From Gareth, on the first draft:
I'd suggest dropping the 'mini' caveat - reads a little defensive and people know
mini models won't perform as well.
The page currently marks codex-gpt-5.4-mini with a dagger and explains in the
legend that it is a deliberately weaker model, run to find the floor rather than to
represent typical usage.
I argued for that marker and now think Gareth is right, for a reason that has
strengthened since: the model's name already says "mini". The note explains something
the label states, and explaining it reads as pre-emptive defence.
There is also a newer argument against it. Loop 1 showed that failures in this suite
are frequently variance rather than capability (#14). A note that attributes a weak
model's failures to its weakness is asserting a cause we cannot currently support. It
is not just defensive, it may be wrong.
Proposal: drop the dagger and the legend sentence. Keep the model identified by
its full name, which is self-describing. Revisit only if a reader actually
misreads it.
Done when the marker and its legend note are gone from the page.
From Gareth, on the first draft:
The page currently marks
codex-gpt-5.4-miniwith a dagger and explains in thelegend that it is a deliberately weaker model, run to find the floor rather than to
represent typical usage.
I argued for that marker and now think Gareth is right, for a reason that has
strengthened since: the model's name already says "mini". The note explains something
the label states, and explaining it reads as pre-emptive defence.
There is also a newer argument against it. Loop 1 showed that failures in this suite
are frequently variance rather than capability (#14). A note that attributes a weak
model's failures to its weakness is asserting a cause we cannot currently support. It
is not just defensive, it may be wrong.
Proposal: drop the dagger and the legend sentence. Keep the model identified by
its full name, which is self-describing. Revisit only if a reader actually
misreads it.
Done when the marker and its legend note are gone from the page.