Here's an interesting little article about Japanese phonology by Kawahara Shigeto arguing (eponymously) that "Lyman’s Law is active in loanwords and nonce words".

Lyman's law, as any fule kno (and as Kawahara xplain), is "a general phonotactic restriction in Japanese which prohibits two voiced obstruents within the same morpheme," famously observable via its blockage of rendaku in cases where the second element of a two-morpheme compound already contains a voiced obstruent. Or, as Lyman himself put it, "the second part of a compound word takes the nigori [Japanese for "voiced form"]; that is if beginning with ch, f, h, k, s, sh, or t, those consonants are changed into the corresponding sonant ones ... [but] the general rule does not apply ... when b, d, g, j, p, or z already occurs anywhere in the second part of the compound" (Lyman 1894).

There's more to rendaku than this, of course, and even Lyman observed as much, but the one-voiced-obstruent-per-morpheme insight is the interesting one. (Naturally, Motoori Norinaga and Kamo no Mabuchi reportedly had it first, although I wasn't able to find the exact terms they framed it in.)

Anyway, the question of whether Lyman's Law applies to modern loanwords is a complicated one, not least because said loanwords often already feature multiple voiced obstruents.However, Kawahara cites three pieces of evidence, from three separate papers, supporting the idea that Lyman's Law is still an "active" process at work in loan words and "nonce words" (pseudowords specifically):

1. Vance (1980), finding via wug test that Japanese speakers appear less likely to apply rendaku to compound nonce words when the second morpheme contains a voiced obstruent, i.e. the rendaku would violate Lyman's Law.
2. Tateishi (2003), observing that the English plural suffix -s appears in Japanese as /zu/ in most cases, but /su/ when attaching to words already containing a voiced obstruent (e.g. the baseball team names /faitaa.zu/ "Fighters" vs /taigaa.su/ "Tigers") (there are a lot of other factors at play here and Kawahara goes into them if you are interested)
3. Nishimura (2003), arguing that the general tendency to avoid repeated voiced obstruents is the reason why loanwords like /baddo/ "bad" can be devoiced to loanwords /batto/ while loanwords like /reddo/ "red" cannot be devoiced to /retto/

Kawahara then describes three experiments he performed to investigate further. The first two are updates on Vance's 1980 wug experiment. Kawahara more or less replicates Vance's results, except that (1) where Vance told his subjects that the words were pseudowords, Kawahara also tried telling his subjects that the words were Old Japanese, thus theoretically testing for and failing to find a difference in how subjects treated "Japanese vocabulary" and other words; (2) Kawahara's experiment did not reproduce Vance's finding that the proximity of an offending voiced obstruent to the compound word break can affect acceptability of rendaku.

In the third experiment, Kawahara investigated the /batto/ vs /reddo/ issue, for both real words and nonce words, and came up with surprisingly clean graphs showing that the "naturalness of devoicing" is, from highest to lowest:

1. Geminate consonant + Lyman's Law violation (e.g. /baddo/ "bad")
2. Geminate consonant (e.g. /reddo/ "red")
3. Non-geminate consonant + Lyman's Law violation (e.g. /bagu/ "bug")
4. Non-geminate consonant (e.g. /hagu/ "hug")

Interestingly, the spread of variation is wider in real words than in nonce words. In other words, it's more acceptable to devoice a real "Geminate consonant + Lyman's Law violation" word than a pseudo-word with the same structure, and less acceptable to devoice a real "Non-geminate consonant + no violation" word than a pseudo-word with that structure. Perhaps this represents the effects of a historical "strong Lyman's Law" preserved in the lexicon combined with an ongoing weakening of the law itself. That is, it might be that real loan words were affected more strongly by Lyman's Law in the past, so that /baggu/ → [bakku] is even now considered much more acceptable than /reddo/ → *[retto], but nonce words, having no history, are subject only to a weaker, present-day Lyman's Law, producing less polarized acceptability judgments.

References

These are the references from Kawahara's article for the other papers mentioned above, plus Kawahara's paper itself for good measure:

• Kawahara, Shigeto. 2012. Lyman's Law is active in loanwords and nonce words: Evidence from naturalness judgment studies. lingBuzz/001344
• Lyman, Benjamin S. 1894. Change from surd to sonant in Japanese compounds. Oriental Studies of the Oriental Club of Philadelphia.
• Nishimura, Kohei. 2003. Lyman’s Law in loanwords. MA thesis, Nagoya University.
• Tateishi, Koichi. 2003. Phonological patterns and lexical strata. In The proceedings of International Congress of Linguistics XVII (CD-ROM). Prague: Matfyz Press.
• Vance, Timothy J. 1980. The psychological status of a constraint on Japanese consonant alternation. Linguistics 18: 245–267.