OK, thanks to ikegami's help, I've fixed the benchmarking code. Here are the results:

# The suffix +/- correspond to "matching" and "non-matching" input,
# respectively.

# input length: 7
          Rate ?<!-  ?!+ ?<!+  ?!-
?<!-  583158/s   -- -38% -54% -55%
?!+   945231/s  62%   -- -25% -27%
?<!+ 1262619/s 117%  34%   --  -2%
?!-  1286220/s 121%  36%   2%   --
# input length: 10
         Rate ?<!-  ?!+ ?<!+  ?!-
?<!- 220554/s   -- -49% -63% -64%
?!+  430349/s  95%   -- -28% -31%
?<!+ 601510/s 173%  40%   --  -3%
?!-  619376/s 181%  44%   3%   --
# input length: 10000
        Rate  ?<!-   ?!+   ?!-  ?<!+
?<!-   342/s    --  -91%  -99%  -99%
?!+   3864/s 1029%    --  -86%  -86%
?!-  26795/s 7730%  593%    --   -6%
?<!+ 28417/s 8204%  635%    6%    --
[download]

use strict;
use warnings;

use Benchmark 'cmpthese';

my $sz      = ( shift || 10 ) - 4;
our $match   = 'N' . 'x' x $sz . '001';
our $nomatch = 'N' . 'x' x $sz . '000';

print '# input length: ', length( $match ), $/;

cmpthese( -1,
          {
           '?!+'  => '$::match   =~ /^N(?!.*00$).*$/',
           '?!-'  => '$::nomatch =~ /^N(?!.*00$).*$/',
           '?<!+' => '$::match   =~ /^N.*(?<!00$)$/',
           '?<!-' => '$::nomatch =~ /^N.*(?<!00$)$/',
           }
        );
__END__
[download]