DCCONSTRUCTIONS
e515ab1b8c
Preserve the completed teach-and-repeat laboratory stage: reference preparation, cascaded acquisition, local tracking and recovery, recording lifecycle, replay qualification, and persistent Rerun scene controls. Document the open grid-picking regression and Rerun upgrade contract. No autonomous driving or loop-closure optimization is claimed.
4.6 KiB
Stationary physical start: initial acquisition did not complete
Read-only investigation of the operator's stationary attempt on 19 September 2026. No replay, numerical recalculation, threshold change, application restart, scanner command or new capture was performed during this review.
Evidence and method
The completed live report, input arrays, per-hypothesis registration results,
receipt metadata and raw transport were inspected together with the current
entry-search and presentation code. A private immutable copy is retained at
data_dir/missions/diagnostics/20260919-physical-start-001: 37 files, 26,806,716
bytes, all SHA-256 verified. Manifest SHA-256:
bf1f5810cf9052fe4766d862c230e323ea1d66b5e66293a31eac2a0d6ea84b20.
Its UTC/monotonic creation time, source identifiers and operator account are
retained privately. The code copy records the inspected dirty working tree,
not independent attestation of the original running process's build.
The selected reference was the previously recorded second pass, first 30.050 m, 43,886 reference points. It differs from the first-pass reference used by the earlier successful archived qualification. The reference trajectory contains the already known 10.407 m step; this review does not establish that the step caused the initial-search rejection.
Observed sequence (UTC)
- 13:33:14.066: collection of the stationary prefix.
- 13:33:24.366: bounded entry search.
- 13:33:51.695:
lost / initialization-rejected. - 13:37:46.709: input ended; no tracking had been established.
The ten-second prefix contained 188 derived source events, 6,098 query points and maximum K1-reported movement of 0.006441 m. The stationary gate passed. This movement is a scanner estimate, not an independent physical measurement.
Search completed 47 of the required 108 hypotheses in 25.234 s against a 25 s
limit. The outer worker finished normally after 27.221 s; its 30 s process
timeout did not fire. The persisted rejection is incomplete-search, with no
selected candidate and no confirmation on fresh windows.
All 47 hypotheses were rejected. Two converged fits had overlap of 96.749% and 96.603%, with inlier RMSE of approximately 0.165 m. Both were rejected solely by the local correction bound: 3.087 m and 4.901 m against 3 m. These corrections are measured at the observed patch centre relative to each artificial seed; they are not the operator's physical error relative to the recorded start. Overlap counts query points within 0.5 m of a reference point. It is not independent localization accuracy or proof of unique place recognition.
Seed iteration is lexicographic over translation and yaw. Indices 0–46 were evaluated. The central translation (0,0), beginning at index 48, was not reached. No accepted cluster existed at termination. Raising the time limit alone is not evidence that the full search would pass support, ambiguity and fresh-data checks. The final report's zero overlap is the rejected-result fallback, not evidence that the actual clouds had zero overlap.
Additional findings
The bootstrap performs one initial search. Its rejection terminates acquisition; waiting longer does not retry it. Both backend and frontend presentation map this initial failure to the generic lost-binding wording. The observed message incorrectly suggests that previously established tracking was lost.
The derived planning ingress recorded 102 overflow drops out of 851 published LiDAR events and two out of 889 pose events. An empty receipt-gap list applies only to delivered pose-window gaps; it does not establish lossless ingestion. These counters do not establish raw-recorder loss, and they have not been isolated as the cause of this rejection. They require a separate bounded check.
Decision and next stage
The physical start has not passed acceptance. It failed in initial acquisition, before tracking or movement, and is not evidence of localization loss during a walk. Preserve the passing stationary observation and the failed result.
Next, use the frozen input to qualify a complete, bounded search, examine seed ordering and repeated preprocessing cost, and retain the multistart support, ambiguity and fresh-confirmation gates. Compare against the previous positive and wrong-reference controls. Distinguish unsuccessful initial binding from loss of established tracking in the shared product status. Verify derived ingress under the same bounded scenario before another physical walk. Do not weaken acceptance thresholds merely to turn this example green.
Only this diagnostic report and private evidence were added. Product code and running device state were unchanged; no new algorithmic acceptance is claimed.