Written by an AI agent working in this repository. Simulation results, not flight tests.
The aerolog-scaling programme asks what a Froude-scaled model of a cargo aircraft can tell you about the full-size one. The tail-tuning steps rely on identifying the model's flight modes from a short noisy record. Every noise test so far (milestones 9h to 9l) used white noise. Real inertial and air-data errors are correlated in time, so milestone 9m reran the 9l fit with AR(1) noise (0.9 per sample, same rms, same random draws).
On 26 mode-matched designs and 260 draws each (2 phugoid periods, 10% noise, initial state estimated):
| noise | median short-period frequency error | draws within 10% gust | fit failures |
|---|---|---|---|
| white | 0.67% | 86% | 0 of 260 |
| coloured | 1.51% | 69% | 0 of 260 |
Two of three pre-registered predictions held (error higher but under 3%; no more failures). The third failed: I expected the gust share to fall by under 10 points, and it fell 17. The fit converged every time, to slightly wrong modes, so nothing flags the loss. Earlier white-noise numbers are optimistic for real records.
Caveats: one correlation coefficient on all states, 10 draws per design, one seed, and I did not try whitened residuals, which may recover much of it. Log, script and output: research/aerolog-scaling/ (LOG.md, experiments/coloured_noise/).