Why the revised procedure failed its checks

An IMU (inertial measurement unit) measures turning and acceleration. To interpret its small level-flight signal, the analysis must know how its axes point relative to the aircraft. This review used the saved 72-case results and small, independently constructed motion examples. It ran no new flight simulations, nonlinear fits, bootstrap or calibration.

Two separate losses of information

The analysis first removes signals that allowed IMU errors or wind could imitate. It then keeps only sufficiently strong independent combinations of the remaining model signals. That second step uses a fixed numerical cutoff. A comparison can lose useful information at either step; changing the cutoff would change the method being validated.

Here, “enough” means that every required model comparison passes the same frozen retention threshold at every model anchor. It is not a calibrated decision or a guarantee of useful power. The revised dynamic-crab failures arise at the cutoff. Its explicit-wind failures already have insufficient information before that cutoff. Lowering the cutoff alone cannot solve both. No magnetic-watchdog segments were excluded in any of these 72 cases.

Turning the IMU while the aircraft turns

The revised schedule places an IMU turn at minute 20, exactly when the simulator begins its first aircraft turn. The aircraft-turn clock includes an initial five minutes and the time spent making each previous turn; it does not follow the nominal leg boundaries alone.

The software integrates the gyroscope during a deliberate IMU turn to recover its new facing direction. The gyroscope sees both the hand movement and aircraft motion. In an independent example, a known 180° hand turn overlapping aircraft yaw of 3° per second produces about 15° of mount-mapping error. No data-gap warning catches it because the measurements are complete. Another example with overlapping bank motion produces about 7.5° of mapping error.

Without overlap, the same independently constructed examples recover the mount change and forward direction accurately. The basic conversion between coordinate frames agrees with independent rotations to numerical precision. These checks isolate a timing hazard, rather than a general sign or axis error.

The revised campaign's estimated horizontal forward direction differs from the nominal mount direction by 4.82–5.21°, compared with under 0.2° for the original procedure. Its reported direction uncertainty is about 1.5°. This comparison is diagnostic: nominal mount geometry does not capture every pitch, bank or accumulated reconstruction error. Saved summaries omit the raw readings and mount matrices needed to reconstruct the historical error exactly. The overlap has not been shown to cause every campaign failure.

A concrete candidate for the next check

A timing-only revision moves the six IMU turns to minutes 5, 25, 40, 55, 70 and 80, while preserving the proposed 90-minute route. It meets the existing spacing rule and avoids all known aircraft turns, including the integration padding. It has not been optimized or tested through the complete recording and analysis process.

Small calculations using the aircraft-turn timeline and known IMU axes pass all model anchors under both wind descriptions with that schedule. The worst retained fraction is about 0.327, against a threshold of about 0.312: a narrow margin. Those calculations use approximate route geometry and coarse exclusions, so the earlier lesson still applies: a promising simplified calculation is insufficient to approve a flight protocol.

Before another flight campaign, define how overlapping turns will be detected and how uncertain orientation will cause abstention. A later, separately budgeted comparison should retain raw recordings, processed bins and mount matrices so the effect can be isolated directly. It should compare timings using paired inputs without relaxing the wind model or information cutoff. Fresh calibration remains premature until a usable protocol and its acceptance rules are fixed.

The reproducible audit tool is audit_processing.py. Its saved technical output belongs to this campaign's separate research record. Eight focused checks cover independent rotations, clean and overlapping maneuvers, schedule timing and the two stages of information loss. Historical campaign results remain unchanged.