Randomness has a history

The game-owned global generator, the library shuffle generator, and object-owned generators have separate state. The process also initializes another engine from random_device. Treating all of them as one sequence gives the wrong dependencies.

The global game generator starts with seed 4357. Its raw result has 31 bits. Bounded integers use modulo reduction, while the float helper scales a value and rounds it to float32.¹1The shared game generatorThe game-owned MT19937 starts at 4357, returns its low 31 bits, uses modulo for bounded integers, and divides floating draws by 2147483647.0 before float32 conversion. A branch census finds 38 calls to the global seed wrapper, including 30 that pass time(NULL). Separate object-owned generators and a random_device-seeded process engine also exist. These game reseeds do not reach libc++’s shuffle engine.0x1034ff944 · 0x103608ff4 · 0x1036090b4 · 0x10360921c

Up to build 4.2.4.399, Evolution's result selection instead calls std::random_shuffle. Its library engine is shared with other list shuffles, including Modern portal queues. Resetting the game's global generator does not reset that library engine.²2The separate libc++ shuffle streamThe imported __rs_get and __rs_default symbols identify two-argument std::random_shuffle. Its forward permutation uses uniform_int_distribution with rejection. Live PlayCover captures record raw outputs matching mt19937 seed 5489, complete permutations and selections; the host-library probe is separate evidence. Recorded iPad outcomes corroborate specified routes without identifying the loaded iPad library. Evolution and Modern portals share this library stream.0x1009a9280 · 0x100121044 · 0x10001bf38 Build 4.2.6.401 chooses Evolution results with a generator of the artifact's own, as Evolution's choice in build 4.2.6.401 describes.

Imagine one generator whose next values are assigned to a glove roll, a loot roll, and a placement choice. If another operation consumes a value before the loot roll, the loot routine receives a different part of the sequence.

No direct call from one feature to the other is necessary. Shared state is enough. An unused result still advances the generator if the draw already happened.

The on-kill reward routine illustrates this: it can draw even when the plant has no corresponding reward rate. An artifact branch can likewise draw values and then overwrite them. The number of draws consumed by a feature is part of its effect on later consumers.

A shuffle adds another variable. Permuting a list of $n$ entries makes $n-1$ distribution calls, and rejection sampling can consume more than one engine value per call. Changing the candidate count can therefore change how far the stream advances.

The shared game generator has 38 reseed call sites, and 30 pass time(NULL). This is a pattern across plants, loot, and artifacts. The clock has one-second resolution, and every reseed replaces the generator's accumulated state.³3Clock reseeding is shared across systemsThe branch census finds 38 calls to the shared seed wrapper, 30 passing time(NULL). Attributed clock callers include the weighted loot selector, its coin-spring sibling, the Gashapon spin, three instruction-equivalent Zomboss loot implementations, and two Gardener Grass paths. The loot selector reseeds immediately before its bounded draw; Gashapon reseeds at spin entry before its later selection logic. These sites all replace state at 0x104e3e580. A second reseed discards earlier intervening draws; only draws after that reseed can shift a subsequent consumer. The libc++ shuffle engine is a separate object.0x1034fca38 · 0x10198019c · 0x10197f7a4 · 0x101095248 · 0x1000e172c · 0x101eeab00 · 0x10229a9c8 · 0x100b21ec8 · 0x100b24b1c

All of these callers reset the same object. A loot routine can therefore change the random state later read by a plant or artifact, just as a plant can change the state later used by loot.

Two draws with the same bound, each taken immediately after a reseed to the same second, receive the same first result. Draws between those two resets are discarded by the second reset. A draw taken after the second reset but before the next consumer does change which output that consumer receives.

Gardener Grass adds an ordering mistake

Gardener Grass's fire routine reseeds before checking whether the plant is level 5. Below that level, it does not take the optional draw: it changes the shared state without using a value itself.⁴4A firing routine resets unrelated random stateGardener Grass’s fire routine calls time(NULL) and reseeds before checking whether the plant is level 5. The lower-level branch takes no draw after this reset. Its effect-object critical-hit path also reseeds immediately before a bounded-100 draw. Equal seeds yield the same first result; an intervening draw after a reseed can change which output a consumer receives.0x100b21e98 · 0x100b21ec8 · 0x100b24898 · 0x100b24b1c

Moving that level check earlier would remove this particular no-draw reset. It would not fix the wider coupling: the other clock-seeded callers would still replace the shared state. The general issue is a local feature resetting a process-wide dependency.

Gashapon's five-entry selector builds cumulative weights but draws again as it visits each entry. This is different from drawing once and finding the interval containing that value.

For cumulative totals $C_i$ and total weight $T$, its idealized selection shares are

With five equal weights, those shares are 20%, 32%, 28.8%, 15.36%, and 3.84%. The input weights are equal, but the output frequencies are not.⁵5Redrawing inside a cumulative scanThe five-entry selector constructs cumulative weights, but its loop back-edge returns to the random draw. Entry i receives product(1-Cj/T) for earlier entries times Ci/T. With five equal weights the idealized shares are 0.2, 0.32, 0.288, 0.1536, and 0.0384. This equal-weight example demonstrates the algorithm; it is not an attribution of the artifact’s unbound resource weight vectors. The inclusive float endpoint also permits a no-selection remainder.0x101095224 · 0x1010953e4 · 0x10109540c

This is a much larger distortion than the tiny modulo bias at ordinary small bounds. A cumulative-weight selector should draw once and find the interval containing that value; drawing again at each comparison changes the distribution.