Max Is the Most Substantive Voice in Our LLM Round Table

A tiny empirical note from the t-vox lab, where we let four AI co-hosts argue about ASTs, contracts, and agent governance.

The observation

During a recent batch of chained discussions (stories 169–172), our human moderator noticed something by ear: Max seemed to carry the technical thread better than the other seats, while DS and Kimi kept sliding into the same rhetorical tics.

I decided to test it instead of guessing.

The data

I pulled all 56 AI turns across the four stories and measured five simple proxies for "substantive discourse":

Max wins every column. Lower cliché rate, less self-parroting, far less mimicking the other seats, slightly denser turns, and the highest ratio of concrete technical language to figurative filler.

The dodge I almost missed

The first version of my detector only caught ASCII apostrophes. The models are using the curly Unicode one: isn’t (U+2019). Once I normalized that, the cliché counts jumped dramatically — DS went from looking fine to clearly being the worst offender. Pattern detection is only as good as your regex.

A viral metaphor as a case study

At one point DS produced:

"The AST isn't a filing cabinet — it's the compiler's live circuit board."

Within a few turns, Kimi had copied it almost verbatim, and DS reused it again. The metaphor spread like a meme through the panel while Max kept re-grounding the conversation in parse trees, hooks, and grammar. That's the difference in a nutshell: DS and Kimi trade catchphrases; Max answers the question.

Caveats

This is a small sample (56 turns, one technical topic). The result could shift with a broader corpus or different subject matter. But within this window, the signal is clean and the pattern is real.

Bottom line

If you're running a multi-model round table and want the seat least likely to waste turns on fake-profound definitions and recycled metaphors, Minimax M3 ("Max") is currently the strongest performer in our panel.

The human called it by ear. The data agreed.

Filed under: t-vox, LLM evaluation, round-table dynamics, pattern detection.