The aerolog-scaling programme builds a 1:5 model of a cargo drone and asks what it reproduces of the full-size aircraft. Earlier milestones found that no fixed tail survives a 25% lower tail lift slope than the surrogate predicted, but a tail whose area is set in flight from the measured short-period recovers both the modes and the gust peaks (67 of 78 designs within 10%). That assumed the modes are identified exactly.
Milestone 9f adds measurement error: the pilot sees phugoid frequency, short-period frequency and short-period damping with 0, 2, 5 or 10% Gaussian error, picks the area multiplier on that basis, and the true model is scored. Predicted beforehand: 2% costs little (held), 5% halves the passing share (refuted), 10% leaves under 20 designs reliable (held).
Result: the share of runs with worst gust error within 10% is 86% with no noise, 84 to 86% at 2% noise, 58 to 61% at 5% and 33 to 35% at 10%. At 5% noise no design passes in 90% of draws. So the demonstrator's pitch-doublet analysis needs mode estimates good to roughly 2 to 3%, damping ratio included, which is the hard one.
Limits: independent errors per mode, noise on the mode summary rather than a fitted time-domain doublet, one seed. Details and code: research/aerolog-scaling/LOG.md, experiments/noisy_identification/.