While independently replicating each others results (just to be safe and prudent), a colleague and I inadvertently uncovered an issue whose import I had previously (but no more) underestimated--stats software updates.
In our specific instance, Stata's move from version 9.0 to 9.1 (I now use 9.2) implicated, among other commands, those germane to clustered standard errors. One result of Stata's update was ever-so-slightly different output (again, only relating to our clustered standard errors). Because we were working off the identical data and do files (but, alas, as we subsequently realized, slightly different Stata versions) any discrepancy alarmed us. An incidental hallway chat with Ted Eisenberg reminded me of Stata's version control command which allows users to run prior Stata versions. Once I did, exact replication (and much relief) emerged. (And for purposes of this discussion, let's set aside disputes about what "replication" might mean in a legal sense.) Because I frequently flip between stats packages (e.g., SPSS and Stata), I've always been mindful of subtle differences between stats software packages. This episode taught me that I need to be similarly sensitive to subtle differences across versions within a software package.
To the extent that this is an issue that will likely persist over time, perhaps we need to develop a norm promoting the disclosure of the software version used to generate results so that those following our footsteps can save some time.