NODEDC_MISSION_CORE/docs/audits/2026-09-20-physical-repeat-007-100m-results.md

Physical repeat ja-sun-007-100m — read-only acceptance review

Reviewed on 2026-09-20 after the operator reported a successful approximately 100 m walk. This is a retained-data inspection, not a replay, new capture, device operation, algorithm change, or navigation acceptance. No Ops mutations.

Identity and integrity

• Live run: ac07a83a-f784-4d31-a908-d1fd526921b4.
• Query: 20260920T054444Z_viewer_live, generation 1, display name ja-sun-007-100m; do not confuse it with the earlier unsuccessful ja-sun-007.
• Reference: JA-SADOVAYA-001, 20260911T085226Z_viewer_live.
• Draft: 672c1d9f-b817-4b78-b84c-b34a4edc0870, revision 1.
• Selected route: indices 0–1320, length 102.59360280377184 m; context map indices 0–1516, route-context-map/v2.
• All 133 declared run artifact hashes verified against report.json.
• Raw capture SHA-256: 1c8161b2d2e22ecb98ccd1bbe04f1dc348b8d79b1cab1a2cf94dcecf58ea929c.
• Metadata SHA-256: 8c8caa3e293fa459e423c815a08c84ced20433cefb55eb81d018e245dd1c5783.
• Raw and metadata hashes verified, all 3,858 message payload hashes verified; full streaming decoding found zero pose/cloud decode errors.
• Empty recovery journal and clock origin verified. The summary declares the sealed session clock mqtt.timeline.session-de7e1779fa341e2e81e0943d8523f09fc1003a0528b2f0342fa4ec86334d445b.json; its digest matches. The mutable transport mqtt.timeline.json is a different envelope and must not be compared to the session-clock digest.
• Query raw data contains 1,625 poses, 1,624 clouds and 5,573,004 cloud points. Capture summary records no rejected messages, recoveries, or transport error.

Evidence root is the canonical NODEDC_MISSION_CORE/.runtime/mission-core: run artifacts under missions/live-tests/<run-id>/, raw source under evidence/sessions/<query-id>/. Private raw evidence remains outside Git.

Timeline and registration

Times below are UTC; local operator time is UTC+3. Durations were calculated from retained monotonic timestamps.

Event	UTC
First pose / first cloud	05:45:11.013 / 05:45:11.024
Stationary prefix complete; initial search starts	05:45:21.048
Provisional initial match complete	05:45:36.390
First fresh validation accepted	05:45:38.881
Third fresh validation; tracking established	05:45:49.087
Last fresh validation accepted	05:47:41.109
Selected route distance reached; tracking authority cleared	05:47:43.034
Run completed	05:47:43.039
Operator STOP correlated response	05:47:53.360
Last raw pose / cloud	05:47:53.585 / 05:47:53.592

Tracking established 38.076 s after the first pose. The initial search took 15.337 s wall time and only supplied a provisional prior; it is not counted as an accepted fresh validation. All 25/25 fresh validations were accepted; there was no rejection, reacquisition, receipt-gap segment change, or loss between acquiring tracking and completing the selected distance.

Fresh-validation metric	Minimum	Median	Maximum
Overlap	96.658%	98.497%	99.638%
Inlier surface RMSE	0.13956 m	0.15095 m	0.16728 m
Worker wall time	0.26009 s	0.35009 s	0.47530 s
Latest input age at acceptance	0.30952 s	0.65603 s	0.92065 s
Change at current query position between accepted transforms	0.00710 m	0.01281 m	0.03758 m
Full rotation change between accepted transforms	0.04392°	0.13181°	0.30600°

The transform yaw ranges from 4.0242° to 4.4371°, a span of 0.4129°. This is the relationship between session frames, not an independently measured heading error. Similarly, overlap and inlier RMSE describe agreement of matched surfaces, not ground-truth rover position accuracy.

Final fit overlap is 98.390%, RMSE 0.13956 m. Its sampled accumulated distance is 98.463 m; the run continues using that fresh accepted alignment until 102.594 m, approximately two seconds later. Do not describe every metre of the last interval as separately re-registered.

Termination is distance-limit: the operator-selected route length, not a restored hard-coded 100 m cap. selected-live-route/v1 has no fixed maximum route length or time ceiling. Terminal tracking_state=lost is the intentional revocation at completion, not an in-motion localization loss.

The raw recording continued for about ten seconds after the planning run: scanner-reported distance at run end was 102.572 m, final scanner distance 104.463 m. Summing all raw positional increments gives 104.912 m, a different metric from the planner's 5 cm filtered path accumulator.

Cloud delivery and retained visual evidence

• Raw pose/cloud average cadence is approximately 10 Hz, including during tracking. It is not uniformly spaced.
• During tracking, cloud gap p50/p95/max = 0.08162/0.24342/1.00042 s; pose gap p50/p95/max = 0.08123/0.23606/1.65884 s.
• No raw pose/cloud gap exceeds 2 s. Across the full capture the longest cloud gap is 1.06501 s. Maximum adjacent raw pose displacement is 0.12333 m; the large timing gaps do not correspond to a coordinate teleport.
• At run completion the planning ingress reports lidar 1,521 published, 1,492 consumed, 27 overflow drops and 2 queued; pose 1,522 published, 1,519 consumed, 2 overflow drops and 1 queued. Raw recording is independently retained; these counts must not be equated with raw file loss.
• Fast display accepted 1,484 cloud updates. Final rolling display has 20 chunks and 39,238 points. It intentionally does not accumulate the entire route cloud.
• A read-only scan of the service access log found 1,703 requests for this run's scene delta endpoint: 926 HTTP 200, 777 HTTP 204, no HTTP errors, one base request. All observed requests kept reference/query/trajectory enabled in 3D.
• Request cursors include 802 distinct live cloud revisions, 842 distinct live pose sequences, and 25 distinct fit-evidence sequences. From the client implementation, subsequent cursors follow a successful send_rrd call. This supports repeated native-channel delivery, not merely production of backend snapshots. Counts are aggregate access-log evidence, not per-display frame timing or a GPU presentation acknowledgment.
• There is no retained measurement of actual browser paint FPS or end-to-end visible latency. The operator's uncertain visual impression cannot be closed as “smooth real time” from these counters alone. Green correspondence evidence changes with the approximately five-second fit, separately from fast cloud.

The camera archive summary is complete, with 1,534 segments and no failure code. Camera media was not fully decoded or independently hash-verified in this review, so no stronger video continuity/presentation claim is made.

Decision and remaining boundary

Positive physical laboratory result: stationary acquisition and continuing registration work across the selected approximately 100 m route. Quality did not collapse with distance, and the earlier reported approximately 10° sudden misalignment is not reproduced in the retained accepted transforms.

Fast cloud delivery is physically evidenced beyond the previous offline qualification, but subjective/paint-level visual acceptance remains open. Ingress drops and occasional 1–1.66 s input pauses are real limitations to measure, not grounds to call the entire acquisition smooth.

This does not establish arbitrary-start relocalization, robustness across mounting configurations, independent absolute localization accuracy, or an acceptable navigation freshness/stopping budget. The laboratory still retains localization_confirmed=false, vehicle_control=false; the 8 s alignment age gate is not a vehicle safety criterion. A longer walk alone does not close those questions. Next useful acceptance is repeated starts with known offsets and a controlled observation of cloud/pose presentation and stale-input behavior.

Only this audit was added. No implementation, runtime configuration, device, recording, Ops card, or service lifecycle was changed during review.