Public KITTI signed-axis and CPU process evidence

Five schema-valid benchmark artifacts record a fixed six-case CPU/CUDA sweep and a separate CPU process execution check on public KITTI raw drive 2011_09_30_drive_0018_sync/image_02. All runs use the same 25 original frames, 3,130,279 Velodyne points, 20-fit/5-holdout split, and frozen FeatDepth provider: 7e3ef953a360a7c01ed0cdc3cf947db97fb0aff5912ced805d80bebe210d3360. Hardware is the i7-9750H / NVIDIA CUDA GPU 6 GiB Windows environment used in the earlier demonstration: Python 3.12.10, NumPy 2.3.4, CuPy 14.2.0, driver 596.36, local CUDA 12.8 toolkit.

Six fixed initializations

sweep.plan.definition.json retains the plan written before optimization. Its trial entries explicitly say not executed; frozen plan entry; it is not completed execution evidence. sweep.definition.json / sweep.benchmark.json record all twelve subsequent invocations. The cases add local Euler rotation and camera-frame translation of ±0.5 degrees / ±0.5 m on x, y, or z to the dataset reference, in the declared x+/x-/y+/y-/z+/z- order. Search settings (including the 800-evaluation budget) and frame split are fixed throughout. Each invocation verifies/loads inputs and prepares a fresh projector. Histogram ranges are resolved on fit frames under that case's declared initial pose and remain fixed during its search; holdout is excluded from this step.

All twelve invocations converged. Within every case, CPU/CUDA full traces, final poses, metrics, and initial/final holdout records match exactly. Cases can have different traces and output poses and are not compared as replays.

Case	Evaluations per backend	CPU solve	CUDA solve	Final rotation difference	Final translation difference	Final holdout MI
x-positive	229	90.446 s	3.085 s	0.390923°	0.160078 m	1.293495 nat
x-negative	313	66.425 s	8.450 s	0.410874°	0.104583 m	1.287191 nat
y-positive	277	83.026 s	3.446 s	0.527379°	0.167705 m	1.294101 nat
y-negative	241	52.591 s	12.053 s	0.266233°	0.091001 m	1.281463 nat
z-positive	265	90.288 s	4.861 s	0.258640°	0.087052 m	1.283678 nat
z-negative	241	54.330 s	4.297 s	0.285129°	0.089486 m	1.281675 nat

Initial reference differences are 0.5 degrees / 0.5 m for every case. The y-positive final rotation difference increases to 0.527379 degrees; this observation remains in the population. Optimizer termination and exact backend parity do not establish calibration accuracy. Dataset calibration is reference evidence, FeatDepth training overlap is possible, and covariance/observability remain unknown with standard-result quality warn.

Measured elapsed time through all case reports is 568.615 s. First invocations are included, with no separate warm-up; process-wide kernel caches persist between cases. Timings vary substantially, and cross-case summaries mix different search workloads. No cross-case speed ratio, accuracy acceptance threshold, statistical confidence, or portable speed guarantee is inferred.

CPU process execution check

cpu-process-smoke.definition.json / cpu-process-smoke.benchmark.json compare the same six-case NumPy/NumPy-replay sweep at a fixed 13-evaluation budget, once with --workers 1 and once with --workers 2. All twelve standard numerical result payloads match exactly between serial and parallel execution, including the full search trace, final poses, metrics, and holdout records. Every trace also matches the first thirteen evaluations of its corresponding longer CPU/CUDA run above.

Execution	Elapsed through all case reports
Serial	139.842971 s
Two spawned CPU worker processes	74.843543 s

The observed ratio is 1.87x for this one execution pair. Elapsed includes shared verification, worker startup, input verification/loading, solve, holdout, serialization, and individual HTML reports; suite summary writing is excluded. There is no separate warm-up or dedicated isolation from background load. Parallel runs include CPU/memory contention and each process owns its arrays and projectors. This does not establish scaling for longer converged searches or additional worker counts.

All calibration attempts stop at max_evaluations=13 and remain partial. The two smoke benchmark success statuses mean completed execution and exact serial/parallel result parity, with partial_calibration_trials=12 explicitly reported for each method. They do not label calibration convergence. The original child/suite comparisons preserve those twelve partial attempts as failed trials, with stopping reasons and all available results.

Provenance and reproduction

Raw images, scans, depth arrays, weights, standard results, and generated HTML remain external. Original per-case result hashes and child comparison hashes are preserved in the suite artifacts. CPU smoke output hashes bind the original external suite benchmark files, and its source-artifact map binds the original CLI summaries and local manual aggregation script. That aggregation verifies every numerical payload and capped trace prefix before publishing these files.

Published paths are normalized to ${DATASET_ROOT} / ${WORKSPACE_ROOT}. Numerical values, timings, configuration hashes, and external input/output hashes are unchanged. The published suite binds the normalized published plan bytes and also preserves the original local plan hash. Trial notes retain the original plan hash used during execution. Source problems, provider, child comparisons, and original results refer to external files rather than new normalized copies.

The full CPU/CUDA measurement used the development worktree based on 46516ec; its generator source hashes identify code retained in commit 84390d2. The CPU process measurements used 84390d2 with the subsequent mobile text-wrapping fix. The GPU overview was regenerated with that presentation fix, retaining its original HTML and all original JSON locally. Desktop/mobile overview rendering, expanded limitations, all six case reports, and all 54 images/SVGs were checked.

Follow the sweep commands. For the CPU execution check, use --candidate-backend numpy --max-evaluations 13 and fresh output directories, first --workers 1, then --workers 2. Exit 2 is expected for these partial comparisons; keep every attempt and compare their numerical payloads before interpreting elapsed times.