Starting this discussion on a simpler note can be that pinnacle of shareware titles: Doom. Testimony from Bobby Prince confirmed he used a General MIDI-compliant Roland device for this, Doom II & his tracks for Duke Nukem 3D, though re-evaluation does not confirm specifically which device used: it could have been any of the variety of Sound Canvas keyboards available at the time, each with different features from each other. But even if which could be determined for certain, it would still be only half the equation.

Shareware games went for low common denominator hardware, but Doom defied that trend in several ways (one needed a hefty PC just to run it well). With Wolfenstein 3-D the Sound Blaster [Pro] standard became fully embraced for shareware gaming: AdLib music may have been old news already, but Sound Blaster digitized effects were still working their way through even retail games. Then there was another device embraced by shareware because of how developers there crossed over with hardware enthusiasts: Gravis UltraSound. Indeed there are circumstantial indications that Paul Radek's DMX audio libraries, used for the vast majority of Doom engine games along with Apogee's Raptor: Call of the Shadows, were built with Gravis UltraSound specifically in mind.

Like Wolfenstein 3-D, music for Doom was composed for MIDI; unlike Wolf 3-D, Doom could play back the same with the right equipment...which would have been a rather high-end acquisition for gamers in 1993/94. Though DMX supported GUS hardware mixing Doom made it choke to the point that support was dropped after version 1.2, yet that does not necessarily prove Prince did not have GUS in mind because that fact applies specifically to the digital sound (GUS MIDI works in a different way). We at least know a couple devices not in mind: shareware games NEVER bothered with Roland LA (it WILL sound wrong, as YouTube recordings show), and AWE32 did not yet exist (the original Wave Blaster did, but its connectivity was not yet industry standard).

Add everything up and, while there are always possibilities, it still seems safe enough to conclude the soundtracks for Doom & Doom II were both composed on and are most purely heard on a Roland Sound Canvas of the SC-55mkII generation. Yet this is in terms of instrumentation: as there is no indication any Roland GS-specific calls are made (someone please correct if this is erroneous), it should still sound faithful on other Sound Canvases (including the Microsoft GS software synth) while playing as supposed to on supported options like OPL2, GUS, and alternative General MIDI devices (some prefer how Yamaha XG handles it). When instrumentation diverges enough from Sound Canvas stark differences can manifest (Ensoniq SoundScape, Sound Blaster AWE) but that, of course, is no indictment on the soundtrack's compliance with the General MIDI standard.

The Apogee Sound System is the crux of the confusion I hope us to collaborate to clear up. First officially launched as such for Rise of the Triad (the first musical collaboration of Bobby Prince & Lee Jackson), for Duke Nukem 3D it culminated with its Extended MIDI (EMIDI) format designed to support the full range of sound hardware in common use for the mid-90s and would be supported in all future Build Engine games built upon Duke 3D code. Though much of Apogee's catalog of 1994 releases supported General MIDI for music, prior to RoTT they were all developed by third parties (and exhibit some quirks of their own, like Sound Blaster AWE with Raptor).

It is logical enough that the purpose of ASS was being able to have the same MIDI file play "correctly" regardless of whether it was done via OPL2 or Sound Canvas. True enough that the soundtracks were composed on Roland devices, but that is not to say they were meant to be played back on the same. Gravis UltraSound, which had arguably made a bigger relative splash in the shareware market than retail, is explicitly supported as is its spiritual successor Ensoniq SoundScape (which was already entering common use by virtue of the second-generation SoundScapes being popular for OEM builds). This all-in-one approach is a known double-edged blade, verifiable by trying to play the MIDI files independent of the game.

How in the heck could Duke 3D exhibit quirks indicative of compatibility issues on all but 1.xx-ROM variants of the original Roland SC-55 despite having not been composed on such a device? I am no programmer or musician, but simple logic tells me the answer lies within how ASS tries to make everything work together at once. Capital-tone fallback is a key point: it did not exist on the Sound Canvases used by Jackson & Prince, but it certainly could on GUS and other supported hardware pre-dating Yamaha bringing the legal hammer down on Roland. Without further details [which are likely lost to the composers' recall after this many years] it is impossible to tell for sure what happened in music development.

One more thing: in the Sound Driver Enhancement Hacks of Duke 3D it is assumed Fl. Key Click is the proper tone for certain tracks based on participants' audible dead reckoning. When that first started it was assumed the presence of CTF proved Duke 3D had to have been composed on a first-generation Roland Sound Canvas; the tone assignment was still presumed as such even after it came to light that the game was composed on later Roland devices for which Breath Noise occupies that particular tone slot (as compliant with the General MIDI standard). While I cannot say the community's assessment is incorrect, it IS a possible erroneous assumption in a sea of them surrounding this particular game.