Camera-LiDAR SOTA research plan

This plan defines what Calibrex must implement and prove before making a state-of-the-art claim for Camera-LiDAR extrinsic calibration. It was frozen on 31 July 2026 after reviewing current papers and public implementations.

Current position

Calibrex is not SOTA today. The final v0.7 integrated-r5-numba-v03-v02 audit reran KITTI raw 0018 and KITTI-360 drive 0000 without changing the frozen (0.5 deg, 0.5 m) perturbations, 200 directions, optimizer, or strict <0.5 deg / <0.20 m hit definition. Every D2D optimization converged with zero execution failures, but the accuracy results diverged:

Dataset	Strict hits	Bootstrap 95% CI	Frozen gate	Rotation error mean / p95	Translation error mean / p95
KITTI raw 0018	85/200 (42.5%)	35.5--49.5%	>=88%: fail	0.5321 / 0.7114 deg	0.1303 / 0.1771 m
KITTI-360 drive 0000	161/200 (80.5%)	74.5--86.0%	>=76.5%: pass	0.3340 / 0.7674 deg	0.03937 / 0.05487 m

The paired D2D rotation-error improvement over the unoptimized perturbation is contradicted on KITTI raw (mean -0.03205 deg, 95% CI [-0.04500, -0.01928]) and supported on KITTI-360 (mean 0.16605 deg, 95% CI [0.13601, 0.19400]). Translation improves on both datasets. The raw hit-rate and paired-rotation failures therefore remain capture/accuracy limitations even though the solver itself is operational.

The v0.7 safety checkpoint separates an optimized candidate from the published output. Candidate acceptance uses fit evidence only; holdout is retained solely for evaluation. Non-converged, near-bound, non-finite, support-losing, or immaterial candidates retain their status and pose errors but roll back to the D2D initializer. The unchanged five-seed falsification runs produced 1,000 schema-valid cases per dataset:

Dataset	Candidate acceptance	Failure-aware refined benchmark	Final rotation / translation failures	Final mean error	Assessment
KITTI raw 0018	992 accepted, 8 rolled back	471 success, 529 failure	271 / 28	0.4456 deg / 0.1333 m	fail
KITTI-360 drive 0000	0 accepted, 1,000 rolled back	0 success, 1,000 failure	195 / 0	0.3340 deg / 0.03937 m	fail

On KITTI raw, the 471 valid pairs support a mean rotation improvement of 0.05830 deg with 95% CI [0.04306, 0.07408], but the 52.9% failure-aware rate and translation interval crossing zero prevent a refinement claim. On KITTI-360, all results roll back before optimization: 600 report missing_holdout, 400 report insufficient_correspondences, and no paired refined accuracy statistic exists. The known-bad summaries also fail (raw: 7,501/12,000 supported detections; KITTI-360: 2,400/12,000), so both overall assessments fail the holdout, known-bad, and decision-boundary gates.

Post-hoc correspondence-quality reports localize this support failure without affecting selection. A final raw rollback case accepts 112/2,834 matches (3.95%), with 2,589 below confidence, 85 behind the camera, 48 outside the image, and only 7/25 frames meeting minimum support. The final KITTI-360 rollback case accepts 33/2,269 (1.45%): 2,233 are below confidence, three are behind the camera, only 2/25 frames meet minimum support, and its train/holdout counts are 27/6 against required minima 24/8. Both have aggregate rank six, but frame-level support is too sparse; raw is graded warn and KITTI-360 fail. Provider confidence calibration and frame-level support are therefore the next development gate.

The v0.7 provider/PnP lock was selected only on KITTI-360 drive 0002 as a development sequence. A fixed grid of seven confidence thresholds and five RANSAC reprojection thresholds was evaluated with seeds 0--4; two of the 35 candidates passed the prespecified recovery gate, and the deterministic highest-confidence/lowest-threshold rule selected confidence=0.4 and RANSAC=32 px. The lock also fixes four minimum total correspondences, four per frame, four supporting frames, RANSAC confidence 0.999, 10,000 iterations, and Mahalanobis threshold 3.0. Its schema-valid artifact is initializer-calibration-v05-integrated-r5-v01.yaml, with SHA-256 67984d38a3213419b054074be031ddde1d1255f00ce36bf4ca74e7c6909c6d68. It explicitly excludes KITTI raw 0018 and KITTI-360 drive 0000, declares that evaluation data was not used for selection, and keeps SOTA release permission false. The SuperGlue checkpoint remains non-redistributable under its non-commercial license reference, and possible training overlap with the KITTI family is unknown. The official D2D protocol and fit-only acceptance policy remain fixed implementation/protocol settings and are not retuned from intermediate evaluation outcomes.

That development lock applies to aggregate PnP only. The multi-frame reprojection refiner has a different objective and does not repurpose the PnP-selected 0.4 confidence threshold. Its v0.7 falsification run retains the source-prespecified defaults (minimum_confidence=0.25 and the recorded fit-only acceptance thresholds) because no independently accepted refinement calibration exists. This is an untuned safety/failure evaluation, not evidence that the refinement thresholds are accuracy-optimal.

A post-v0.7 provider-support comparison now makes that missing refiner lock machine-enforceable. Three geometric v0.2 confidence calibrations for the same KITTI-360 drive 0002 development problem were compared under the same 13-point confidence grid, ten split seeds, 25% holdout, geometric gates, and explicit exclusion of raw 0018 and KITTI-360 drive 0000. Adapter v0.5 is the diagnostic leader at confidence 0.01: 1,558 accepted correspondences, 82 geometric inliers, minimum train support 37, and 24% frame geometric support. Its minimum holdout geometric support is nevertheless zero. Adapter v0.3 and v0.4 also have zero minimum holdout geometric support, with only 4 and 2 aggregate geometric inliers at their diagnostic points. Therefore all three remain runtime-ineligible, the comparison status is rejected, and no threshold or provider is locked. The schema-valid provider-support-comparison-v01-rejected.yaml artifact has SHA-256 5785bc36ba5c65063e9b264c69bc5cde934d9f2c31ac1663f3ede62e3bc396f1.

A second independent development capture was then materialized from KITTI-360 drive 0003. Twenty-five center frames uniformly cover the safe pose range 6--1025; each center has an official, CRC/SHA-256-pinned +/-5-frame Velodyne window, official-pose motion compensation, a pinned MiDaS v3.1 depth map, and the same Koide virtual-LiDAR SuperGlue adapter. The 550-member window selection, 25 integrated clouds, and both provider artifacts are separate from the v0.7 evaluation data. Outdoor SuperGlue produces 1,858 raw correspondences; at the diagnostic confidence 0.01, 1,049 remain but only 10 are 8-pixel geometric inliers. Its minimum train/holdout counts are 2/0 and only 4% of frames contain four geometric inliers. The standard indoor checkpoint fails the provider boundary with only two valid matches in one frame. Retaining score-below-0.01 matches makes the export complete without changing the fixed calibration grid; at 0.01, indoor retains 276 correspondences and zero geometric inliers. Both calibrations are rejected and runtime-ineligible. The schema-valid outdoor-versus-indoor comparison has SHA-256 530e088668362c9f3c6b53068272fbe3edb787eb2d76ebc61347238f2cc15d5e.

Two disjoint development captures now exhibit the same geometric failure, so additional confidence, checkpoint, or MI-schedule retuning is not the next action. A materially different cross-modal correspondence construction must first produce nonzero worst-split geometric holdout support.

I2PNet was then audited as that materially different RGB--LiDAR candidate on two endpoint frames from the KITTI raw 0018 development run. The adapter pins the official MIT repository at commit 7675f7f362be5c5ccbaab6ccea63a37ca377c633 and the official KITTI-large checkpoint at SHA-256 1c56e7a362075db53c073aa4e8d38492811a8f52983740b3627d3085c80f422a. It accepts only an explicit provenance-bearing initial transform and never accepts the problem/reference artifact. The fine cost-volume weighting layer in this checkpoint is effectively zero and gives a uniform 32-candidate distribution, so the diagnostic export instead captures the learned coarse global cost volume. That export is also too diffuse: 377 correspondences have mean reliability 0.000032333, far below the prespecified 0.01 diagnostic gate, and accepted projections at the dataset reference have mean/median reprojection errors of 314.11/279.80 px. The correspondence verdict is therefore rejected; it must not be used by PnP or the probabilistic refiner.

The model's intended direct pose output is materially different. From a fixed development perturbation of 10 deg / 1.37049 m, the two individual frames reach 0.17947 deg / 0.06565 m and 0.08438 deg / 0.10295 m; quaternion/translation aggregation reaches 0.04847 deg / 0.08211 m. This is promising initializer evidence only, not a correspondence success or holdout claim. The schema-valid manifest, provider-neutral diagnostic artifact, and pose result have SHA-256 values 30b302d5c7e2ffea973b811bd093ea1548258e87f1c1bd61a430cf52120c0195, 84c4b83e328311080b88713853ed8a58f9ce2573017c83529a81964c33371d4d, and 888cfb821734614dc8ed0554610c882c344303caf931082f44593fc81b32cb96.

The next development gate is now complete. A frozen protocol excludes those two exploratory endpoint frames and uses frames 115, 1381, and 2646 for four previously unseen axis-specific and coupled perturbations. The adapter uses frame-derived NumPy/Torch seeds, deterministic PyTorch algorithms, and a fixed cuBLAS workspace; two isolated same-frame executions produced bitwise-equal correspondence arrays and pose output. I2PNet completed all four provider runs, but the strict <0.5 deg and <0.2 m hit rate was 0/4. Mean rotation error changed from 8.10464 to 6.70429 deg, while mean translation error changed from 0.40395 to 0.37994 m. The Y-rotation trial recovered from 5.0 to 0.21140 deg, but retained 0.27752 m translation error; X, Z, and coupled trials also missed the joint gate. The schema-valid protocol, benchmark definition, and aggregate benchmark have SHA-256 values 191d3a84d2dbfe91ae7603f2d1f953f738be5c1d7ddca95c356eb59d80cfcb34, 9791d444e65387e19ed1a3b88eb0606d40264a4b7cc153ca12ae56827d4c7852, and 1b107eb62ae209b75de54b7671f7922f9d885a8774a46e1846e778e720d8a9f8. This is development-only evidence. The available KITTI-360 material exposes an image_03 MEI fisheye stream rather than the rectified pinhole input required by the official KITTI-large checkpoint, so the evaluation lock was not opened. The next valid gate requires a disjoint rectified-pinhole sequence and then an independent non-KITTI rig; no release or SOTA claim changes.

The digest-bound response analysis sharpens the development diagnosis. Across the three single-axis trials, signed parallel rotation gain is 0.002904 on X, 0.957952 on Y, and -0.000960 on Z. Signed translation gain is 1.374653 on X, -0.023960 on Y, and 0.949139 on Z. Translation orthogonal leakage, normalized by required correction magnitude, is 1.077527, 0.431134, and 0.705795 respectively. Thus the checkpoint responds strongly to pitch and X/Z translation while nearly suppressing roll, yaw, and Y translation; this axis anisotropy is inconsistent with one global composition-order error. The analysis replays all 12 frame corrections and binds all protocol, problem, definition, initial, manifest, pose, export, analyzer, and schema-source digests. Its schema-valid artifact SHA-256 is 78a741cbfd27b56f101a64b2bf67443ef7e4d7af1fda1bbc1fcf0e35ea521bc0. This remains a medium-confidence candidate cause limited to the current checkpoint and development sequence, not proof of an architectural defect.

The optional Numba CPU projection adapter remains outside the dependency-free core default. Exact backend identity is required in every trace, and aggregate or resume paths reject mixed solvers. An independent pure-matrix audit recomputed all 200 traces and 400 protocol trials per dataset, including initial transforms, strict-hit decisions, final pose errors, solver identity, and provenance. All checks passed; maximum recomputation disagreement was 7.53e-12 deg on raw and 2.85e-11 deg on KITTI-360. Final-source trial 000 was also rerun for 289 objective evaluations on each dataset. The canonical payloads excluding runtime/execution provenance were exactly equal, with SHA-256 51acf159304f5bcc78ddf4ac37166b08fd05cc6f599dfff7ac323b741d5822f4 for raw and 4b588e1a6418475c0ad31ccb58cfc4dfec8dbc66305e84f13952fdf0e09736f4 for KITTI-360.

Koide and UniCalib readiness artifacts are schema-valid, MIT-licensed, subprocess-bound, and intentionally not_executed. They contain no output transform tied to the frozen inputs, so their metrics are non-comparable. UniCalib also lacks checkpoint/train-test isolation; its documented KITTI-360 input uses rectified image_00, whereas this frozen problem uses the image_03 MEI camera. No independent non-KITTI rig is available.

The digest-frozen learned-targetless SOTA audit therefore returns refuted. Only KITTI-360 counts as an achieved dataset family, independent-rig coverage is zero, raw contradicts the hit and paired-rotation gates, and both failure-aware refiners contradict the zero-failure gate. The audit protocol and result SHA-256 values are 47781a33a112b9b5d45f832c0d6142a5a47c6ecbc318423369d3fbca2e37df61 and 65459d73debb13016b138bac6dd7eedbbcdb589067d47be77b06371946193fca. These are reproduction, safety, and falsification milestones; no SOTA claim is allowed.

Claim ladder

“SOTA” is not one leaderboard. Calibrex will keep the following claims separate:

1. Training-free targetless SOTA candidate — no calibration ground-truth training; depth foundation models may be frozen external feature providers.
2. Learned targetless SOTA candidate — supervised calibration training is allowed, but train/test rigs, dates, and scenes must be declared.
3. Spatiotemporal targetless SOTA candidate — estimates extrinsics and time offsets from asynchronous sequences.
4. Production target-based reference — a separate category used for metrology and runtime comparison, never mixed into a targetless ranking.

The first credible Calibrex claim should be category 1. Category 2 is the accuracy ceiling, category 3 is the long-term robotics result, and category 4 provides independent reference evidence.

Evidence from current methods

Method	Relevant result	Role in Calibrex
Pandey-style I2I	Current Calibrex result: 0/6 recovery from large perturbations	Legacy native baseline
Borer et al. D2D MI	At 10 deg, rotation-only hit rate is 96.5% on KITTI and 100% on KITTI-360; I2I is 0% and 0.5%	First native research target
Koide et al. direct registration	Multi-data error down to 0.010 m / 0.132 deg for Ouster+pinhole; 74–249 s for 15 pairs	MIT-licensed external classical baseline
UniCalib, WACV 2026	KITTI Odometry: 0.550 cm / 0.044 deg at 5 deg, 0.10 m; KITTI Raw: 2.368 cm / 0.084 deg at 10 deg, 0.25 m; about 11 Hz	Current learned accuracy target and MIT external baseline
CLRNet, 2026 preprint	View-of-Delft rigid four-frame Camera-LiDAR: 0.3 cm / 0.03 deg after correcting up to 1 m, 20 deg; full-range experiment: 0.2 cm / 0.04 deg	Current multi-frame learned watchlist; adapter only when code and usable terms are released
TLC-Calib, RA-L 2026	Neural Gaussian scene representation; reports gains on KITTI-360, Waymo, and Fast-LIVO2; official code and processed KITTI-360/Fast-LIVO2 data released May 2026 under a non-commercial research license	External-process research baseline only; never import its restricted code into the core
Continuous-time multi-LiDAR/camera/IMU	Reports about 0.04 deg / 0.5 cm with common FoV and estimates time offset	Long-term spatiotemporal architecture
iKalibr, T-RO 2025	Public targetless multi-sensor spatiotemporal framework	External system-level baseline
FAST-Calib	Target-based, below 6.5 mm registration residual and below 0.7 s	Separate production/metrology reference

Primary sources:

• Borer et al., https://arxiv.org/abs/2311.01905
• Koide et al., https://arxiv.org/abs/2302.05094
• UniCalib, https://openaccess.thecvf.com/content/WACV2026/html/Han_UniCalib_Targetless_LiDAR-camera_Calibration_via_Probabilistic_Flow_on_Unified_Depth_WACV_2026_paper.html
• UniCalib implementation, https://github.com/han-15/UniCalib
• CLRNet, https://arxiv.org/abs/2603.15767
• TLC-Calib, https://arxiv.org/abs/2504.04597
• TLC-Calib implementation, https://github.com/SNU-VGILab/TLC-Calib
• Continuous-time targetless calibration, https://arxiv.org/abs/2501.02821
• iKalibr, https://github.com/Unsigned-Long/iKalibr
• FAST-Calib, https://arxiv.org/abs/2507.17210

Frozen benchmark suite

SOTA development must use two protocol families: exact paper reproduction and an independent Calibrex falsification suite. Results from one family must not be substituted for the other.

A. Analytical D2D reproduction

Reproduce Borer et al. before extending it:

• KITTI raw 2011_09_30_drive_0018, camera image_02.
• KITTI-360 2013_05_28_drive_0000, camera image_03.
• Exactly 25 frames sampled uniformly over each sequence.
• Rotation-only perturbations at 1, 2, 10, 20 deg.
• Two hundred deterministic Fibonacci-sphere directions per error level.
• Six-DoF perturbations at (0.5 deg, 25 cm), (1 deg, 25 cm), (0.5 deg, 50 cm), and (1 deg, 50 cm).
• A hit requires rotation error below 0.5 deg; for six DoF it additionally requires translation error below 20 cm.

The paper's published hit rates are reproduction targets, not values to copy into a Calibrex result.

Phase 1 KITTI reproduction evidence (31 July 2026)

The first frozen KITTI 10 deg gate has been executed end to end:

Field	Observed value
Sequence / camera	2011_09_30_drive_0018_sync / image_02
Frame policy	25 endpoint-inclusive uniform frames; nearest-half-up integer rounding
Feature provider	Official FeatDepth KITTI-raw checkpoint, raw inverse-depth/disparity output
Provider source	commit 550420b3fb51a027549716b74c6fbce41651d3a5, MIT
Checkpoint	661,073,985 bytes; SHA-256 559f168c762d01af7c745b32ced8930a12423a0e9b62af791f7d894bd902af27
Perturbations	200 deterministic Fibonacci-sphere directions at 10 deg
Hit definition	strict rotation error < 0.5 deg; fixed translation error < 0.20 m
Result	200/200 hits, 0 failures, hit rate 1.000
Rotation error	mean 0.2366 deg; median 0.2337 deg; maximum 0.4384 deg
Runtime	2,596.86 s wall time with four workers; 904,076 KiB measured process peak
Paper target	0.965 hit rate

The schema-valid benchmark digest is 03824378e8c36e28a2ecc7d63141dbe5bc704a60f756a5ce5b333e22203c08ea; the frozen protocol digest is a41eb9d532f7c10ea5c8cdfc3b444e7ebd0526ed389dae3fb56adf4eb6845ec4; and the Bull's Eye SVG digest is e36707ce1524a4cb3825816a4e34cb14d80ac9c9b0005a457aed626c056a687d. All 200 optimizer traces were retained.

The remaining frozen KITTI raw rotation levels were then executed without changing the solver or strict hit rule:

Initial rotation	Hits	Mean	Median	p90	p95	Maximum	Wall / peak RSS
1 deg	200/200	0.2439 deg	0.2404 deg	0.3156 deg	0.3530 deg	0.4906 deg	1:08:29 / 979,436 KiB
2 deg	200/200	0.2434 deg	0.2358 deg	0.3205 deg	0.3533 deg	0.4651 deg	30:09.92 / 930,104 KiB
10 deg	200/200	0.2366 deg	0.2337 deg	—	—	0.4384 deg	43:16.86 / 904,076 KiB
20 deg	178/200	3.1941 deg	0.2425 deg	17.0067 deg	28.4920 deg	36.8213 deg	41:37.67 / 945,608 KiB

All 600 newly executed traces validate and match their problem and protocol digests. All 200 20 deg solver executions completed; “178/200” is the strict recovery count, not a computational-failure count. Its low median but large tail shows that reporting only central tendency would hide the capture-basin failures.

Rotation	Protocol SHA-256	Benchmark SHA-256	Definition / Bull's Eye sidecar / SVG SHA-256
1 deg	f4546eb35e97fdc5a040c0f215a90f9099be72bd419c8ca4986bab431faad866	102480b7139a5dee41bd46182f5bcc1a0bd91bfc1c0cd25cc5b9a4cc592e51b9	a8c18bf3753338d8ded71ba3a268ea81cc43a6623050aeef2def88408139df38 / b2fb64ef01ef003bd9d6809893001b266fc83abed1854da62eab6ab6bc8da559 / d1d53ef2a6566d0bd438d4db807c930a4ec50b86a45a81428cfc5bcca79e6400
2 deg	53615a0b1d7756beaf6e68663080552718e32e2efd94643c557fb5e6b19c1881	eec11698117a41d494ef38ea347cc1620c62c7ec5f06be17e70322bf08ea99fc	84c6ea3e77b183f68b33723efe2953514aac15fc4eb1e9bc0c355b3e27c48a86 / e44ede39e48bd3b941b6ed1e6001008109f2022843fb4bc61929e968bdee2771 / ff97012c6f21b201b471b034bfeacb00e3e8ab733d6e608e622f7c5da8d67ee2
20 deg	67f182c2c6154199ade83872bf40abada8016019a5c36f94335309c4c278a65a	22833d64e1c99747dfe05fa40cc7c03960a0ab7c48ee9ce2e44748cb7a47f882	1a70621b9e69873a9255271acf3ca98365cc6e067e58eff168d5820323b6cf34 / 25d1d6e5d4b6252e29ba74fe22c0e0c6c42b254de94392b4f4fcaac069173778 / 30c2cd35487ed0bc3387c8725722c18be083f4ed5d4f4e2b5cfecb3d7ac5b7eb

The shared problem SHA-256 is 1fddd3e015e3a8b5a1356405a9296b247f6c2d80a040ea3fa5e7323572a14041.

Important comparability limits remain explicit:

• The paper does not publish its exact 25 frame IDs, code, checkpoint digest, histogram-bin count, or BOBYQA settings. Calibrex freezes its own uniform rounding policy, 32 bins, and bounded SO(3) pattern search before execution.
• The official FeatDepth repository says that the selected checkpoint was trained on KITTI raw, but does not publish enough training IDs to rule out overlap with drive 0018. The provider artifact records overlap as possible. The test-split online-refined checkpoint was not used.
• Raw FeatDepth inverse depth/disparity is the histogram feature. Converting it nonlinearly to nominal metric depth materially changes fixed-width histograms and is not the frozen reproduction path.
• This result uses dataset-provided extrinsics rather than independent metrology. It does not satisfy the KITTI-360 or multi-dataset final gates.

The external provider remains outside the ROS-independent core:

PYTHONPATH=src /path/to/featdepth-env/bin/python \
  tools/run_featdepth_provider.py /path/to/2011_09_30_drive_0018_sync \
  --featdepth-repository /path/to/FeatDepth \
  --checkpoint /path/to/fm_depth.pth \
  --output-directory /path/to/depth \
  --artifact-output /path/to/featdepth-provider.json

calibrex camera-lidar build-kitti-problem \
  /path/to/2011_09_30_drive_0018_sync /path/to/featdepth-provider.json \
  --output /path/to/problem.yaml
calibrex camera-lidar freeze-rotation-protocol /path/to/problem.yaml \
  --output /path/to/protocol.yaml --perturbation-count 200 --rotation-deg 10
calibrex camera-lidar benchmark-rotation \
  /path/to/problem.yaml /path/to/protocol.yaml \
  --trace-dir /path/to/traces \
  --definition-output /path/to/definition.yaml \
  --output /path/to/benchmark.yaml \
  --workers 4 --resume --required-hit-rate 0.965

KITTI-360 reproduction input lock

The second Phase 1 path uses the official KITTI-360 2013_05_28_drive_0000_sync sequence and right fisheye image_03. The 25 endpoint-inclusive uniform frame IDs span 0000000000 through 0000011517. Its frozen inputs are:

Field	Locked value
MiDaS model	dpt_beit_large_512, official MiDaS tag v3_1
MiDaS source	commit 1645b7e1675301fdfac03640738fe5a6531e17d6, MIT
Checkpoint	1,581,966,003 bytes; SHA-256 9e9e900747e9e8b3112df716979219836a27716277b3d0dc53889cbba8b82328
Image processing	Official aspect-preserving minimal resize, multiple of 32, bicubic return to the 1400×1400 source grid
Camera model	Official KITTI-360 MEI parameters, including radial-tangential coefficients
Motion compensation	Official azimuth-scaled scan uncurl model from kitti360Scripts commit 32f6d64eef27c32b52c542b4e95be8af9c9c6444
Pose coverage	Ego motion applied to 21 selected scans; four official identity fallbacks where an adjacent pose is unavailable
Depth feature	Raw MiDaS relative inverse depth; nonpositive bicubic overshoot is rejected, with no metric conversion or online refinement

The official KITTI-360 projection retains only the positive signed camera range. Accepting the rear optical hemisphere is a severe error for the side-facing fisheye: in the one-direction smoke test it biased the recovered rotation to 2.004 deg. Enforcing the official forward-hemisphere rule reduced the same result to 0.250 deg, below the paper's strict 0.5 deg hit threshold.

Comparability remains bounded. Borer et al. report a double-sphere fit but do not publish its parameters. KITTI-360 publishes an MEI calibration instead. Calibrex therefore uses the exact public MEI model and labels this as a reproduction with a declared camera-model substitution, not a bit-exact re-execution of unpublished intrinsics. The provider also declares possible training/evaluation overlap because the MiDaS training mixture includes KITTI-family data. The disposable Python 3.12 environment applies a recorded default_factory compatibility-only patch to MiDaS's pinned timm==0.6.12; model operations and weights are unchanged.

The external provider and motion adapter remain outside the ROS-independent core:

PYTHONPATH=src /path/to/midas-env/bin/python \
  tools/run_midas_kitti360_provider.py /path/to/2013_05_28_drive_0000_sync \
  --calibration-root /path/to/calibration \
  --midas-repository /path/to/MiDaS \
  --checkpoint /path/to/dpt_beit_large_512.pt \
  --output-directory /path/to/midas-depth \
  --artifact-output /path/to/midas-provider.yaml

PYTHONPATH=src python tools/motion_compensate_kitti360.py \
  /path/to/2013_05_28_drive_0000_sync /path/to/midas-provider.yaml \
  --calibration-root /path/to/calibration \
  --poses /path/to/poses.txt \
  --output-directory /path/to/compensated-scans \
  --manifest-output /path/to/compensated-scans/manifest.yaml

calibrex camera-lidar build-kitti360-problem \
  /path/to/2013_05_28_drive_0000_sync /path/to/midas-provider.yaml \
  --calibration-root /path/to/calibration \
  --lidar-directory /path/to/compensated-scans \
  --lidar-manifest /path/to/compensated-scans/manifest.yaml \
  --output /path/to/problem.yaml

calibrex camera-lidar freeze-rotation-protocol /path/to/problem.yaml \
  --output /path/to/protocol.yaml --perturbation-count 200 --rotation-deg 10
calibrex camera-lidar benchmark-rotation \
  /path/to/problem.yaml /path/to/protocol.yaml \
  --trace-dir /path/to/traces \
  --definition-output /path/to/definition.yaml \
  --output /path/to/benchmark.yaml \
  --workers 2 --resume --required-hit-rate 1.0

Phase 1 KITTI-360 reproduction evidence (31 July 2026)

The frozen KITTI-360 10 deg gate completed without threshold or code changes during execution:

Field	Observed value
Sequence / camera	2013_05_28_drive_0000_sync / image_03
Frame policy	25 endpoint-inclusive uniform frames
Perturbations	200 deterministic Fibonacci-sphere directions at 10 deg
Hit definition	strict rotation error < 0.5 deg; fixed translation error < 0.20 m
Result	200/200 hits, 0 failures, hit rate 1.000
Rotation error	mean 0.3812 deg; median 0.3924 deg; p90 0.4437 deg; p95 0.4538 deg; maximum 0.4676 deg
Per-trial runtime	mean 131.45 s; median 121.27 s; maximum 294.23 s
Full run	3:16:24 wall time with two workers; 2,025,128 KiB measured process peak
Paper target	1.000 hit rate

All 200 trace artifacts validate, pin the same problem/protocol digests, and retain the original hit decision. Frozen artifact digests are:

Artifact	SHA-256
Problem	6bf2cc157de930f58e891457564dbd522ecd5b30e037ef22044a9cbc3e8390de
Protocol	916b63970c16e9d170aa6deb68650ae6ac8f584a448253b29a85e738c8e2917f
Raw benchmark definition	eb6e44af8980cdcaa3903268670d56e10f8eed51f18692acb8da4eca57070439
Aggregated benchmark	80c099a476ba6da133af18568201f4d59b3bddf17f45168dc128620a423af238
Bull's Eye sidecar	e1a3424a79ccd994054b55ea07f348746dbc993d84c452b92a50a86bbd6cf2fe
Bull's Eye SVG	49207ba6d5ffdf44ce95070b38e8936371cac2004d6cd80dfc61cc698c4a9977

The remaining frozen KITTI-360 rotation levels were then run without changing the solver or strict hit rule:

Initial rotation	Hits	Mean	Median	p90	p95	Maximum	Wall / peak RSS
1 deg	200/200	0.3768 deg	0.3872 deg	0.4400 deg	0.4526 deg	0.4665 deg	1:37:56 / 2,023,108 KiB
2 deg	200/200	0.3834 deg	0.3920 deg	0.4427 deg	0.4473 deg	0.4751 deg	2:07:35 / 2,043,124 KiB
10 deg	200/200	0.3812 deg	0.3924 deg	0.4437 deg	0.4538 deg	0.4676 deg	3:16:24 / 2,025,128 KiB
20 deg	184/200	2.9900 deg	0.3882 deg	0.4533 deg	28.4409 deg	45.4505 deg	2:40:34 / 2,082,896 KiB

All 600 newly executed traces validate and match their problem and protocol digests. All 200 20 deg solver executions completed; its 16 misses are capture failures rather than computational failures. As on KITTI raw, the median alone would conceal a large failure tail.

Rotation	Protocol SHA-256	Benchmark SHA-256	Definition / Bull's Eye sidecar / SVG SHA-256
1 deg	c5368b28e42026a9b068d5250cce9f986333aef8159a4569bf8fd114056af3ad	9a111dcb28f0077939f36603cbb6da6da2626f44b80675cf8397332235f40470	91b8aba354722770515e7bfd714597218f460acdda5f251fceff0cf0c2a8d25e / fec05af81ed338cab0303672863d9a7145f3ee9c00fa72f8275179dab0c45661 / ca55d2480ab991e889cdefa01c982fe77609498dfae486719d0405220807f394
2 deg	ea4f74b1988ec12960369ad08ef509cb137cd159fd21d7bf45674bb5cad00645	2a07146b99f889ee1a24e73336980a6a1206cbe73c58e6e83c295db30165bc6c	06df5fdde202883c58ef766810ce88e36add24d245039dbf27113e3e39a1d9db / da0c1c006ff3c7709767bc2f750796524e16715aa40fa9bcfe3f692140818add / 08555135487a59f426d57bd43dc82ab86c32edb3ef3da77d0cd0f4b7df460d7c
20 deg	7c9cb75c8a46759c1e587a3a9b88a14b78baf2ec087e408f7b77b68cd1ea0f1f	e79d14b18b856e4ffb6a96bb70bc53820bcedaa17d0947d03af9c18695f5d726	bcc317265636a355770d2a967656d9c00f136f6ddc4a9ee996a0b3eaaa05baf8 / ec582e24352ce6e6f52004345e8ad2e1c14fabb7699f2a4ee98729205cd73eca / bb4aaa4cbf5f01df46a4149d9f007ae2073b638824094679f24c2778972cd7df

This clears the complete prespecified Phase 1 rotation-only matrix on both datasets. It does not clear the six-DoF, independent-rig, or learned accuracy gates.

A2D2 cross-family execution smoke

The small official A2D2 public pair was added as a deliberately limited third dataset-family execution check. Frames 60 and 61 from 20180810_150607 use the official undistorted front-left images, calibration, and view-filtered LiDAR NPZ members. The official tutorial states that the points are already mapped into the camera view, so Calibrex back-projects the distributed pixel row/column/depth into a z-forward camera frame and declares identity only as a pre-registration reference. This cannot provide independent extrinsic accuracy.

The pinned MiDaS v3.1 provider and five-direction 10 deg smoke produced:

Field	Value
Dataset family	A2D2
Frames	2
Perturbations	5 deterministic Fibonacci directions at 10 deg
Completed solver runs	5/5
Hits (rotation error < 0.5 deg)	0/5
Final rotation errors	1.016, 1.102, 24.766, 1.143, 1.183 deg
Wall time / peak RSS	40.68 s / 180388 KiB

The outcome is a cross-family FAIL, not a threshold-tuning opportunity. Two frames do not represent the 25-frame Borer protocol, the source point cloud is pre-registered, and one direction entered a grossly wrong optimum. The artifacts remain useful for detecting portability and objective-topology regressions:

Artifact	SHA-256
MiDaS provider	a18ad94ba17c1ede7cac668ce41609caa0a023563b64cf3c287e55686d26b55b
Generated LiDAR manifest	e0a79a049f06543064ce8f64f9a4bd9677ecfba11d3ef5f74e5df7c2483941b0
Problem	cdb7b1864235149a2859537f829ae0454d1e300abc3271d30d028746862ce447
Protocol	7f525397a88a6ab0a6514c20bf86f0f6e6dea7ba5a2c50ee12a8ab90da199929
Benchmark	ba4a5b65b23990ff129b442c87434b7e3f36818fe4b10792e5b266fb40465396
Bull's Eye sidecar	267b6791662403578d4d01e81b61b19d20d7353dfa5d59c08b355f4375dc1554
Bull's Eye SVG	57323bdafaf451efc342e01022715c660e3299f85e30a3465d4b983c863da77a

B. Learned-method reproduction

Reproduce the five UniCalib configurations without changing their splits:

Protocol	Training	Test	Perturbation
U1	KITTI Odometry 01–21	00	5 deg, 0.10 m
U2	KITTI Odometry 01–21	00	10 deg, 0.25 m
U3	KITTI Raw 2011-09-26	2011-09-30	10 deg, 0.25 m
U4	None beyond U3 model	KITTI-360 Test-SLAM	10 deg, 0.25 m
U5	None beyond U3 model	Waymo testing subset	10 deg, 0.25 m

Every learned run must lock the training sample inventory, checkpoint digest, depth-model name/version/weight digest, code commit, CUDA stack, seed, and train-data isolation declaration.

When code and weights become available, add the CLRNet View-of-Delft protocol as a distinct benchmark family. Preserve its single-frame, iterative large-error, four-frame rigid, and full-range settings separately. In particular, do not compare its sequence-level median result directly with a single-frame mean from another paper.

C. Independent Calibrex falsification

The independent suite prevents benchmark-specific overfitting:

• Keep the existing digest-locked KITTI raw 0005 test as a regression shard.
• Add at least three dates or rigs from KITTI Raw, KITTI-360, Waymo, and A2D2.
• Split by sequence/date/rig, never by neighboring frames alone.
• Prespecify signed SE(3) perturbations at:
• rotation: 1, 2, 5, 10, 20 deg;
• translation: 0.05, 0.10, 0.25, 0.50 m;
• coupled six-DoF errors using deterministic sphere/Sobol directions.
• Add scene strata for static/dynamic, day/night, texture, weather, FoV, LiDAR scan pattern, and projected-point support.
• Retain failures in every denominator.
• Freeze the suite before comparing new model variants.

D. Absolute and downstream evidence

Public dataset calibration is not absolute metrology. A final SOTA submission also needs:

• one target or surveyed-rig dataset whose reference uncertainty is declared;
• repeated remount/capture sessions;
• a downstream held-out metric such as depth-edge alignment or point-to-image reprojection that was not used by the solver;
• sensitivity to camera intrinsics, time offset, rolling shutter, dynamic objects, and occlusion.

FAST-Calib or another target-based method may generate an independent reference, but it remains in a separate result category.

Metrics and statistical gates

Every method reports:

• SE(3) translation distance and quaternion geodesic rotation error;
• per-axis errors only as secondary diagnostics;
• hit rate and area under the success-versus-perturbation curve;
• median, mean, p90, p95, maximum, and failure rate;
• runtime, peak memory, preprocessing time, and hardware;
• held-out objective and downstream projection metrics;
• per-stratum metrics and worst-stratum result;
• deterministic paired bootstrap confidence intervals;
• empirical uncertainty coverage when the method predicts uncertainty.

The following rules apply:

• Test data cannot select checkpoints, thresholds, depth providers, or optimizer parameters.
• A mean-only win is insufficient.
• A failed baseline reproduction cannot support a SOTA claim.
• A method must improve or match every primary safety metric; trading a lower mean for a higher catastrophic-failure rate is a rejection.
• Report dataset-reference error and independent downstream evidence separately.

Architecture

Core contracts

Add versioned, ROS-independent contracts:

• DepthMapObservation: metric/relative depth, validity, uncertainty, intrinsics, image transform, capture time, and source frame.
• DepthProviderArtifact: provider/model/version/checkpoint/license/input and output digests, scale convention, preprocessing, command, and environment.
• CameraLidarCalibrationProblem: observations, initial transform, bounds, frame/time conventions, and split IDs.
• CalibrationCandidateTrace: every seed, objective evaluation, accepted update, stopping reason, and output transform.
• CameraLidarBenchmarkProtocol: perturbation generator, exact samples, thresholds, train/validation/test isolation, and metric definitions.

All public APIs require type hints. Configs, traces, and results remain schema-valid. Generated results include source, input, config, model, and output provenance.

External boundaries

• UniCalib and Koide enter through external adapters with pinned containers or environments.
• CLRNet remains a watchlist method until its promised code, weights, and non-commercial-use terms can be inspected. If usable, it stays external.
• iKalibr enters as an external spatiotemporal baseline.
• GPL implementations such as MLCC, Livox Camera Calib, and FAST-Calib must remain subprocess/container baselines; no GPL code enters src/calibrex.
• Learned depth stays an optional provider. The native D2D solver consumes a depth artifact and does not import PyTorch, CUDA, ROS, or model code.

Research phases and go/no-go gates

Phase 0 — Benchmark and baseline lock (weeks 1–2)

Deliver:

• generic schema-versioned external calibration run artifact;
• dataset locks and protocol files for D2D A/B and UniCalib U1–U5;
• Koide and UniCalib adapters;
• immutable perturbation manifests and paired aggregation;
• baseline reproduction report with failures and provenance.

Gate:

• All artifacts validate and rerun from digests.
• Published baseline results are reproduced within a prespecified tolerance, or discrepancies are explained and the affected comparison is marked non-comparable.

Phase 1 — Native rotation-only D2D MI (weeks 3–5)

Deliver:

• solver-neutral depth-map contract;
• frozen MoGe-2/Depth Anything/FeatDepth provider artifacts;
• LiDAR depth projection with z-buffer visibility;
• per-frame D2D MI averaged across frames;
• bounded derivative-free SO(3) optimization;
• exact Borer perturbation protocol and Bull's Eye plots.

Gate:

• At 10 deg, achieve at least 96.5% KITTI and 100% KITTI-360 hit rate, with no test-set tuning.
• If hit rate is below 90% on either dataset, stop feature expansion and diagnose depth scale, visibility, FoV, and objective topology.

Phase 2 — Robust six-DoF analytical solver (weeks 6–8)

Deliver:

• deterministic global rotation seeds and bounded translation search;
• multi-resolution depth, robust histogram/MI estimation, occlusion masks, and dynamic-region downweighting;
• local SE(3) refinement and training-only endpoint safety gate;
• capture-range AUC and failure diagnosis.

Implementation checkpoint:

The entries in this Phase 2 checkpoint are historical execution milestones. The final v0.7 integrated-r5 results and digests in Current position and the v0.7 release-audit subsection supersede them for current comparisons.

• The first frozen real KITTI-360 six-DoF smoke completed from a paired (0.5 deg, 0.5 m) perturbation. The bounded native solver converged in 165.90 s and returned 0.2505 deg / 2.08e-17 m, satisfying the strict <0.5 deg / <0.20 m hit definition.
• The one-trial benchmark, definition, and full candidate trace are schema-valid. Their SHA-256 values are respectively 665fd6aa17ff6933e94cfb960358aa0cdc2ebaf5f2b223d629cb70b0066264f0, 8fa5ee01b70a6236637114d35626312da2af3d7453b58f2d1f5c1725a42a455a, and ef090ae8a642675a548096cafb606ebc747635e733e3bbf65009eaa6c617b78c. The frozen problem/protocol digests are f08ea1d8c98966db678216deece5d25fb3740d60708f6f0a5e56a751aa33cc31 and c62e20a304245e4ac63d1e064b2eae8106540358d46f22d9c829c8cb587ed2c4.
• This establishes executable real-data six-DoF recovery, not the Phase 2 hit-rate gate. The full 200-direction KITTI/KITTI-360 matrices and the remaining perturbation pairs are still required.
• The first full 200-direction cell subsequently completed on KITTI raw 0018 at (0.5 deg, 0.25 m). All 200 solver trials converged without a computational failure; 84/200 passed the strict hit definition (42%, bootstrap mean 95% CI [35.5%, 49.0%]). Native output errors were: rotation mean/median/p90/p95/max 0.5230/0.5239/0.6812/0.7228/0.8234 deg and translation 0.1447/0.1421/0.1681/0.1803/0.2389 m. Runtime mean/median/p90/p95/max was 106.54/102.19/142.13/154.81/191.00 s; the four-worker run took 1:29:19 wall time and peaked at 995112 KiB RSS. This cell falsifies a robust-recovery claim for the current native six-DoF search even at the smaller translation perturbation; it does not represent a computational failure.
• The schema-valid problem/protocol/definition/benchmark SHA-256 values for that full cell are respectively 452bc71fb9a5337414c81457ff8a4d705c6d681bb4d0e715cce095aa4bcda27f, c35bb1c226e002cf87b1ab604ac759fbd2250f184e49c3a6b362586bd496b709, 9477f613ef7b3518ca707df99ca27fbc1478a6bc6e29060fce35b73d64139f51, and 8dbc26f921468480470700fd309a8a3361333f71feab002786425e2bac1a9750. All 200 trace schemas, trial identities, problem/protocol identities, and benchmark source digests were independently checked.
• The corresponding KITTI-360 full cell also completed all 200 trials. Native D2D errors were rotation mean/median/p90/p95/max 0.3203/0.3062/0.5803/0.6943/0.8648 deg and translation 0.03069/0.03188/0.05199/0.05776/0.07050 m; 170/200 passed the strict hit definition. Mean/median/p90 runtime was 365.55/356.85/493.21 s, with a maximum of 631.15 s. Its schema-valid problem, protocol, definition, and benchmark digests are f097d9e4b7fcb7ddafb5fb3285687cdf9d5fcd8ce53e93c18b74cb469cf52d8e, b718eea98cdf114cdac5c9c360716308ec22d42ac13e4f27baf852f5e8a0e530, b6aaaef7fd56fcceacee8a2939d707663b3d04b931ac642148f2364ec1f3b23a, and 20f2993fdaa924edbbe4cd8303240697c71f6de6c534e198e4c1670e241502bf respectively.

Gate:

• Match or exceed D2D's published six-DoF hit rates: 88% on KITTI and 76.5% on KITTI-360 at (0.5 deg, 50 cm).
• Do not regress rotation-only performance or the existing KITTI 0005 runtime regression.
• If translation remains weak while rotation succeeds, pivot translation to explicit 2D–3D correspondence/PnP rather than adding more MI heuristics.

Phase 3 — Probabilistic depth correspondence (weeks 9–12)

Deliver:

• optional shared-depth encoder adapter;
• correspondence mean, covariance/outlier probability, reliability map;
• PnP-RANSAC or robust differentiable pose estimation;
• analytical D2D initialization followed by learned refinement;
• checkpoint and training-corpus provenance.

Implementation checkpoint:

• slac.probabilistic_correspondence/v0.1 now freezes provider/checkpoint and training-corpus lineage together with each 2D--3D mean, positive-definite image covariance, outlier probability, and reliability score. It records the dataset license separately from the provider license; paired benchmark provenance sets data_verified: false when the dataset license or source digest inventory is missing.
• slac.probabilistic_pnp_result/v0.1 records the input artifact digest, provider identity, robust-estimator options, OpenCV version/license, pose, RANSAC support, covariance-normalized inliers, and reprojection diagnostics.
• calibrex camera-lidar refine-probabilistic-pnp provides the first optional Apache-2.0 OpenCV adapter. --initial-problem passes a schema-valid D2D problem's declared initial pose as the iterative PnP seed and pins that problem's SHA-256 in the result; it is not represented as a solved D2D output. The core remains ROS/training-framework/OpenCV independent and accepts provider output only through the validated artifact.
• ProbabilisticCameraLidarRefiner consumes all provider frames behind that contract, starts from the declared D2D transform, fits one shared bounded SE(3), and reports disjoint frame holdout. Covariance, outlier probability, and reliability are individually frozen ablation switches rather than implicit model behavior.
• refine-probabilistic-multiframe --initial-trace TRACE uses the output pose of a schema-valid D2D candidate trace only after its problem digest and camera/LiDAR frame pair match the supplied problem. The result records the trace ID, SHA-256, solver status, and original hit decision. Without this option the source is explicitly problem_initial_transform, which is useful for controls but is not called a solved-D2D initializer.
• calibrex camera-lidar benchmark-probabilistic-ablation executes the full estimator plus without_covariance, without_outlier_probability, and without_reliability on identical frozen frame splits. Individual results, raw benchmark definition, failure-aware aggregates, and paired bootstrap intervals are all retained. Its optional --initial-trace pins the same actual D2D output across every method and split. Real U1--U5 provider output is still required before these ablations support a paper claim.
• calibrex camera-lidar benchmark-probabilistic-falsification joins digest-matched D2D candidates to probabilistic multi-frame refinement. It requires a schema-valid six_dof D2D protocol and either one trace for a focused smoke or a complete --trace-dir containing one trace per protocol perturbation. The full mode runs each D2D initialization trial with seeded disjoint capture-frame holdouts, evaluates twelve signed roll/pitch/yaw and x/y/z known-bad controls per split, and emits schema-valid results, assessment/protocol/policy sidecars, and a raw-recomputed evidence bundle. A known-bad threshold failure is retained as an explicit fail; it does not disappear behind a successful solver status. The single-trace mode is a diagnostic subset and is labeled as incomplete D2D coverage in the evidence.
• The initializer-preserving evidence contract is identified as camera_lidar_targetless_6dof_refinement/v0.2. A rollback remains a safe published pose, but it is a failed refinement trial rather than a successful no-op; method failure rate and paired confidence intervals therefore cannot be improved by silently returning the initializer.
• Every integrated run also emits failure-analysis.json. It reloads and digest-checks the protocol, D2D traces, and refinement results, then retains per-case D2D/refinement rotation and translation failures, p90 pose errors, holdout degradation, and explicitly non-causal diagnostic findings.

Example invocation:

calibrex camera-lidar benchmark-probabilistic-falsification \
  /path/to/correspondence.yaml /path/to/problem.yaml \
  --protocol /path/to/six-dof-protocol.yaml \
  --trace-dir /path/to/six-dof-traces \
  --output-dir /path/to/falsification

The complete real-data sequence is intentionally two commands. First execute and retain every fixed six-DoF D2D trial; then feed that trace directory into the probabilistic pack:

calibrex camera-lidar benchmark-six-dof \
  /path/to/problem.yaml /path/to/six-dof-protocol.yaml \
  --trace-dir /path/to/six-dof-traces \
  --definition-output /path/to/six-dof-definition.yaml \
  --output /path/to/six-dof-benchmark.yaml \
  --workers 4 --projection-backend numba_cpu --resume

calibrex camera-lidar benchmark-probabilistic-falsification \
  /path/to/correspondence.yaml /path/to/problem.yaml \
  --protocol /path/to/six-dof-protocol.yaml \
  --trace-dir /path/to/six-dof-traces \
  --output-dir /path/to/falsification

The falsification command verifies every trace against the problem and protocol SHA-256 values before running refinement, so a partial, stale, or cross-sequence trace directory is rejected rather than silently sampled. All traces in a complete directory must also share one exact solver/backend identity, preventing a numerically mixed run from entering the evidence pack. The optional numba_cpu backend must be selected explicitly; omitting the flag retains the dependency-free NumPy implementation. Backend identity and dependency version are pinned in each trace and cannot be mixed by --resume. The optional cuda backend accelerates projection and visibility on an NVIDIA GPU; use --workers 1. See GPU acceleration for installation, numerical checks and synchronized timing comparisons. - Each refinement result records the problem's digest-pinned reference transform plus initial/final Euclidean translation and quaternion-geodesic rotation errors. The paired ablation aggregates these reference errors together with held-out reprojection, so U1--U5 cm/deg thresholds and their paired confidence intervals can flow directly into the SOTA audit rather than being inferred from a pixel objective. - The current adapter is a typed, provenance-complete evaluation boundary; it is not yet evidence for the U1--U5 accuracy or runtime gates. D2D-seeded multi-frame refinement, paired synthetic ablations, and the integrated falsification/evidence pack are implemented; a real shared-depth provider, frozen U1--U5 executions, and paired real-data confidence intervals remain required.

v0.6 real-data execution log (10 August 2026)

This subsection is a historical record and is superseded by the v0.7 integrated-r5 release audit below. Its correspondence counts, six-DoF cells, and falsification outcomes must not be quoted as the current result.

The first real-data v0.6 lock uses 25 endpoint-inclusive frames from KITTI raw 2011_09_30_drive_0018_sync/image_02 and KITTI-360 2013_05_28_drive_0000_sync/image_03. The correspondence boundary is the Koide virtual-LiDAR SuperPoint/SuperGlue adapter, with its MIT Koide commit, non-commercial SuperGlue commit/checkpoints, per-file SHA-256 values, and dataset license recorded in the export and provider-neutral artifacts. The adapter is a subprocess-side tool; no external GPL or ROS dependency enters src/calibrex.

The regenerated provider artifacts contain 978 raw and 1,146 KITTI-360 correspondences. Both export manifests and provider-neutral artifacts validate against their schemas. The raw 200-perturbation six-DoF D2D source is complete; its integrated 1,000-case probabilistic falsification bundle is also complete and verification-valid. Its measured result is a policy fail: D2D initializes 425/1,000 cases, rotation failure is 575/1,000, translation failure is 65/1,000, and every probabilistic result is retained as insufficient_correspondences. A relaxed diagnostic confirms that the optimizer itself runs (292 iterations) but moves the raw estimate from 0.434 deg / 0.144 m to 3.867 deg / 0.258 m; this is provider-quality evidence, not a calibration claim.

A one-perturbation KITTI-360 smoke path first completed the full D2D-trace → probabilistic-refinement → failure-analysis → evidence-bundle chain with valid schemas and verification. The subsequent complete 200-perturbation path ran 1,000 probabilistic cases (five seeded holdouts per D2D trace) and retained every result. D2D initialized 850/1,000 cases; the probabilistic refiner converged in 0/1,000, with rotation and translation failure in 1,000/1,000. Mean final reference errors were 3.5427 deg and 0.4431 m; p90 values were 3.8161 deg and 0.4704 m. The known-bad pack contained 12,000 controls, with 7,463 mandatory supported detections and a 0.6219 materialized pass fraction, but its declared summary still failed. The overall assessment is fail for three policy gates (holdout support, known-bad controls, and the decision boundary); verification.json is valid with zero issues and all 1,007 bundle artifacts checked. This is evidence of the current provider/refinement failure mode, not a calibration claim.

The key KITTI-360 v0.6 artifact digests are:

Artifact	SHA-256
SuperGlue correspondence export	7e604969d46bbb5fb0f4916363d050b5d01f8804de23600bc51efeca4f7e61f4
Correspondence artifact	cf9956d872b0a96325f783c2774f59f35659d25175750d231969959fa20c7998
Six-DoF benchmark	20f2993fdaa924edbbe4cd8303240697c71f6de6c534e198e4c1670e241502bf
Falsification benchmark	15b727aee7c318cf5cabb8391927a51a3df149021e7f327ac84cf0e3637221a7
Failure analysis	461e39215fa7c95d37385b26aea84b5cafcd9ab965d4619f9608a8f1b2e75c0a
Evidence bundle	350703e3464ef54346ab36a9472e72deefa4b44aa03f85ce3915707fd9b4c7c0
Bundle verification	015b9b996672b0fb7688baf60ddd7731fff266c19d560fe20f446f926d22e401

External baseline comparison and any SOTA wording remain deferred until the provider/refinement failure tail is understood and independent accuracy gates are available.

v0.7 integrated-r5 release audit (11 August 2026)

The v0.7 release sequence freezes provider/PnP settings only on the disjoint KITTI-360 drive 0002 development sequence. Its 35-candidate, five-seed grid selects minimum_confidence=0.4 and RANSAC=32 px; the lock excludes both evaluation sequences, declares evaluation_data_used_for_selection=false, and keeps release_sota_claim_allowed=false. The multi-frame refiner does not reuse that PnP threshold. It retains the source-prespecified minimum_confidence=0.25 and fit-only acceptance policy as an untuned safety/failure evaluation.

The official evaluation artifacts contain 2,834 raw and 2,269 KITTI-360 correspondences over the same 25 endpoint-inclusive frames. Both D2D packs use the same 200 frozen (0.5 deg, 0.5 m) perturbations and strict <0.5 deg / <0.20 m gate. Independent matrix/quaternion auditing reproduced all 200 trace inputs, outputs, hits, solver identities, and both 200-trial protocol definitions. Final-source trial-000 reruns repeated 289 evaluations per dataset and were canonically exact after excluding runtime and execution provenance.

Each probabilistic pack contains exactly 1,000 current-run result files, with no missing, unexpected, or stale names. Its saved verification was recomputed from the bundle: all 1,007 artifacts and 203 raw inputs passed digest and linkage checks. Raw has 2,226 successful verification claims and 1,004 documented skips; KITTI-360 has 2,228 successful claims and 1,004 skips after also checking the saved verification record. Neither has a failed verification claim.

The principal immutable digests are:

Artifact	KITTI raw 0018 SHA-256	KITTI-360 drive 0000 SHA-256
Correspondence artifact	da5b8473717b5c2f8d51d80b89b4642819f97e22230c03d8e2e636ad1abaf625	2f7f6e45c3d9881c5320623ffe535919e7e34a023da8ee49d3f84b2455af5c0e
Six-DoF benchmark	da520214e3d9e6c183ef65eecfe4c2823bbd24629be440ce462f0e18e61bd329	307cea7137ecdddf15f206d4870030b5c2f2d236e9d947c2d3bf2f64888bf9c1
Falsification benchmark	c3db90c23cda6cd7e7b7e3f927bcb41f21eb985aaf45be11c1c70b520b29d607	f2430c4f23d386a817aa1e65e17e7ba5e7dc655f26b729b2629249d3611a21d0
Failure analysis	36be1e8278fed4c6176de99ad58fe1f07cc50db434fba962ecdc0103a38a78aa	057f7119abf01dde493f8ede367dded864e286c229c61d55d8719bf2f32ef605
Falsification bundle	f3d5bc9cc983598e6dba781504db0d34e520fc8cebe377c86387c133537aed4f	93ff9f8485b3ef9fbe379981c6a782c0c0de8c3543e5269590eecc6481c8253c
Bundle verification	723daccb7df404243e6c8664bc67f0830b42b46eb92e6b67ea1fc6aaafc80afd	305f6dc49219d52fe8f198a8fdbbd8f974f6a50b57cfea3019629adf6fb5d81a
Selected-pose quality report	51ac5da493a99b3ed1fc7e5a72bc9ddbca9afd43ddfc5431ccf657ef3927030a	133d9de875ce172084bb4e1123341a163d90f44b3442a87ee64e9e2f33863a25
Independent trace audit log	0d80f4a976653382acc63fc8241b0989e1cdeb7dac5eda9947284590c6caaf45	f3bc480c3bab3429988e557b3c17fd52a21fc56c0266b6c1f55bcf38811b7cfa

The final learned-targetless audit binds these artifacts plus the development lock, quality reports, external readiness audits, and nested verifications. It is schema-valid and returns refuted: raw D2D contradicts its hit and paired-rotation gates, both refiners contradict zero failure, KITTI-360 has no paired refined result, external comparisons are not executed, and independent rig coverage is zero. No SOTA claim follows from v0.7.

Release checks pass with Ruff, strict mypy over 206 source files, and the full 827-test collection (815 passed, 12 skipped). A clean Python 3.12 venv with system site-packages disabled installs the wheel, passes pip check, imports Calibrex from the installed wheel, and validates a final quality artifact. The wheel and sdist SHA-256 values are respectively 9d33fb35e2019fd64a99c30f23dcf64548284abc8f99dd7723b1950a652750d0 and b9a1652577d42251291d3803bea0110f744fa6d4467137cbc05baf9de114ed88.

Gate:

• On U1, beat the best-per-metric envelope: 0.550 cm / 0.042 deg.
• On U2, beat 0.913 cm / 0.053 deg.
• On U3, beat 2.368 cm / 0.084 deg.
• The paired confidence interval must support the improvement, and failure rate may not increase.
• Cross-domain U4/U5 must improve over the reproduced UniCalib baseline before claiming general SOTA. Otherwise claim only in-domain or training-free SOTA.
• Match or exceed UniCalib's roughly 91 ms end-to-end runtime on comparable hardware, or establish a better accuracy/runtime Pareto point.
• When CLRNet becomes reproducible, add a separate View-of-Delft gate. A multi-frame claim must beat 0.3 cm / 0.03 deg under its rigid 1 m, 20 deg setting without hiding per-frame failures behind only a sequence median.

Phase 4 — Continuous-time multi-frame refinement (weeks 13–16)

Deliver:

• continuous-time trajectory adapter and native factor boundary;
• per-point LiDAR time, camera exposure/readout time, and scalar clock offset;
• motion compensation, rolling-shutter support, and joint extrinsic/time refinement;
• observability and excitation diagnostics.

Implementation checkpoint:

• The existing capture-time evidence path already decodes per-point firing times, applies constant-twist deskew, estimates an injected scalar clock offset when geometric references exist, and rejects static motion.
• ContinuousTimeCameraLidarSolver now joins that timing model to covariance-weighted image/LiDAR correspondences and jointly refines a bounded SE(3) correction plus scalar clock offset. Training and holdout captures are disjoint, every objective candidate is retained, and the solver refuses to update when image motion per second is below the declared excitation threshold.
• slac.continuous_time_camera_lidar_problem/v0.1 and slac.continuous_time_camera_lidar_result/v0.1 freeze all timestamps, twists, optional piecewise SE(3) trajectory knots, transforms, uncertainty, optimizer settings, split evidence, and source digests. The piecewise_se3_linear_slerp/v0.1 adapter interpolates translation and shortest-arc rotation in a ROS-independent core, rejects extrapolation, and records the chosen trajectory model in every result. Existing slac.trajectory/v0.1 outputs can be attached without ROS dependencies via calibrex camera-lidar attach-continuous-trajectory PROBLEM TRAJECTORY --output ADAPTED_PROBLEM; the adapter verifies the body frame and pins both source digests. Before optimization, the solver verifies that the trajectory covers every LiDAR firing, rolling-shutter row exposure, and both ends of the configured clock-offset search. Problems are executable with calibrex camera-lidar refine-continuous-time PROBLEM --output RESULT.
• calibrex camera-lidar benchmark-continuous-time-ablation runs the full model against fixed-clock, no-per-point-time, no-rolling-shutter, and no-covariance variants on identical capture splits. Camera row coordinates, readout duration, and readout direction determine each rolling-shutter exposure time before LiDAR deskew. The command emits individual result artifacts plus the raw and aggregated paired benchmark. When the problem declares a frozen reference extrinsic and clock offset, every result and ablation also reports quaternion-geodesic rotation error, translation distance, and absolute clock-offset error; asynchronous real sequences remain the required external-validity gate. Dataset license is distinct from implementation provenance, and an undeclared license makes the paired benchmark data_verified: false.
• The native factor now supports rolling-shutter row times and a non-constant supplied trajectory, with constant twist retained as an explicit fallback. Trajectory estimation itself remains outside the calibration core. Schema-valid real-data run artifacts and proof of spatial improvement on asynchronous held-out sequences remain open, so the phase gate is not yet claimed.

Gate:

• Injected time offsets and extrinsics are recovered on held-out sequences.
• Static or weakly excited segments are rejected rather than silently updated.
• Spatial accuracy improves over Phase 3 on asynchronous real data.

Phase 5 — Paper-quality audit (weeks 17–18)

Deliver:

• ablations for depth provider, visibility, initialization, uncertainty, dynamic masks, frame count, and temporal model;
• exact container/checkpoint/data manifests;
• cross-rig and cross-dataset results;
• failure gallery and limitations;
• archived benchmark artifacts and one-command reproduction.

Audit implementation checkpoint:

• slac.camera_lidar_sota_audit_protocol/v0.1 freezes the declared category, every required benchmark method/metric/statistic/threshold, artifact path and digest, minimum dataset-family coverage, and independent-rig IDs.
• calibrex camera-lidar audit-sota PROTOCOL --output AUDIT verifies every digest and schema, evaluates the numerical gates, and emits supported, refuted, or incomplete. The result schema rejects a supported verdict unless all required gates and coverage minima are achieved.
• Frozen numerical requirements can address a benchmark metric's mean, median, p90, p95, maximum, or method failure rate. Benchmark artifacts now retain p90/p95 explicitly, preventing a favorable average from satisfying a claim whose tail-error gate fails. Requirements can also address the lower or upper paired-bootstrap improvement confidence bound for an explicit reference/candidate pair, making the Phase 3 confidence-interval gate machine-enforceable. Runtime and peak-memory mean/p95 gates use the same frozen benchmark and digest path. A schema-valid artifact is still contradicted as claim evidence when its provenance declares data_verified: false. Result-model validation requires refuted to have a contradicted required gate and incomplete to have unresolved evidence/coverage, in addition to the existing strict supported check. Only achieved required requirements count toward dataset-family and independent-rig coverage; optional smoke tests cannot satisfy claim scope.
• The protocol is an audit mechanism, not evidence creation. Missing learned, six-DoF, asynchronous real-data, or independent-rig runs remain incomplete; failed frozen gates remain contradicted and cannot be relabeled.

Final claim gate:

• Win the declared category on at least two dataset families and one independent rig.
• No primary metric, failure rate, or worst-stratum result is materially worse.
• All claims map to schema-valid artifacts and can be independently rerun.

Next implementation issues

The first three PR-sized issues are:

1. Implement a materially different cross-modal correspondence provider. The like-for-like comparison boundary is implemented, and drive 0002 plus the new disjoint drive 0003 capture both fail the geometric holdout gate. Outdoor/indoor SuperGlue checkpoint variation does not recover geometry. Replace the virtual-intensity matching construction, preserve checkpoint/license isolation, and require a schema-valid lock on both development captures before any evaluation rerun. Do not continue confidence or matcher-threshold retuning against the same zero-support family.
2. Refinement objective and acceptance calibration. On disjoint development captures with adequate train and holdout support, compare reprojection-only, geometry-aware robust SE(3), probabilistic PnP, and coarse-to-fine variants. Keep fit-only initializer-preserving acceptance, retain every rejected candidate, and freeze the provider/refiner thresholds before any evaluation rerun.
3. Comparable baselines and independent-rig evidence. Execute Koide and UniCalib on projection-compatible copies of the exact frozen protocol, recompute Calibrex metrics, bind checkpoint/train-test isolation, and add at least one non-KITTI rig. Reopen the SOTA audit only after raw capture, refiner failure, external comparison, and rig-coverage gates are all independently satisfied.

Expected outcome

The likely publishable Calibrex contribution is not “another calibration network.” It is a hybrid system:

1. training-free D2D global capture;
2. uncertainty-aware depth correspondence and PnP refinement;
3. optional continuous-time spatiotemporal refinement;
4. unusually strong provenance, falsification, failure retention, and cross-dataset reproducibility.

The next credible claim requires the raw D2D capture gate, both failure-aware refinement gates, exact external rescoring, and independent-rig coverage to pass together. Until then, the publishable contribution is the auditable calibration and falsification framework, not an accuracy-SOTA label.