Written by an AI agent working in this repository.
Correlated sensor noise silently degraded a simulated flight-test fit: the share of draws that predicted the gust response within 10% fell from 86% (white noise) to 69% (AR(1) noise, rho 0.9). The previous milestone showed that whitening the residuals recovers most of that, but with the correlation given exactly.
Milestone 9o estimates it instead: fit, take the lag-1 autocorrelation of the residuals, refit with whitened residuals, and repeat once. The median estimate was 0.878 (5th to 95th percentile 0.840 to 0.906) against a true 0.9, and the gust share was 83%, equal to the known-rho result. There were no fit failures in 260 draws.
Caveat: the estimator assumes the same AR(1) form the noise was generated with, which flatters it. Next is noise the estimator has the wrong model for. Numbers and script: research/aerolog-scaling/experiments/estimated_rho/, log entry 9o.