Temporal agreement does not resolve the wrong real surfaces

Frozen 53c39a61, evaluated2026-10-30. See protocol, plan, 18 reprojection controls, all18 centers/36 neighbor checks and schema.

Reject this temporal verification as a remedy for the remaining surface failures. All18 exact-render translation/visibility controls pass, minimum visible-point consistency98.34%, above the frozen95% threshold. Real data, however, retains incorrect surfaces after both-neighbor temporal agreement and loses useful support.

Across18 centers	Predictive surface baseline	Verified in both neighbors
Centers passing declared quality/support screen	15/18	13/18
Median domain retained	24.77%	19.43%
Median normal failure	11.84%	11.70% among17 nonempty centers
Median distance failure	9.14%	11.08% among17 nonempty centers

The candidate has one empty center, so its conditional precision median excludes that undefined value and must not be read as an all-case improvement. Full screen counts include the empty center as a failure. Dense points are not independent recordings; no significance or calibration improvement claim follows.

Original failures remain

Frame	Both-neighbor selected points	Domain retained	Normal agreement	Distance failure
0005/frame38	49861	11.49%	45.24%	20.69%
0005/frame114	99418	22.92%	76.02%	34.75%
0009/frame111	80656	18.59%	76.98%	10.26%

All three remain below80% normal agreement. Additional failures are0013/frame107, with both motion pairs unsupported and therefore zero verified points, and 0015/frame74, whose selected distance failure is21.79% despite adequate normal agreement. No missing support is relabeled as success or omitted.

The report partitions every selected point per neighbor into invalid projection, out-of-view, missing depth, consistent, nearer observed surface, farther observed surface or unsupported motion. Category totals equal baseline counts in all36 checks. Nearer observed surface can be real occlusion; rejected temporal support is not automatically incorrect geometry. Both-neighbor verification deliberately requires both tests; no one-sided or reference-assisted fallback was introduced.

Interpretation

Temporal projection uses measured predictive surface XYZ, estimated camera-only motion and raw neighbor stereo depth. The target camera/LiDAR reference transform is never used in that classification. Reference distance/normal maps score the result afterward. Motion and depth nevertheless share stereo images and errors; a coherent geometry error can be temporally consistent. Current results do not prove which failures arise from motion error, correlated stereo bias, occlusion, LiDAR normals or reference/timing assumptions. Exact controls validate projection on static clean scenes, not every real dynamic/noisy condition.

Four focused tests, repository Ruff and strict documentation build pass. Tests check projection direction, positive-depth/out-of-view/missing categories, nearer/ farther surface sign and separate error denominators. Checks validate report/schema, all hashes,18 control outcomes,18 real centers,36 neighbor-count partitions and baseline-versus-selected population totals. No extrinsic fits, ICP runs, fresh data, default changes or SOTA claims occurred.

Decision

Close the current self-consistency-filter sequence: plane predictability, reverse stereo and temporal agreement cannot establish reliable geometry on all18 frames. Do not tune tolerances or add another mask based on these reference-error maps. The next useful work is reference-assisted error attribution using independent measurement channels, explicitly evaluation-only. First audit available OXTS motion/timing and the camera-frame conversion uncertainty against the estimated motion; reference-based conversion is allowed for diagnosis but must never feed selection. Such a proxy is not independent metric ground truth and cannot reopen an accuracy/SOTA claim. If this cannot localize the error, pursue a different measurement construction rather than extending the filter chain or ICP iterations.