[contingencies]

It strikes me that the use of diverse systems to reinforce assumptions of trust within a given subsystem is an architectural paradigm not limited to compilers. The key problems are implementation feature-set or edge-case differences and overhead (real time and maintenance/up-front development). In fact, it would be ideal with multiple client versions/implementations on any service (particularly distributed or financial) and indeed I have done this in the past. Not sure if this paradigm has a name... anyone? I suppose you could just say consensus-based hedging.

[nickpsecurity]

It's in fact a long-established technique in high-assurance systems going back to I think aerospace or security-critical where the triple-modular redundancy trick was reapplied with separate teams building each one. Security through diversity also re-emerged relatively recently as a very active sub-field of INFOSEC/IT. If you're interested in that stuff, use these terms in various combination when you're Googling: "artificial diversity," "automated obfuscation," "moving target software security," "security diversity software." Also, including "pdf" helps given most good ones are papers. The word "survey" will occasionally land you on a pile of gold with references all in one spot. :)

Happy hunting!