Keep/discard review of the 84 GPS-promising files

This local review addresses storage usefulness. It does not make an Earth-model claim or require unresolved files to satisfy a scientific orientation gate before they can be kept. Of the 84 files, 82 use IMU6, whose physical conversion needs separate investigation; two use IMU21 and pass the earlier physical checks as fine-alignment decoder diagnostics.

Final decisions

Keep 83 files; discard one redundant copy. Every kept file has at least 60 seconds of uninterrupted raw IMU measurements inside a GPS-selected stretch. Hash and framing checks passed for all 84. The kept files represent about 7.02 hours of GPS-selected geometry across potentially overlapping streams; this is not independently confirmed level-flight time.

Twenty-two kept files also have at least a minute compatible with the stricter fused-navigation motion limits when the alignment label is omitted. These are descriptive spans, without a new buffer around every internal bank boundary; they are not accepted scientific windows. The other 61 still contain complete measurements and useful trajectory context, so missing orientation/decoding knowledge is insufficient grounds to discard them. Keeping them does not establish that a usable level-flight interval can ultimately be recovered.

The remaining 83 comprise 81 IMU6 files and two IMU21 files. Their reported median heading uncertainties within the GPS stretches range from about 0.050 to 0.755 degrees. Those are internal fused estimates using an inferred older layout, not demonstrated accuracy. None gains scientific eligibility. Units/axes for IMU6 and independent orientation/correction behavior still need work before an Earth-shape test.

The review verifies each original's hash and packet checksums, checks for at least 60 seconds of uninterrupted raw IMU measurements inside GPS-selected stretches, and reports the navigation solution's roll, motion and internal uncertainty. Those navigation quantities are fused context, not independent orientation evidence. Unknown scales and fine-alignment flags do not by themselves justify throwing away useful measurements.

An initial IMU6 sample has much stronger gyro agreement after a 90-degree horizontal-axis remapping. This is diagnostic evidence that an assumed axis convention can cause failed comparisons. It is not a validated scale table or a completed decoder.

The catalog's 21 and 22 April 2010 LVIS510 files contain identical original bytes and both embed 21 April timestamps. The correctly dated entry was retained. The redundant owned compressed original was removed after verifying both hashes and writing durable prepared/removed disposition records. This reclaimed 6,984,885 bytes (about 6.66 MiB). Other useful or uncertain measurements stay intact. The corpus now holds 232 originals: 80 earlier candidates and 152 alignment-blocked files. There are 93 historical no-level rejections and one later duplicate removal; previous reports retain their original counts.

The audit uses Applanix's interface document for framing/status structure. Full-navigation status incorporates user accuracy requirements; it is not a universal accuracy threshold for this experiment. The manufacturer's operation guide also describes user-selected accuracy indicators. Neither status nor internal RMS estimates independently establish actual orientation accuracy.

The separate manifest, atomic reports and process lock live in data/ilvis0-retention-84-20261007/. No prior ledger or frozen study is rewritten. Scientific eligibility remains unchanged; no downloads, synthetic campaigns or Earth-model fits run. The worker has completed and released its lock. inventory.csv lists each filename, decision and reason; completion.json, summary.json.gz and cleanup.jsonl preserve the evidence. The manifest and source snapshots preserve the implementation. Seven focused retention tests passed, including protection against a changed canonical original and interrupted deletion.

Any future selection must honor the cleanup journal: the removed 22 April catalog entry maps to the preserved 21 April entry. Do not restart older completed workers that expect both copies.