What is not built

The original design specified considerably more than what shipped. Naming the difference is more useful than a feature list, and a reader deserves to know which parts of the original claim survived contact with data.

Measured and removed

The design called for a posterior probability that an order is informed, computed from six on-chain microstructure signals. Two of those — realised variance and order-flow imbalance — were built, tested, and wired into the swap path.

Calibration then measured their incremental value over simply reading the reference price. It came in at −0.008 and −0.002 across two independent windows: negative in ten of twelve horizon and threshold combinations. They were making the classifier no better while billing every swap for the gas.

They were removed from the swap path. The libraries remain in the tree, pure and tested, for a milestone that can show they earn their cost. Two earlier errors had hidden this result — a markout horizon that silently resolved to a wider window than intended, and a comparison of univariate rather than incremental discrimination.

Contracts that do not exist

There is no rebate ledger, no ERC-6909 accounting, no EIP-712 router attestation, no on-chain logistic classifier, and no growth-optimal fee curve lookup table. The design described all of them.

The growth-optimal curve in particular is blocked on three parameters — baseline volume, volume decay in fee, and arbitrage arrival intensity — that have not been estimated from data. Shipping a curve with invented parameters would be worse than shipping a mechanism whose single parameter is honest.

On the interface side: pools, positions, portfolio and activity surfaces are not built, and there is no light theme. Swap execution is real — the SwapCard drives the actual approve-then-swap flow against the live pool — but the Simulation surface’s demo controls (replay speed, a developer/calldata view, presentation mode) are not.