Dread Majesty

Fortresses raise Dark Legions. Legions take ground that becomes Warrens.
Warrens breed Minions. Minions generate Evil.

Generators produce other generators, so production compounds.
That cascade is the game, and showing it is the point of the interface.

Play it →

---

The chain

Five rungs. Each one spends a cycle and pays out in the rung below it, so buying at the top pays at the bottom several steps later.

	Tier	One cycle pays	Cycle	First one costs
	Thrones	1 Fortress	90m	1.6e12 Evil
	Fortresses	1 Dark Legion	30m	6e8 Evil
	Dark Legions	1 Warren	10m	6e7 Evil
	Warrens	5 Minions	60s	3,000 Evil
	Minions	5 Evil	4s	40 Evil
	Evil	—	—	what everything is bought with

Nothing runs until somebody makes it run. Every tier starts manual: you rouse it from its node on the chain, it runs one cycle, pays out and stops. Appointing that tier's Overseer, for Evil, automates it for good. That is the opening loop, and it is bought off one tier at a time.

---

Running it

pnpm install
pnpm dev        # http://localhost:5173
pnpm check      # typecheck + lint + test
pnpm build      # production bundle
pnpm harness    # headless balance run

Node 22 or newer.

---

Layout

flowchart LR
    web["<b>apps/web</b><br/>React + Vite<br/>renders state, sends intents"]
    engine["<b>packages/engine</b><br/>pure TypeScript<br/>no DOM, no React, no I/O"]
    types["<b>packages/content</b><br/>balance, copy, art slots"]

    web --> engine
    engine -.->|types and ids only| types
    web --> types

    style web fill:#1e1a26,stroke:#c9a227,color:#ece8f1
    style engine fill:#1e1a26,stroke:#ea580c,color:#ece8f1
    style types fill:#1e1a26,stroke:#6b8f5e,color:#ece8f1

Loading

Dependencies flow one way. The engine takes the content types and the id vocabulary; the balance numbers arrive as a function argument, and lint bars the engine from importing them. Engine tests run against fixtures, never against shipping content, so a balance change cannot fail an engine test.

---

How it works

The engine is pure

No DOM, no React, no I/O, no clock. Time and seeds enter as arguments at the boundary. That single rule buys most of what follows: every engine test is a plain assertion, and replay verification is available later for free.

One simulation, fixed timestep

Elapsed milliseconds are banked and spent in whole slices, so a cycle finishing depends on how much time passed, never on how the browser scheduled frames. Coming back after four hours runs the same step the same number of times as sitting there for four hours would. There is no second code path for offline, no aggregate shortcut, no closed form.

flowchart LR
    bank["elapsed ms<br/>banked"] --> slice{"a whole<br/>slice?"}
    slice -->|yes| read["read the<br/>snapshot"]
    read --> write["write into<br/>a delta"]
    write --> commit["commit at<br/>slice end"]
    commit --> bank
    slice -->|no| bank

    style bank fill:#1e1a26,stroke:#2b2536,color:#ece8f1
    style slice fill:#16131c,stroke:#c9a227,color:#ece8f1
    style read fill:#1e1a26,stroke:#2b2536,color:#ece8f1
    style write fill:#1e1a26,stroke:#2b2536,color:#ece8f1
    style commit fill:#1e1a26,stroke:#ea580c,color:#ece8f1

Loading

Nothing produced within a slice affects anything else within it. That is why tier order does not matter and why there is no tie-break rule to get wrong.

Every count is a Decimal

break_eternity.js, never a JS number, not even early on when the values are small. The game reaches 1e30 in a week of play and past 1e300 eventually. Mixed arithmetic is the bug you find six months later.

Saves migrate, and shipped steps are never edited

Ten versions so far, each a pure function from one blob shape to the next, applied in sequence. A mistake gets another step appended rather than a fix in place, because there is no way to tell which saves already passed through the old one. Version 10 exists for exactly that reason.

Balance is measured, not guessed

pnpm harness runs the real engine headless for seven simulated days and reports when each tier arrives, when each goes obsolete, what the growth exponent is, and whether the prestige loop converges or runs away. Every number in packages/content was fitted against it.

It earns its keep on the failures nothing else catches. A prestige curve that read as balanced diverged to ×634 after two resets. Its replacement was correct in every figure but had lost its zero, and paid 213 souls for thirty-four Evil. Neither was visible by reading the code.

The harness is a script and never gates CI — it is a measuring instrument, not a test.

---

The first run

The front door	The tour

The title screen shows only on a genuinely fresh run. Anything with a save behind it gets the return summary instead.

Then a line at the foot of the screen names the next thing to do and holds back every other control until it is done, then clears and leaves the player alone until the next moment worth teaching. Six moments are gated that way. A seventh gates nothing and simply remarks on the Minions the first Warren delivered without being asked, which is the whole cascade in one sentence. The tour is skippable at the first prompt and never returns.

The first time you smite, something else starts talking. It wants you to do it again immediately, which is the wrong move, and it argues better the longer you refuse.

---

Where things stand

M1 through M4 have had a first pass. The game is playable end to end.

The engine is complete and tested: fixed-timestep simulation, an exact cost curve and max-buy, purchases, manual cycles and Overseers, milestone and prestige multipliers, achievements, unlock latching, offline catch-up, and versioned saves.

apps/web is the designed interface, not a shell — the live chain diagram, the buy rail with exactly one accented spend, the prestige panel, the deeds wall, the ledger, and the offline-return screen. Saves go to IndexedDB and export as a pasteable blob. Sound is synthesized in code and muted until asked for.

There is a dev workbench at the foot of the page: jump to any point of progression, set any resource, appoint or dismiss every Overseer, simulate an absence. It is stripped from production builds and it deliberately looks nothing like the game.

Every player-facing string lives in packages/content/src/v1/copy.ts. Adding an achievement without copy fails typecheck.

Economy tuning is ongoing. The numbers have had several measured passes, and the current ones are the best answer so far rather than a finished one: Warrens at 3 minutes, Dark Legions at 14 minutes, Fortresses at 1h11m, Thrones at 2h12m, first prestige at 34 minutes. The first Minion is bought by hand at 33 seconds, eight taps in. Expect them to move. They want real players more than another harness run. Re-run pnpm harness after touching any number in packages/content.

Not done: real art. Every slot renders a generated SVG fallback — including the marks in this file, which are the same drawings the game ships.

---

Reading order

1. The design spec — the decisions and why each one went the way it did. Everything else amends it.
2. CLAUDE.md — how to write code here. The five engine rules, in full.
3. docs/ui-sensibility.md — the interface rules the genre usually breaks, and why this one keeps them. Normative for apps/web.
4. docs/ — every spec and plan, indexed, with what each one still governs.

---

Using this

No license, on purpose. The source is here to be read, not reused: all rights reserved. Read it, learn from it, quote it with attribution. If you want to do something else with it, ask.