A periodical from The Ronin Advisory Group
GunnerBench
A scoreboard for model behavior under controlled variation. Each issue is a multi-page investigation that does the thing a leaderboard cannot: it shows the work — the actual prompts, and the models' actual answers.
Issues
The unit of analysis is the diff between cells, not any single score — change-detection, not absolute measurement. Same tasks, same rubric, judged the same way; what we publish is the difference.
Everyone Passed. Now What?
In about two years, expert legal reasoning went from a frontier capability to a commodity. Four models now write a perfect 20/20 bar-exam memorandum — and the price of that identical grade spans roughly fifty-four times, from eighteen cents to nearly ten dollars. When capability stops being the variable, price and trust become the whole question. We prove it by showing the work.
MikeOSS versus the unmodified prompt
A performance ranking of the MikeOSS tool against the same models prompted directly, on identical tasks.
The security of MikeOSS
A look at the security posture of MikeOSS and its current variants.
What this is
GunnerBench is published as findings, not released as a product. The reader we have in mind is the legal professional — sophisticated, skeptical, and answerable to someone who is counting on the decision. A leaderboard hands that reader a number. A periodical hands them the memorandum the number was built from, so they can read the lawyering themselves and decide what to trust.
Each issue follows the same discipline: a battery of tasks held constant across every model, a single judge, channel-corrected costs, and the artifacts published in full. The apparatus is in service of the knowledge, and the knowledge is what we put on the page.
GunnerBench is built and operated by The Ronin Advisory Group LLC.
The first dispatch is out. The next question goes here.
© 2026 The Ronin Advisory Group LLC