reikhelm/deepfall

DEEPFALL

Reikhelm, the deep hold, is sinking into the Maw. Delve fast, hold ground, and know when to die. The deepest grave is the brightest glory.

A tabletop game design + a balance-by-bot simulator, built as a self-contained experiment in the reikhelm repo (sibling to combat/). It mashes up nine games the designer loves and tries to keep their best parts while filing off their worst.

The dungeon you delve is not invented — it is a real reikhelm dungeon (reikhelm-core's room/theme/graph model), so a physical printing could ship seeded, reproducible boards straight out of the generator.

---

The one-paragraph pitch

You are rival delve-companies racing into a collapsing megadungeon. Each company is a Lineage × Calling smash-up (your race × your class/faction) that defines a bag of dice you build mid-game. You push your luck rolling that bag to descend, fight, crack vaults, and seize rooms — which score for whoever holds them when their floor collapses. The Maw devours the dungeon from the bottom up, so the richest deep rooms die first and the clock is brutal. Every party faces one decision, over and over: cash out (send them "to ground" — they keep their territory as silent dead and you draft a fresh party), press on (greedy, keep the engine, risk the Maw), or go for a Worthy End (dive suicidally deep and let the Maw take you in a blaze for the single biggest payout in the game). Most glory wins.

The nine games, and what we took / what we fixed

Game	What we kept	What we filed off
Smash Up	Two-half faction mashup; control rooms (bases) that score at a breakpoint	The "take-that" hand hate & swing — conflict here is positional, never "discard your stuff"
Small World	Race×Power combo drafting; the Decline mechanic (its single best idea)	Crowded-map turn-order tyranny — the Maw, not opponents, is the pressure
Blood Rage	Card-drafted strategy per age; escalating doom (Ragnarök); dying is glorious	Coin-flip combat reveals — outcomes here are dice you build and mitigate
Quarriors	Dice-bag building; in-game engine growth; push-your-luck	"Dice gods hate me" whiffs — the Decline safety valve + Faith mitigation soften variance
Dragon Rampage	Push-your-luck escape tension; one escalating shared threat	Player elimination — nobody is ever out; a dead party becomes glory
Catan	Tile production when you hold ground; spatial placement	The robber feel-bad & resource starvation — the Maw is the only "robber," and it's fair to all
EverQuest	Holy-trinity party comp (Steel/Cunning/Faith); faction reputation; loot chase	The grind & downtime — turns are fast, simultaneous-ish, no camping
Daggerfall	Custom class via Lineage×Calling with advantages and disadvantages; procedural world; faction web	Sprawl & getting lost — bounded floors, a shared revealed board
Eye of the Beholder	Grid dungeon of room tiles; party of four; mechanism puzzles gate the best loot	Mapping tedium — the board is shared and revealed, not memorized

The load-bearing fusion: Small World's Decline + Blood Rage's glorious death. That one combination kills the runaway leader, abolishes player elimination, and inverts loss-aversion — three of these games' worst problems, solved at once.

Layout (mirrors combat/)

deepfall/
  README.md            — this file
  DESIGN.md            — the full rules + the simulation model (the spec the sim implements)
  Cargo.toml           — self-contained workspace (does not touch the root reikhelm graph)
  configs/default.json — the tunable balance table; editing it IS the balance act
  deepfall-core/       — pure, deterministic, headless game engine (ChaCha8, no I/O)
  deepfall-sim/        — batch harness: runs sweeps → CSV/JSON
  analysis/            — Python (pandas/matplotlib) over the exported data

Build & run

cd deepfall
cargo test                                  # engine determinism + rules tests
cargo run -p deepfall-sim -- --out out      # run the balance sweeps → out/*.csv, out/*.json
python analysis/analyze.py out              # (optional) plots → out/plots/*.png

What the simulator is for

It is not the game — it is a truth serum for the design. Bots play the three exit strategies (Cash-Out, Press-On, Worthy-End) plus an adaptive Balanced bot, across all 36 Lineage×Calling combos, thousands of seeded games. We read the data to answer the only questions that matter for a design like this:

1. Is there a dominant strategy? (We want a rock-paper-scissors band, not a king.)
2. Are any Lineage×Calling combos broken? (Asymmetry should be fair, not solved.)
3. Does the Maw end the game in a satisfying window? (~8–14 rounds.)
4. Is it a runaway? (Margin of victory should be tight.)
5. How often does a party die for nothing? (The feel-bad rate — keep it low.)
6. Is the Decline decision real? (Do winners vary their cash-out timing, or is there one provably-correct moment?)

See DESIGN.md §"Simulation model" for exactly how the sim abstracts the rules, and the committed sweep notes for what the data actually said.