← Maker leaderboard
n=18 · SMALL SAMPLEMaker Proof Card
MM-R1maker-arena-v1 · forward-markout quote-quality falsification · replay · config f997d5…6f33hash of the exact sealed maker-arena configuration that produced this result — identifies which configuration was scored, never a performance claim.
FALSIFICATION RESULTpairwise bootstrap test of the Δ between two agents; SEPARATED = whole 95% CI above zero, INCONCLUSIVE = CI spans zero, INVERTED = whole CI below zero (the candidate is reliably worse).the claim · lead with this
Separated
txline-fair-mm is less toxic than the naive-mm control
SEPARATED_QUOTE_QUALITY · whole 95% CI above zero → the difference is real
Δ DELTA+43.0 bps
95% CI[34, 52]
UNIVERSE
18fixtures scored
n=18 (Polymarket-resolved cp1), small sample
WINDOW MARKOUT AGGREGATENOT CLV
WINDOW MARKOUT1050 bps
ACTIONS308,826
maker markout vs future TxLINE FV; labeled aggregate, NOT a CLV rank axis
CAVEATS
Trades ≠ Fills
Observed trades are not the maker's own fills — no fill attribution at MM-R1.
NO EDGE CLAIM
No executable-edge / PnL / fill claim — real_executable_edge_bps is null by construction.
DIAGNOSTIC
Mean markout is geometry (≈ half-spread/ref), not quote quality — never the rank axis. Rank is toxicity loss, lower is better.
Trade-aware diagnostic
future · operator-gated (real on-chain trade capture not run)
Fill-assumption bracket
not present at MM-R1
Anchor / verify
not yet surfaced by the API