Methodology

Blind matched-context battles, append-only votes, Bradley-Terry ratings.

Your question

PDFTASKBACKGROUNDOBJECTIVECONSTRAINTSASKTASK BUILT
A question becomes a structured task

Blind battle

SHARED TASKABA IS BETTERB IS BETTERTIEBOTH ARE BAD
Two anonymous ideas answer the same task in the same format

Vote, then the reveal

A IS BETTERB IS BETTERTIEBOTH ARE BADMODEL 1MODEL 2YOUR PICK
The vote locks, then the names appear

The thread continues

THIS ROUNDABFOLLOW-UP, NO VOTEVOTE, THEN FOLLOW-UPABCDSAME PAIRCONTINUESFRESH PAIRRESAMPLED
Follow-up keeps an unvoted pair; a vote brings a fresh pair
ABACBC
Three blind pairings rank all three ideas, one page at a time

Ranking

1ST2ND3RD
p(i beats j) = si / (si + sj) - rating = 1000 + 400 log10(s)

Every pairwise vote feeds the same Bradley-Terry fit with its recorded sampling propensity; nothing else ranks.

Why this design

Blind and matched-context

Model names are hidden until the vote is cast, and both ideas answer the same task in the same format, so preference is about the ideas.

Append-only votes

Every vote is stored in full and never rewritten; the leaderboard is a pure function of the vote log that anyone can recompute from an export.

Propensity-weighted ratings

The sampler steers votes toward under-battled pairs, so each vote records its draw probability and the Bradley-Terry fit reweights by its inverse.

Harnesses are disclosed

Every ranked row is a model plus its harness, and after you vote the revealed idea carries its full tool story: searches run, sources read, turns and cost.

Blind matched-context battles, append-only votes, Bradley-Terry ratings. The full protocol is in the methodology.