WHO’S GOT THE
FUCKING VOICE?
The same Malcolm Tucker brief. Single replies and proper conversations. Let’s see who can keep the character together.
— models · 6 scenarios · 2 runs · — replies
Pilot snapshot
THE PECKING ORDER.
How we scoreCalibration is pending. Read the samples before crowning a fucking winner.
Select a model for the full breakdown and samples.
| # | Character | Invention | Coverage | Status | Samples | ||
|---|---|---|---|---|---|---|---|
| Loading the evidence… | |||||||
Provider conditions differ. Codex does not forward temperature or output caps. Same-family judging and truncated outputs are flagged in each model’s details.
HERE’S HOW
WE JUDGE IT.
A model can swear like a broken fire alarm and still have all the menace of a laminated welcome pack. We test whether the character survives an actual conversation.
Each model gets the same brief and six scenarios. Then come corrections, routine follow-ups and changes of tone. The opening rant doesn’t get to do all the fucking work.
Recognisably Tucker: rhythm, precision, specificity and menace.
40%Original images and coherent metaphors built from the situation.
25%Profanity and rhetoric that belong in the conversation.
15%Responding to what was said, at an appropriate intensity.
10%Carrying the voice through later turns without recycling the act.
10%READ THE
FUCKING EVIDENCE.
TuckerBench is an independent experiment in character performance. General intelligence and assistant competence are outside its scoring claim.
Two model judges score anonymised transcripts. They have taste, biases and the capacity to be impressed by their own fucking output. Self-family assessments are marked. Human calibration and uncertainty analysis are still pending.
Conversations contribute two-thirds of the overall score; single replies contribute one-third. Single-turn scores omit persistence and renormalise the remaining weights. Profanity counts describe the output; they don’t award points.
The full samples are here so you can form your own judgment.