Complete prepared-input GNSS–LiDAR evaluation reuse

This archive retains the selected six fresh-process runs, prepared problem, generator, original public OXTS packets/timestamps and download provenance, complete core source capsules/manifests, measurement code and schemas, official results, full residual/optimizer histories and validation receipts. index.json binds every publication byte other than itself.

All 2,762 OXTS poses of KITTI 2011_09_30_drive_0018_sync are used. The antenna track and 29 odometry windows are generated with the declared transform, lever arm, clock and seeded noise. There is no raw LiDAR registration or independent accuracy evidence. This is a prepared-input performance/parity control; its unchanged policy verdict is inconclusive for the lever arm.

The final whole-process median is 13.871298 → 4.896686 seconds (2.832793x). The complete prepared evaluation phase is 11.028193 → 2.085993 seconds (5.286783x). All eight default jackknife fits and four held-out controls remain. See the benchmark for timing boundaries, all pairs, work counters and limitations.

history.json.gz retains every requested residual's fit, window IDs, pair and epoch counts, exact parameters and full residual digest. It also retains every least-squares solution, active mask, termination fields and full residual/ Jacobian digests. The audit replays both bounded caches and checks actual interpolation, every queried epoch and SVD work. Real full and abbreviated Git identities and imported source bytes are independently verified before comparing all six complete results/histories.

Raw input/source hashes and archived numerical payloads are exact. Only offline regeneration from OXTS allows platform math roundoff: rtol=0, 2e-8 m for translation and 1e-13 for rotation-matrix elements. Times and IDs stay exact.

The qa directory includes the two failed endpoint fixture attempts and their repairs. They concerned single-axis angle array shape and the artificial fixture's GNSS gap bound; the final regression and complete reuse suite pass. No measured core or protocol was repaired after the selected six-run cohort.

KITTI attribution: Andreas Geiger, Philip Lenz, Christoph Stiller and Raquel Urtasun. Original data and generated derivatives retain CC BY-NC-SA 3.0: dataset and license. Calibrex source snapshots and measurement code retain Apache-2.0.

Audit offline from the repository root:

python -m tools.benchmark_gnss_lever_arm_reuse \
  --audit-only docs/assets/gnss_lever_arm_evaluation_reuse