Maestro vs Phantom

Random sample (1,000 a day per Solana row, 6,000 per EVM row) of the swaps each terminal routed, every attempt counted for the fail rate. Reference = the pool's state before the swap (its reserves, or the previous trade within 60 s); swaps without one keep their split, no loss figure. Published from 50 priced swaps pooled over the product's chains at 25 or more (a single-chain product from 25), ranked from 100; the median's 95 % interval is in the JSON. Compare medians, not single swaps.

Same sample, window and reference as bench 268; read that page for how a swap is priced. Here each swap's loss has the terminal's fee subtracted before the median is taken. Subtracting the two published medians gives a different number wherever an app's fee varies across its own swaps, which is why this is measured, not derived. A swap reads negative when a fee was rebated or the reference drifted the user's way; those stay in. Compare medians with their intervals, never single swaps.