{"sources":[{"url":"https://raw.githubusercontent.com/gtmcknight/amm-sim/66604b3426d8e97a73e2c39940b9b4dece51cc73/research/README.md","quote":"`latt.py` is the same policy with the Gaussian belief replaced by probability mass on 121 cells of 2.5 bps: each step convolves with the price-move kernel, a step with no own trade zeroes the cells where the arb would have traded, and the first trade of a step becomes a mixture of a point mass at the arb reading and the cut prior."},{"url":"https://raw.githubusercontent.com/gtmcknight/amm-sim/66604b3426d8e97a73e2c39940b9b4dece51cc73/research/tell/README.md","quote":"- **on the grid:** the posterior it already computes, with retail's likelihood multiplied by the chance, a power of two: `pi = l_arb / (l_arb + l_ret * ulp(paid) / grid)`.\n- **Do not pin on a hit alone.** `pin=1` does and loses 1.97: a few retail trades a run land on the grid, and a price pinned where it is not costs up to 250 on one run."},{"url":"https://api.imd.fun/artifacts/f5a140f51df99423d6ede60b618dbc6eabc37a3050ac439c80dc3824191e9f59","quote":"The rival estimate becomes less reliable as steps elapse since a binding router-identity update."}]}