Radar clock-profile reference indexing comparison

All six fresh-process runs are retained in their original alternating order. The whole prepared-kinematics calibration workflow changes from a median 90.695003 to 11.618504 s (7.806083x). The solver API changes from 88.714816 to 6.589154 s (13.463765x). The host was unreserved; variation is material and every timing is reported. No warmup, slow run or failure is omitted.

The input is all 2,762 OXTS samples from public KITTI raw drive 2011_09_30_drive_0018_sync. The 2,758 Radar velocities are generated execution controls from those kinematics, with a declared lever arm, rotation, clock offset and deterministic noise. They are not independently measured Radar. This comparison establishes performance and parity, not Radar calibration accuracy. The archive contains no locked accuracy-cohort evaluation.

Timing includes process startup, imports, JSON parsing, object construction, all 201 offset profiles, least squares, observability, unchanged holdout and eight signed probes, full trace capture, serialization and saving. Source freezing, input generation, downloading and external auditing are untimed. Both variants use the same dependencies, one BLAS thread and no measurement budget. Every profile candidate and complete result match exactly across all six runs. Only the separate provenance wrapper has a variant-specific commit.

protocol.json pins the source commits, capsules/manifests, control and actual launcher/helper/schema bytes. Each worker records its actual imported source hashes and normal Git identity. The only changed runtime module is calibrex/solvers/radar_spatiotemporal_lever_arm_solver.py. The source capsules omit scratch Git directories. input-download.json binds the original ZIP members with SHA-256 and CRC-32; input-fetcher.py.txt is its exact generator. control.json binds the files read to generate the control and records the field convention and generation code identity. index.json binds every other published byte, including empty stderr files and regression receipts.

The source data and derived numerical control are attributed to Andreas Geiger, Philip Lenz, Christoph Stiller and Raquel Urtasun, KITTI, and are provided under CC BY-NC-SA 3.0. The official KITTI raw page and original archive URL are retained in the control/download provenance. Source-code capsules retain the project's Apache-2.0 license.

Audit without a GPU, raw dataset download or calibration run:

python -m tools.benchmark_radar_reference_indexing --audit-only docs/assets/radar_reference_indexing

See the benchmark report for all timings, scope, regression coverage and remeasurement commands.