I did the following in order to investigate the bit depth of various record modes and the SPDIF digital ins/outs of the Roland VS880ex and VS1880 recorders.

First I made an observation of the mixer/rec bus interface which shows some nonlinear behavior. Then I used this nonlinear (yet predictable) behavior to perform a couple experiments. 

Here they are:


OBSERVATION:
When recording a signal, or more precisely, when routing a signal to the record bus, the record section will not allow attenuation of the signal if it is already so small as to be near the LSB (least significant bit) limit. In other words, once a signal's dynamic range has been reduced to the LSB, it cannot be reduced any further regardless of fader reduction or attenuation when being recorded. It would be expected that 6 or 12dB attenuation would produce silence, but it does not. A very small signal will always be recorded and remain audible (very distorted). Only when the fader is at zero, will the signal disappear. A signal already existing near the LSB limit will remain mostly unchanged (only disappearing if the fader is at zero). 
However, during playback, when the signal is on the 24bit track mixer, it can be attenuated below the LSB of 16 while monitoring (of course, with 24 bit being the absolute limit in the mixer). 

When a signal is routed to the record bus (i.e. routed to a track), it is altered to fit the record mode. This is true even when the signal is not being recorded. Just being routed to the record bus alters it. The record section truncates (no truncation in MTP) and data-compresses (no compression in MAS or CDR) the signal based on record mode. Coming out of the record section, it is re-expanded to 24 bits uncompressed on the mixer.



EXPERIMENT 1 : to determine the smallest signal recordable in MT1 mode.
I recorded a clean 1KHz sinewave onto a track in the VS880ex at 0dB FS using MAS mode. I then bounced the track down to another track at a fader level of 1 and ATT set to -12dB. The signal then resembled a square wave. I  bounced this small signal several times with the same settings(fader at 1 and ATT set to -12dB). After the first bounce, the signal never got smaller or silent. It remained unchanged. This was also the case even when actual recording wasn't taking place. Just being routed to the record bus had the same effect. It is a relatively safe assumption that this small signal is being formed by the 1 or 2 Least Significant Bits (LSB) of 16, since nothing lower can be recorded in MAS mode. 
This experiment was repeated in MT1 and was also repeated on a VS1880. Results were the same in all cases. Since MAS is 16 bits, it can be safe to say MT1 is also equivalent to 16 bits of dynamic range. 
Repeating this experiment in MTP mode showed MTP was able to record a signal much smaller than that of MT1 or MAS.

EXPERIMENT 2 : to determine if the digital outs and ins are greater than 16 bits. 
First I recorded a signal on the VS880ex in MAS mode at the minimum level that could be recorded (LSB of 16). I then attenuated it gradually from fader level 100 to 0. I could audibly hear the change by turning the headphone level to maximum. I could also watch the master output on an oscilloscope to visually compare levels. It decreased smoothly. The mixer clearly can reduce the signal well below the minimum level that can be recorded in MAS (which is 16bit) as long as the signal is not routed to the record bus. To find out if this change would be transmitted digitally via SPDIF, I connected the digital out of the VS880ex to the digital in of the VS1880. I could hear (and see) the same results on both machines indicating that the digital in/out handles more than 16 bits. 


Next, for verification, I tried to record this fade from the 880ex to the 1880 using MAS, MT1, and MTP modes. MAS and MT1 behaved identically. The LSB of 16 signal coming into the digital input on the 1880 was recorded with input mixer settings nulled. As I faded the signal on the VS880ex, the 1880 (in MAS and MT1) recorded the fade as a jump from a steady small signal to zero signal (when the 880ex fader hit zero level). In MTP mode, the fade was recorded as a smooth transition from a small signal to zero signal. 

I also noted that the mixer could not produce a signal so small that MTP could not record it, or digital could not transmit it. This would imply that the dynamic range of the mixer matches that of the digital out/in and MTP rec mode. 


CONCLUSION:
Neither experiment can quantify MT1 as 16 bits, but shows MT1 to be of the same dynamic range as MAS. If MAS is 16 bits, it would be safe to say MT1 is also 16bits as a measure of dynamic range.

Neither experiment can quantify MTP or the digital out to be 24 bits, but did show them to be greater than 16bits and of equal dynamic range to that of the mixer. If Roland's claim of 24 bit internal processing means the mixer is 24 bits, then it would be safe to say the digital in/out and MTP are 24 bits.


Roland VS recorder record modes:
MAS -  16 bit uncompressed 
CDR -  16 bit uncompressed (non-separable pairs)
MTP -  24 bit compressed (RDAC)
VSR -  24 bit compressed (RDAC) presumably since it is equivalent to MTP
MT1 -  16 bit compressed (RDAC)
MT2 -  16 bit more compression (RDAC)
LIV1 -  16 bit heavy compression (RDAC)
LIV2 -  16 bit very heavy compression (RDAC)
Other than CDR and MAS, the bit numbers represent the equivalent dynamic range RDAC is capable of, and not actual stored bits. 


Converters:
VS880vx -  18 bit
VS880ex, VS1680 -  20 bit
VS890, VS1880 -  24 bit


Digital input/output:
VS880vx - ?
VS880ex - Always 24 bit SPDIF
VS1880 - Always 24 bit SPDIF 


Mixer section:
VS880vx - ?
VS880ex - Always 24 bit processing
VS1880 - Always 24 bit processing