Method

Nothing is hardcoded but factory addresses. For a pair the router asks Uniswap V2, SushiSwap and BaseSwap for their pair, Aerodrome for both its stable and volatile pool, and lists the fee tiers of each concentrated-liquidity deployment — Uniswap V3 and PancakeSwap V3 — then repeats all of that through WETH and USDC as intermediates. A two-hop route is a candidate on the same footing as a direct one.

The candidate set is deliberately wide and pruned by price rather than guesswork: every candidate is quoted once at full size, and only the best six get the full ladder. Laddering all of them would be roughly a hundred and eighty contract calls.

Constant-product venues are priced off-chain from reserves, with the fee numerator that fork actually charges — BaseSwap takes 25bp where Uniswap V2 takes 30. Once reserves are known the whole ladder is arithmetic, multi-hop included, since a two-hop route is the same function applied twice. All arithmetic is bigint.

Aerodrome and the V3 deployments are quoted on-chain. A Solidly stable curve and a concentrated-liquidity tick walk can be reimplemented off-chain, and a reimplementation that drifts by one tick is worse than none.

Each venue is quoted at a geometric ladder of sizes, producing an output curverather than a number. Because a pool’s output is concave in size, handing each successive slice to whichever venue offers the best marginal rate converges on the optimum — the water-filling argument.

Interpolation between rungs is piecewise-linear, which on a concave function underestimates. That is the safe direction: the solver will never believe a venue is deeper than it is.

Splitting is then charged for what it costs. A split that wins 3bp on a trade whose extra hop costs 6bp of gas is a loss, and the recommendation needs a full basis point of daylight before it changes.

Comparing a split to a single route needs both sides in one unit, and the extra cost is in ETH while the benefit is in the token being bought. Rather than a price feed, the router converts gas through the same pools it already quoted. The extra-hop cost is 70,000 gas, measured on a mainnet fork as the marginal cost of a second swap.

Prediction quotes a set of trades at a pinned block, forks that exact block, executes them against the deployed routers, and compares. Tolerance is 1bp; measured drift is zero.

Unit tests cover the arithmetic at edges the chain rarely visits — empty pools, one-wei trades, collapsed ladders, the interpolation lower bound the splitter depends on.

Address and token checks assert every hardcoded address still has bytecode and every token’s on-chain symbol and decimals match the table. A right address with wrong decimals misprices by a factor of a thousand, silently.

Gas profile measures the constants the router makes decisions with, so a guessed number cannot quietly bias every routing choice.