A same-machine replay comparison catches the four desync causes one computer can reproduce

Start with iteration order. It sets a constraint on how the second pass runs. Rust randomizes hash seeds per process, so the second run walks the same entities in a different order, and in an order-sensitive physics step the two runs diverge ^[9]. Two runs inside one process share a seed, so a HashMap walk can agree by luck; the replay has to be a separate process for that class to surface at all ^[19]. Tickwise ships iteration order as one of four chaos modes for this reason ^[9].

The post calls time leaking the case that hides well ^[10]. A subsystem reads the wall clock or a frame delta instead of the fixed tick, and in the post's example the recorded frame took 16.2 milliseconds against the replay's 15.9, which was enough for the simulation to take a different branch ^[10]. On a fast machine the two numbers look close enough to be invisible ^[10].

Two of the four classes are about state nobody cleared. A scratch buffer left dirty between ticks, or a lazily built cache that exists on the second pass and not the first, starts the replay from a slightly different world than the recording ^[11]. A static counter or a shared random generator leaves the same kind of fingerprint across sessions ^[12].

Take the published verdict as a claim about the tool's own plumbing. The --chaos flag injects a known class of non-determinism at a tick you choose, and it exists so the tool can prove it works ^[14]. Injecting stale-value at 4021 produces a compare verdict of first divergence at tick 4021, caught by the light hash, confirmed by the full hash at tick 4200, last agreement at tick 4020 ^[15]. Confirmation trails detection by 179 ticks ^[16]. There are two hash tiers in that verdict, a cheap one per tick and an expensive one at intervals ^[21], so for the same shape to repeat on your code the cheap hash has to cover whatever field went stale.

The same check can happen during the replay instead of after it. Replayer::open takes a ReplayConfig, and with verify_hashes: true every tick of the replay is checked against the hash stored in the file ^[18].

The argument rests on the cost of the test. The post asserts that a large share of desync bugs were never about the second machine and that nobody had checked the single-machine case ^[2], and it does not offer a measurement ^[20]. It does give a procedure you can run alone, offline, in a few seconds, before any player is involved ^[8]. Lockstep netcode depends on the multi-machine version of the same promise ^[3]. The author wrote that the self-check is "the first thing I would run on any simulation that claims to be deterministic" ^[5].