Sparse screw trajectory evidence¶
complete-evidence.tar.gz contains 876 indexed files in 4,319,551 bytes.
It retains both complete 401-module source trees, frozen measurement tools,
result and record schemas, all twelve complete input/result/fit/measurement
records and logs, the derived public geometry fixture, and packaging source.
The original recording is not included.
inventory.json binds the archive, index and original summary.json bytes.
The maintained baseline is 5a80e7470487ceab148f7d49b560f48c1158665e and
the measured candidate is 38b98d03be01919abb5d7a12cab5c939001eef95.
The median of three paired complete-API ratios is 2.366735x for synthetic LiDAR–IMU joint recovery and 3.142851x for the public GLIM geometry fit. Fresh-worker paired ratios are 2.203778x and 2.878685x, respectively. Every input, fitted trajectory/calibration/covariance and full saved result agrees exactly within each pair, excluding provenance only.
These are spent-data performance comparisons. The geometry case assigns
timestamps and fixes capture-local plane associations; it is not a physical
bag trajectory or an end-to-end LiDAR–IMU calibration. Joint seed 20261011
fails the same holdout gyro policy in both variants, and seed 20261010
retains its existing Windows max_iterations status. No quality threshold
or outcome is changed to obtain a performance pass.
Replay the bounded, digest-indexed archive with the current CPU-only auditor:
python tools/audit_screw_residual_archive.py docs/assets/screw_residual_evaluation/complete-evidence.tar.gz
It validates record/artifact schemas, source and tool bytes, imported-module bindings, all observations and logs, complete pair equality and the recomputed medians/gates. It does not import Calibrex, execute archived code, or require the original bag or a GPU. Temporary extraction stays below the selected audit directory and is removed after replay.