	30%	20%
Bank interest	0%	1%/day compounding
Price ceiling multiplier	4x	6x
Trap House income	$200/day	$500/day
Warehouse income	$800/day	$2,000/day
Mansion income	$3,000/day	$8,000/day
Cook Lab income	$500/day	$1,200/day

The bank interest is the game-changer. Deposit $100K (pay 20% fee = $80K). By day 30 it's $107K. By day 100 it's $216K. By day 365 it's $2.94M. Combined with Mansion income ($8K/day), the path to $100M+ in Endless mode is now real.

Results: Before vs. After

We re-ran the full 1M simulation after the balance changes:

Metric	Before	After	Change
Best median NW (Classic)	$315K	$558K	+77%
$1M milestone rate	15-43%	43-65%	+20-22pp
Max NW (Endless)	$22.7M	$121.6M	+435%
Death rate (heat-manager, Classic)	70.3%	65.3%	-5pp
Death rate (banker, Classic)	90.1%	85.1%	-5pp
heat-manager $10M milestone	2.1%	6.4%	+4.3pp

The game is still deadly — that's intentional. But now smart strategies are meaningfully rewarded, the economy has room to grow, and the heat system actually responds to player choices.

New Features

Cooking/Manufacturing System

Buy precursors at market prices, cook them in upgradeable labs, and produce higher-purity product. Six cookable drug recipes, each with its own precursor chain. Lab upgrades improve yield and purity, but disasters can destroy batches and damage your health.

Teammates

Renamed from "plugs" to better reflect the system. Hire up to 4 teammate types:

• Runners — sell drugs passively while you travel
• Enforcers — combat damage bonus, reduce incoming damage
• Chemists — cooking yield bonus, reduce disaster rate
• Lookouts — heat reduction, early warning on encounters

Testing Infrastructure

304 Unit Tests with Vitest

Every game system has dedicated test coverage:

Test Suite	Tests	What It Covers
game-trading.test.ts	42	Buy/sell, purity, price ceilings, bulk deals
game-combat.test.ts	28	Damage calculation, flee mechanics, death
game-finance.test.ts	35	Bank deposits, loans, interest, debt
game-travel.test.ts	31	Travel, heat decay, property income, events
game-shop.test.ts	24	Weapons, armor, heat items, cooldowns
game-teammates.test.ts	22	Hire, assign, collect, auto-assign
game-cooking.test.ts	18	Precursors, labs, cook jobs, disasters
game-getters.test.ts	15	Net worth, inventory, score calculation
data.test.ts	12	Data integrity, drug/location/weapon validation
ui-rendering.test.ts	11	DOM rendering, HUD display, nav tabs
simulation/simulate.test.ts	6	Balance check — 4 strategies, 100 games each

Tests run in ~400ms locally using Vitest with the V8 coverage provider.

npm test              # Run all 304 tests
npm run test:coverage # With code coverage report

3 E2E Suites with Playwright

End-to-end tests using Playwright verify the actual browser experience:

• Navigation (9 tests) — Every tab switches correctly
• Market Trading (2 tests) — Buy drugs, verify cash decreases, error toasts
• Game Flow (3 tests) — Start game, travel, complete a session

Plus the screenshot spec that generated every image in this post.

npx playwright test   # Run all E2E tests

Deterministic Test State

A custom createTestState() factory (state-factory.ts) generates deterministic game states with fixed prices, purity values, and starting conditions. Zero Math.random() calls — every test is reproducible.

The Map Coordinate Tool

The travel map shows 19 cities across 4 regions, plotted on a CartoDB Dark Matter basemap. Getting the coordinates right required a custom tool: scripts/map-coords.html.

It's a standalone HTML page that:

1. Loads the game's world map image
2. Displays a crosshair cursor with live percentage coordinates
3. Lists all 19 cities in a sidebar
4. Click a city, click the map — coordinates are recorded
5. Exports a ready-to-paste TypeScript LOCATION_COORDS object

The coordinates are stored as percentage-based { x, y } values (not pixels), so they scale correctly at any viewport size. This approach means the map works on mobile without any responsive breakpoint logic.

Tech Stack

This entire game is built with zero frameworks:

• TypeScript — Strict mode, ES2022 target, erasable syntax
• Eleventy (11ty) — Static site generation for landing, history, blog, i18n
• Vitest — Unit testing (304 tests, ~400ms)
• Playwright — End-to-end browser testing
• anime.js v4 — Screen transitions and particle animations
• CartoDB Dark Matter — Map basemap tiles
• IBM Plex Mono — Monospace font for the terminal aesthetic
• GoatCounter — Privacy-friendly analytics
• Cloudflare Pages — Hosting and CDN
• Claude Code — AI pair programming (built this entire release)

No React. No Next.js. No Webpack. No Tailwind. Just ES modules, CSS custom properties, Canvas API, and Web Audio API.

What's Next

The simulation tool opens up a whole new development workflow: change a number, run 1M games, see what happens. Future balance patches will follow this data-driven approach.

Planned for v0.5:

• Leaderboard system (global high scores)
• More cooking recipes and precursor chains
• Expanded Endless mode with prestige mechanics
• Mobile-optimized touch controls

Play Drug Warz AI now and see how your strategy stacks up against 1,000,000 simulated games.

---

Built with Claude Code. Tested with Vitest and Playwright. Hosted on Cloudflare Pages.