I hadn't thought of coupling, the distributed nature, so you have a point. However we can think of splitting up the system into pieces, and try to localize the update. For example, for a worldwide distributed system keep everything else the same but replace this one mainframe with modern hardware and code so that the rest of the system doesn't get effected. And so one.

As for the couple million, that's for me :) But I guess if you factor in cost of supporting hardware/software/infrastructure, it's really past 5 million (maybe 10 million), just as an estimate.

To elaborate on my 'unconventional way', I'm talking about using all of:

- complete view of the source code,

- runtime behavior on the screen (even recorded with cameras, or even better, display debug hooks)

- runtime storage logging,

- runtime network logging,

- modern inference methods beyond parsing (semantic inference, logic programming, maybe even some statistical analysis).

- But most importantly, automation-first approach (no, we're not doing things by hand. 'The Machine TM' will do the work for us). E.g., I won't be reading the COBOL code line by line. The code, plus the runtime behavior, would be the input to the inference engine, the output would be 'explanations, documentation, etc'.

- and so on

It would be a very challenging project, hence the 5-10 year timeframe. But I'm not convinced it's intractable. And I'm not convinced the cost is in the billions.