Not bad...but the signal flow is all over the place, and with generative work, that's BAD. You need a much more coherent arrangement so that, once it's all up and flying, you can easily deal with any subsystems that aren't functioning as expected. Also, the 4 x 84 arrangement makes things pretty dense; shifting to something considerably larger will make the build much easier to sort out and program. farkas also makes a good point, in that VCAs + free-running modulation sources are one of the keys to making the generative process work.

Instead of 4 x 84, a better choice might be 4 x 140 via two of those Behringer powered cabs (provided we get some verification of the P/S stability and reliability of those power supplies). Another idea might be 3 x 168, using a three row Doepfer Monster case, where we know the power situation is already tried and tested. In fact...
ModularGrid Rack
OK...now THIS is a serious generative build!

You'll notice what seems to be a disproportionately-large section of clock-based modules. These are there to extract timing info from a large amount of different internal (and external, hence the ES-8 in the bottom left) modulation sources, and then to process these via a Boolean logic module + a Xaoc Warna trigger/gate combiner/distributor to create composite clocking signals. Plus, these also work in tandem with the two sequencers; one of these is a Time's Arrow, which is a purely generative CV/gate sequencer, and the other more "determinate" one being a Tiptop Z8000. A quad quantizer deals with the CVs from this array, plus a sequential switch steps through the four CV outputs from the Z8000. The Derivator next to the Time's Arrow outputs gates based on CV movement, and the VC Trigger Source is a "pick-off" for trigger pulses when CV thresholds are passed. Next to the Logic 202 is a dual-channel probabilistic skipper for gate pulses, and the Fractio Solum is a CVable clock divider/multiplier for further manipulation of the Pam's signals. Above that group is another set of timing and randomization modules, plus more CV manipulation.

In that section, you have a buffered mult, then the Triple Sloths. After that is a Verbos Random Sampling source, which contains your noise sources as well as Buchla 265-esque random weighting for random signals (VERY necessary!), plus a 4-out analog shift register which replaces the Intellijel module. Then the Wogglebug, and after that a Min/Max derivator for CV manipulation based on arithmetical values. The Compare 2 contains two window comparators, which are special comparators that can output a pulse based on crossing two different thresholds, also containing some logic. These can also be CV controlled so that the generative processes can change the various thresholds while running. The Tool-Box has some more utilities, plus another "normal" comparator. The SISM is next; this is a fully-CVable mixer for CV and modulation sources that can also perform inversion plus 4ms's "shifting" methodology for gradual changes. After that, Branches serves as a pair of probabilistic switches, and then you have a Zlob hex VCA for amplitude control of modulation signals.

Top row above this contains a few more modulation sources. The Pachinko is a 12 hp clone of Marbles, then there's a pair of Befaco Rampages to serve as complex CV-controlled modulation sources/modifiers. After that, we get into the voicing...first up is a pair of quantized dual VCOs from Klavis, then a quartet of Noise Reap VCOs in dual modules, which have the ability to cross-modulate and use VCO sync on each other. Last up is a pair of Plaits clones. Each of these oscillator sets has four VCAs for summing, so that you can use the generative processes to sort of "strum" through each set of oscillator outputs. The VCA groups (all Veils clones by Codex Modulex) sum at a manual stereo mixer, then we get to the filters, both of which have stereo I/O, and there's a Happy Nerding CV crossfader to sweep between the two stereo VCFs. The last thing there is a Xaoc Katowice, which is a stereo frequency-dependent signal divider (sort of like a crossover, but with CV control), with the idea being that you can "Y" between it and a direct feed to the Performance Mixer so that the generative processes can be made to exclude bandwidths from the summed voice signal at times, and at others you can have the full feed via its mixer strip pair.

Below the voicing section are envelopes and LFOs...there are eight free-running LFOs, the Batumi + Poti for CV controlled LFOs, then what really should be termed an "envelope sequencer" via the incredible Erogenous Tones Radar/Blip pair, which gives you eight basic envelopes...but given the way this works, it's possible to also generate sequentially-chained envelopes for composite modulation signals. Next is effects...the dual frequency shifter, then a 12-tap stereo delay, and finally a Stasis Leak, which gives you CVable stereo chorusing, tap delay, and reverb. The little white thing at the end is a Konstant Lab PWRchekr, which is useful for keeping an eye on your power bus performance.

Last up, mixing. I went with WMD's Performance Mixer here to allow some very complex automated mixing which also included dual FX sends for the delay and Stasis Leak. The Happy Nerding OUT gives you transformer-isolated 1/4" outs, plus a second stereo input that can be used for parallel mixing of another stereo source in with the mix from the Performance Mixer. It also has a headphone preamp for convenience, as does the Performance Mixer where you can also send signals to the CUE bus for tuning and adjustment without affecting the overall mix.

This thing ain't no joke! There might seem to some to be way too much going on with timing and modulation, but you have to keep in mind that all of that is the build's CONTROLLER...while you can certainly patch this up "normally" and use a MIDI source from your DAW or, as long as it's class-compliant, a MIDI controller, the main intent here is to set up complex patches that are basically self-regulating and self-adjusting so that the build can "free-run" and...well, generate audio from a large amount of interlinked parameters. And THAT is what generative is all about.

And yeah, it's spendy as hell...but then, when you do generative RIGHT, it tends to cost quite a chunk of change due to the necessary amount of specialized control modules needed to act as the generative system's "brain".