A gearbox that shifted four times in half a second
A gearbox bug made lifting off the accelerator at high revs shift up through every remaining gear in under a second, because the upshift threshold is set from acceleration rather than a fixed rpm, and acceleration collapses the instant a driver eases off. The fix is a flat 500 millisecond cooldown between shifts.
Caught on a preview clip
Building the ten-second clip for the V10 preset means holding the car at a steady speed, under load, then lifting off to catch the way the engine falls back on itself for the recording. In one take, the throttle lifted a few hundred rpm short of the preset's 4,600rpm shift point, and the gearbox answered by shifting up, then up again, then twice more, landing in a gear with no business appearing at that road speed. The whole climb took less time than the waveform editor's shortest visible marker.
On a normal drive this is nearly impossible to catch. Four shifts inside half a second do not read as four events on the rpm needle; they read as a stutter, the kind of thing a driver blames on a GPS glitch rather than on gearbox logic doing exactly what it was told. It only became unmistakable on a clip recorded for playback, where the same half second sits on a timeline and can be scrubbed back and forth until the four steps are obvious.
Reproducing it on demand turned out to be simple: hold the speed a little under a shift point at full load, then lift off cleanly, and watch the gear counter rather than listen for it. It happened on every engine tried, every time, with the same shape: one legitimate shift immediately followed by others that had no reason to exist.
Why the threshold falls with the driver's foot
The upshift point a driver meets while driving moves with load. It starts at 62 percent of redline as a cruising baseline, then climbs as inferred acceleration increases, gaining 1,500rpm for every extra metre per second squared, up to the wide-open-throttle case the car reaches at 2.5 metres per second squared. Ease off, and the same formula runs backward: the threshold falls exactly as fast as the acceleration estimate does, because there is no pedal telling the box what the driver meant, only the acceleration worked out from GPS speed a moment before.
That is stable while a driver holds steady pressure, because acceleration barely changes from one moment to the next. It stops being stable at the exact moment recorded on the clip: near redline, the driver lifts off, acceleration swings from its full-load value towards zero inside a single update, and the threshold falls with it, arriving below the engine's current rpm before the engine has had time to react. The box reads that as licence to upshift, immediately, and if the rpm left over after the ratio change is still above the newly lowered threshold, it reads the same licence again.
Smoothing already sits underneath all of this: speed is filtered with a 130 millisecond time constant, and the rpm needle with 95, both there to stop noisy GPS-derived numbers looking noisy. Neither one is built to catch a real, fast change in driver behaviour, and a lift-off is exactly that: a sudden change that the smoothing correctly lets through, straight into a threshold formula with no floor on how quickly it can fall.
The 500 millisecond cooldown
The fix does not try to make the acceleration estimate smoother, or teach the threshold formula to tell a lift-off from a stumble. It adds a floor underneath the whole thing: once a shift happens, the gearbox will not consider another one for 500 milliseconds, whatever the threshold and the current rpm say to each other in the meantime. Four shifts in half a second becomes, at most, one shift every half second.
This is a cruder fix than solving the underlying problem, and it does not pretend otherwise. The gearbox already treats acceleration and braking differently for downshifts, dropping back at 1,500rpm while cruising and holding on until 2,750rpm under braking, and each of those thresholds could in principle be caught out by the same kind of rapid load swing. A single cooldown, applied after every shift in either direction, covers all of it without a special case for each.
It is also the more honest option given what the drivetrain has to work with. There is no pedal signal in the car; the whole model of load is a number derived from how a smoothed GPS speed happens to be changing at that instant, and that number will occasionally be wrong for a fraction of a second. A cooldown accepts the error and limits its damage, rather than pretending the acceleration estimate can be trusted down to the millisecond.
Where it is easiest to hear
The bug was loudest, in an odd way, on machines that already make a sound on lift-off. The turbo hot hatch preset fires an overrun pop when the driver comes off the throttle, at a pop probability of 70. Before the fix, one lift-off near a shift point could trigger that pop once per phantom shift, stacking three or four pops into the time it takes to say the word overrun. After the fix there is one lift-off and one pop.
Ratio count decided how far a cascade could run. Contrast the eight close ratios on the tuned V8 against the four long ratios on the muscle V8, two presets already compared for what their firing order does to the growl: here, the number of gears on the table decided how many rungs a falling threshold could reach in half a second.
The clip that caught the bug in the first place now does what it was meant to. Rpm climbs under load, the driver lifts, and the engine falls back through one gear, cleanly, rather than through every gear left in the box.
The cascade only ever lasted half a second, needing a specific combination of load, revs and a foot coming off the pedal at the wrong instant, but it is the kind of fault a bench test would never surface and a single recording session found on the first attempt. Building a gearbox from a GPS trace and no pedal signal means the load number is always an estimate dressed up as a measurement, and the honest response to a bad estimate is a hard limit on how much damage it can do in a row, rather than a cleverer model that promises never to be wrong.