# ADDENDUM to PREREG_rushmore_round2 — leg 10b re-run (29 September 2026) Leg 10b FAILED in run 1: worst |best lie − formula| = 5.10e-05 rad against 1e-6. That result stands in RUSHMORE2_CROWDS_run1_2026-09-29.json. **Diagnosis, a defect of mine.** A diagnostic on 4,000 fresh crowds put the worst gap at 8.1e-09, in a crowd where the liar's aim sits inside the reachable cone, so the true best distance is 0. There the distance is V-shaped in the report angle β. Scipy's bounded Brent stops at about 1.5e-8·|β| in β, and the distance error is that multiplied by the slope dψ/dβ, which grows without bound as the others' pull |R| falls toward 1. The criterion was set without asking the optimiser's precision, which is the same error I have named before. **The leg is unchanged:** same crowds (same seed, same order of draws), same formula, same criterion 1e-6 rad. Only the measurement changes. Where the crowd's direction ψ(β) − aim changes sign on the grid, the best distance is found by root-finding (brentq, xtol 1e-15), which gives the crossing to machine precision. Otherwise it is the smooth minimum at the tangent point, found by bounded Brent as before. The re-run also prints |R| of the worst crowd so the diagnosis can be checked.