Hey.

Loading the live ledger…

Each row swaps only the direction. Universe, daily budget and the σ^1.5 × confidence weights are identical, so the column is a clean read on direction alone. bandit* is the Thompson bandit rebuilt around a Normal posterior over the spread rather than a Beta over a win rate: same arms, but it learns the expected payoff instead of which side wins more often. That one change moves it from −$2,519 to −$204.

The voter set was chosen on redundancy, not on this column. Over 91 days v3's daily standard deviation is $243, so a 91-day total carries a standard deviation of $2,317, and every drop-or-add variant tried spanned about $1,500, which is 0.65σ. What separates the priors is how much they agree: bandit* matches mean14 93% of the time, mean10 92% and mean7 89%, because a recency-weighted posterior mean is a trailing mean. Voting four near-duplicates is one voter with extra confidence. So mean10 was dropped as a 92% duplicate, and persistence added as the only prior that genuinely disagrees with the rest, at 56–66% agreement, despite being weak alone.

Every settled day since the live deployment began. Daily P&L on the left, running total on the right.

The vote itself is untouched. What changes is what the voters are trying to be right about. v2's three voters all predict a sign and are graded on how often they are right, and on this record the most accurate rule is also the one that loses the most money, because the book is paid in dollars, not in hit rate. v3 keeps the LLM and the bandit but has each statistical voter estimate the expected spread instead, and lets the LLM's confidence set position weight, which is the one thing it measurably does well. bandit* is the same two arms rebuilt around a Normal posterior over the spread instead of a Beta over a win rate.

Every margin in the paper, ours included, is measured against a level that a handful of days determines. v3 states this in the abstract rather than leaving it implicit.

The forward premium inverts, volatility weighting makes the loss larger rather than smaller, and both statistical corroborators turn negative. Only DPDS, which sizes its own capital, survives.