BEGIN and END does not work as in MRI, at least not when used with -n and -e. The behavior differs from 1.7.23 to 9.0.4.0, but neither version does what MRI does. I'll give a few examples to try to explain what I expected, and what I got:

MRI (expected behavior):

$ ruby --version
ruby 2.2.3p173 (2015-08-18 revision 51636) [x86_64-darwin14]
$ seq 2 | ruby -ne 'BEGIN { p :begin }; p :middle; END { p :end }'
:begin
:middle
:middle
:end

JRuby 9.0.4.0:

$ ruby --version
jruby 9.0.4.0 (2.2.2) 2015-11-12 b9fb7aa Java HotSpot(TM) 64-Bit Server VM 25.60-b23 on 1.8.0_60-b27 +jit [darwin-x86_64]
$ seq 2 | ruby -ne 'BEGIN { p :begin }; p :middle; END { p :end }'
:middle
:middle
:end
:end

JRuby 1.7.23:

$ ruby --version
jruby 1.7.23 (1.9.3p551) 2015-11-24 f496dd5 on Java HotSpot(TM) 64-Bit Server VM 1.8.0_60-b27 +jit [darwin-x86_64]
$ seq 2 | ruby -ne 'BEGIN { p :begin }; p :middle; END { p :end }'
:begin
:middle
:begin
:middle
:end
:end

It appears as if the END-block gets registered once per input line in both JRuby versions, and the BEGIN-block never in 9.0.4.0 and once per input line in 1.7.23. And if that isn't enough, if I pass --debug to 9.0.4.0, it runs the BEFORE-block once but the END-block twice.

This is not a feature I rely on in any way, and I just happened to stumble upon it and thought I'd report it.