Research

Fill drift is the live buy price against the paper mark that triggered the entry. It answers the one question a paper test cannot: do real fills land where the simulation assumed? Everything else in the forward test rests on this being small.

Net P&L is after gas (priority fees on both legs); venue fees are already inside the quoted amounts, so counting them again would double-charge. Gas % is the real drag on a small stake — at $5 a ticket it is the difference between an edge and a rounding error. Streak is the run still in progress: a long losing run is the signal to stand down and re-measure, not to size up.

The same columns as above, one row per UTC day, last 7. Streaks are measured within each day, so a run that spans midnight appears in both — days are a reporting convenience, not a natural boundary for a strategy.

Only the timed rule cohort tests the paper strategy — it is the same sell-at-60-minutes rule the +6.6% pumpswap edge was measured on. Manual closes measure the operator taking an intra-hour peak, which is a different and equally legitimate question. Mixed together they would answer neither, so they are counted apart.

Combinations the books stopped betting but the loop keeps grading. A policy exclusion can only be revisited if the excluded thing is still measured — otherwise "we don't trade that" quietly becomes folklore instead of a claim with an expiry date.

Every number here survives four honesty rules learned the hard way: sellable depth is what a real sell quote returns (never market cap, never reported "liquidity"); peaks only count with real depth behind them; young mark-to-market positions are labelled, not celebrated; and any result carried by a single token is called an anecdote. Three separate "edges" in this project turned out to be value nobody would actually pay — including one venue where buying worked and selling was impossible.