The fair, on-theme edge lever for slow chains (from the calibration work): a run
now descends `maxDepth` levels instead of one map, so cumulative run length makes
the block-timer bite WITHOUT a giant single dungeon or a cheating-fast monster.
Engine (game.js):
- depth/maxDepth; _generateLevel(depth) builds each level (fresh dungeon + monster
+ FOV), _descend() steps down. Reaching a non-final exit descends (event
"descend", stays active); only the final exit escapes (event "escaped").
- Treasure lives only in the vault (final level) — intermediate levels race to the
stairs. Single-level (maxDepth=1) is byte-identical to before.
- Provably fair per level: levelSeed(masterSeed, depth) (provablyFair.js) — level 1
= master seed, deeper levels salted; the whole descent regenerates from the one
committed seed. Verified: game levels == regenerateFromSeed(levelSeed(seed,d)).
Pacing (difficultyConfig.js): NETWORK_TUNING now sets LEVELS ∝ block time (GRIN 1,
XMR 2, LTC 2, WOW 4, BTC 8) — replacing the size-scale hack; size + monster stay at
the fair preset. DUNGEON_LEVELS operator override; NETWORK_TUNING_DISABLED → 1 level.
Wiring: socketHandlers afterPlayerMove gets a "descend" branch (no game-over, no
monster move on arrival). Client gameState.js clears the accumulated map/explored/
visible tiles on a depth change so the old level does not ghost through the new one.
Sim: explorer bot resets fog-of-war on descent so multi-level sims stay honest
(full re-calibration of the level counts is the pending follow-up).
Tests: multiLevel_ultra (descent, vault-only treasure, per-level provably-fair,
single-level default, DUNGEON_LEVELS) + networkTuning rewritten for the levels model.
Suite 61 suites / 432 green; end-to-end 4-level WOW run descends 3x then escapes.
The instrument the design discussion said was missing. src/sim/ drives the actual
game engine (game.js — same dungeon gen, monster AI, movement as prod) with
headless bots to terminal outcomes, then measures what was only ever DECLARED:
- Per preset: escape/treasure/caught/stuck rates + completion-move/-time pctiles.
- MEASURED house-win-rate per network, integrating out the random block deadline
analytically: for D~Exp(mean=blockTime), P(house wins an escape at T)=1-exp(-T/BT),
so no Monte-Carlo over deadlines.
- Two bot skill models (omniscient beeline / explorer fog-of-war) as an upper band;
neither evades the monster, so true house-win is likely below both.
First findings (explorer, 200 runs, 320ms/move): presets UNDERSHOOT their declared
targets (casino 57% vs 70%, normal 47% vs 55% on WOW) and the edge swings ~15pts
across chains from IDENTICAL dungeons (casino 48% BTC .. 65% GRIN) — because
cryptoType never actually sizes the dungeon. Exactly the two gaps we predicted.
Files: src/sim/{pathfind,simBots,simulate}.js + test/sim_ultra.test.js (structural
guard, not a balance assertion). Run: node src/sim/simulate.js --help-ish flags.
Suite 59/422 green.
jwinterm