Skip to content

SOTA-v6 recovery plan after the SOTA-v5 blind result

Files this record names under configs/sota_v6/, configs/voxel_slam_v33–v41 and scripts/audit_v44g_raw_shadow_failure_profile.py were removed after v0.11.0. They remain in the v0.11.0 tag.

The SOTA-v5 evidence bundle is complete and negative. The result is not eligible for a SOTA claim: trajectory, runtime, memory, and map-geometry gates failed. Completion and the verified-false-loop gate passed. The independent visual track also failed and remains separate from the primary LiDAR–IMU claim.

The exact machine-readable plan is configs/slam_benchmark_profiles/sota_v6_development_recovery_plan.yaml. Its evidence source is the SHA-256-bound SOTA-v5 bundle at /home/user/workspace/old_~2026/lidarslam_ws/sota_v5_evidence_bundle_6284e1d6.

What the result says

The candidate geometric-mean APE was 24.9877 m; Voxel-SLAM was the strongest eligible suite rival at 2.7160 m, for -820.02% improvement. NavINST missed real time at RTF 1.13475. Candidate peak RSS was 916.48 MB; Point-LIO used 291.75 MB, so the limiting ratio was 3.1413, above 1.20.

Postblind stage scoring rules out graph correction as the trajectory-error source. On repetition 1, raw frontend and final corrected APE were respectively 2.81947/2.81947 m on NavINST, 6.97955/6.97953 m on Oxford, and 792.83954/792.83954 m on UrbanNav. Frontend accuracy is therefore the first workstream. Conversely, the Oxford process-tree peak attributed 744864 KiB of an 798.49 MB run peak to the graph node, making backend storage and queue state the resource workstream.

Development and promotion policy

NavINST Indoor02, Oxford Spires Keble 05, and UrbanNav HK Tunnel 1 are now development/regression data. They must never be described as blind in a later claim. Development uses one global candidate configuration, three repetitions, and the same trajectory, completion, loop, resource, and map checks.

Frontend promotion requires GM APE <= 2.63451 m, 100% completion, and no single sequence more than 10% worse than its best fixed rival. Backend promotion requires peak RSS <= 330 MB, RTF <= 0.85 on every development sequence, zero verified false loops, and no trajectory change above 1e-6 m for resource-only changes.

Only after both workstreams pass do we audit the complete dataset exposure history and choose at least three fresh cross-dataset holdouts. Their sensor and ground-truth identities, candidate/rival revisions, thresholds, machine, and run order are preregistered before execution. Ground truth remains closed until all three-repetition runs are terminal and raw outputs are hash-sealed.

The leo-drive/color-point-cloud project is retained as a later reference for calibrated color-map visualization in the independent visual track. It does not affect the primary trajectory claim or the closed SOTA-v5 result.

Initial NavINST development screen

These are single-repetition, map-save-skipped development screens, not promotion evidence. The common frozen reference and interpolating scorer were used after each runtime-only run completed.

Global case APE RMSE (m) RTF peak RSS (MB) Decision
v5 plain solver 2.81585 1.09813 519.42 reject
older balanced profile 3.18072 1.21282 968.72 reject
0.35 m voxel, 20 points/voxel, v5 budget 3.26439 1.08259 739.52 reject
0.35 m voxel, 8 points/voxel, v5 budget 3.17611 1.10097 565.37 reject
0.35 m voxel, 4 points/voxel, v5 budget 2.98495 1.09504 510.26 cross-dataset screen, then reject
0.35 m voxel, 1 point/voxel, v5 budget 3.18896 1.05518 467.94 resource signal only

The initial table previously mixed the scorer's mean column with the preregistered RMSE primary metric. The table above is corrected to RMSE. None of the retention cases improves the SOTA-v5 NavINST RMSE (2.81947 m). The 8-point case does reduce the fine-voxel case peak by about 23.5%, but still fails the development resource gates. Its peak snapshot assigns approximately 357 MB to RKO-LIO and 137 MB to the graph node, so frontend voxel retention and graph input-cloud size must be reduced independently.

The unchanged 4-point configuration was screened on all three development datasets. RMSE was 2.98495/5.80056/780.74836 m on NavINST/Oxford/UrbanNav, respectively: -5.87%/+16.89%/+1.53% against the v5 candidate. Its geometric mean improved only 4.67%, from 24.98771 to 23.82169 m, while remaining far from the fixed rivals and regressing NavINST. The configuration is rejected.

Symmetric loop-verification screen

The relaxed 8 m correction-cap diagnostic showed that source-only overlap was unsafe: five nonlocal edges entered the final graph, but four failed the frozen reference-relative residual limits. The backend now offers an opt-in harmonic forward/reverse overlap gate. The historical source-only behavior remains the default. The trajectory runner was also corrected to retain a /modified_path published by a loop event when map saving is skipped; previously it silently replaced that corrected path with raw odometry.

On the same NavINST 4-point frontend, the symmetric 0.60 gate produced two accepted proposals that deduplicated to one final nonlocal edge, 140 -> 421. Its reference-relative residual was 0.61465 m / 3.79443 deg, within the frozen 1 m / 15 deg limits, so the final graph had zero verified false edges. The raw trajectory was byte-identical to the source-overlap run. The corrected APE RMSE improved from 2.98495 m to 2.90371 m (2.72%), with 577/577 submaps processed, RTF 0.82812, and 510.32 MB peak RSS.

This validates the symmetric gate as a backend safety improvement, but not the 4-point configuration as a promotion candidate: it remains 2.99% worse than the SOTA-v5 candidate on NavINST and exceeds the 330 MB backend development memory gate. No fresh blind data may be opened from this result.

A follow-up combined the same symmetric backend with the more accurate plain frontend. It completed 574/574 submaps at RTF 0.82920 and 525.00 MB, and its raw trajectory was byte-identical to the earlier plain screen. No loop passed the mutual 0.60 floor, so corrected and raw RMSE were both 2.81585 m. The closest rejected candidate measured 0.599791. The threshold is not relaxed post hoc on exposed development data; this combination is also rejected.

Plain-frontend resource screen

The accurate plain frontend was then held byte-fixed while resource costs were removed one at a time. Limiting the lossless offline LiDAR read-ahead buffer to four frames reduced the 525.00 MB mutual-backend run to 497.54 MB. Spilling graph submaps to a run-isolated PCD cache and disabling the unused benchmark static-TF helper reduced the combined peak to 417.34 MB. Both runs retained the same raw trajectory SHA-256, 3bfd377131cd7a15bd479453a9ba77e1f18b31cde177a70c4d76c951610c70fd. The exact frontend input to these runs is now checked in as configs/sota_v6/development/rko_lio_plain_memory.yaml; its SHA-256 is 426ee9e45be23ab24cac35a0c53361fb2d1d6bd9d7945b2422537309d1e355c1.

RKO-LIO previously reserved the full configured 20-point capacity on the first point inserted into every voxel. Starting each vector at one point and allowing normal vector growth preserves the accepted points and their insertion order. The NavINST run remained byte-identical at APE RMSE 2.815850 m, completed 574/574 submaps with no loop event, and measured RTF 0.82645 and 354.91 MB. This is a 32.40% peak reduction from the mutual plain baseline. At the peak, launch/RKO-LIO/graph contributed 40.71/226.46/87.73 MB.

This is a useful resource-only improvement, but it still fails the preregistered 330 MB development gate and is not promotable. An inline one-point voxel representation increased the peak to 382.20 MB and was reverted. Accuracy- changing shortcuts were also rejected: 50 m range reached 263.80 MB but raw RMSE regressed to 3.15948 m and introduced one verified false edge; 75 m range reached 306.43 MB but RMSE regressed to 3.05200 m; and four points per voxel reached 304.48 MB but raw RMSE regressed to 3.21283 m. No fresh blind data is opened from this screen.

Cross-dataset driving-bridge screen

A single-repetition frontend-only screen tested whether the default-off anchor-decayed inertial bridge could recover UrbanNav's tunnel distance without changing the global configuration per dataset. These are exposed development results, not promotion evidence. The first 12.5 m/s vehicle-envelope run reduced UrbanNav RMSE from 792.83954 m to 129.46159 m, but the same configuration misclassified the walking NavINST sequence and regressed its RMSE to 40.54768 m. Raising the minimum speed to 7 m/s preserved NavINST exactly (2.81947 m, zero bridge attempts) but established UrbanNav anchors too late and regressed it to 864.58873 m.

The bridge now has a default-compatible minimum-IMU-density gate. Unit coverage confirms that a 30-sample minimum rejects a 13-sample interval and accepts 30/40-sample intervals; observed initialization intervals were 13 on NavINST and 40 on UrbanNav/Oxford. This is retained as a general safety mechanism; the default of one sample preserves historical behavior. It did not by itself make the vehicle bridge globally applicable: the corresponding UrbanNav/Oxford single runs measured 179.10298 m on UrbanNav and 229.77000 m on Oxford, versus the v5 Oxford baseline of 6.97955 m.

Additional range-scene applicability screens were negative. A 0.12 near-return fraction plus 30 consecutive trusted scans preserved Oxford at 6.97824 m with zero bridge attempts, but its 60-second cooldown also disabled UrbanNav (792.83613 m). Removing the cooldown improved UrbanNav only to 416.75242 m; a 0.15 threshold measured 428.60424 m. Thirty-scan moving averages measured 499.67152 m at 0.12 and 713.94928 m at 0.18. One-shot initial-scene latching rejected UrbanNav at both 0.12 and 0.18 and was reverted together with the moving-average implementation. Horizontal-only bridge corrections (216.92329 m) and gravity-window alignment (646.307 m) were also worse than the initial vehicle-envelope run and were reverted.

The checked-in configs/sota_v6/development/rko_lio_driving_inertial_bridge.yaml is therefore an explicitly retired reproduction control, not a candidate. No fresh blind holdout may be opened from this screen. The next frontend candidate must solve applicability without using dataset identity and must pass all three exposed development datasets before the preregistered three-repetition promotion gate.

Texture and physical-observability frontend screens

Two further default-off frontend directions were evaluated on the same consumed development inputs. Neither is a promotion candidate, and no fresh blind holdout was opened.

The oriented reflectivity/height disagreement matcher had valid UrbanNav input (intensity float32 and reflectivity uint16), but failed its pre-accuracy diagnostic gate. At full weight it corrected 3,243/3,964 scans, produced 160 search-boundary shifts, raised peak RSS from 221,928 KiB to 472,860 KiB, and shortened the estimated path from 1,263.53 m to 1,121.41 m. A zero-weight observation run preserved the baseline path but still measured a negative longitudinal intensity-minus-ICP velocity median in every 500-scan segment, with 2,674 negative versus 481 positive disagreements. It also used 394,236 KiB. The signal therefore points opposite the required tunnel-motion recovery and violates the memory gate; accuracy was intentionally not scored.

A point-to-plane residual exposed the geometric null space hidden by the historical point-to-point translation block and improved UrbanNav APE RMSE from 792.83613 m to 596.12606 m (24.8%) at RTF 0.205 and 218,660 KiB RSS. It was not globally applicable: NavINST regressed from approximately 2.82 m to 42.78942 m. Separating the objectives—retaining point-to-point pose optimization and using point-to-plane only for observability—reduced that NavINST regression to 3.17056 m but did not remove it. Weight 0.10 measured 3.14657 m. A 120-scan persistence gate still admitted one NavINST intervention and measured 3.18856 m. Raising it to 150 rejected UrbanNav too: zero interventions and 793.09148 m RMSE. Weak-axis/motion alignment did not provide a separator (median cosine 0.914 UrbanNav versus 0.935 NavINST). The entire experimental implementation was therefore removed rather than retained as a dataset-tuned switch.

Covariance-aware registration screen

An opt-in scan-to-model GICP backend based on pinned small_gicp source was then screened on the consumed UrbanNav development input while preserving RKO-LIO deskew, IMU pose prediction, state handling, and the legacy map and recovery paths. The best sparse setting (0.5 m correspondence distance, 1,500 keypoints) completed all 3,982 poses and improved the baseline RMSE from approximately 792.84 m to 707.59656 m (10.75%), at RTF 0.247 and 304,676 KiB peak RSS. It was nevertheless substantially worse than the already rejected point-to-plane screen (596.12606 m) and remained far outside the accuracy required for promotion.

Two bounded follow-up screens were negative. Increasing the correspondence distance to 1.0 m measured 720.73303 m RMSE, RTF 0.248, and 309,464 KiB; dense sampling with 3,000 keypoints measured 723.20938 m, RTF 0.317, and 311,860 KiB. The candidate therefore failed the exposed-development pre-accuracy gate before NavINST or Oxford evaluation. No fresh blind holdout was opened, and the backend, dependency, tests, and development configuration were removed rather than retained as repository complexity.

Independent scan-to-scan diagnostic

A default-off diagnostic then integrated a second trajectory by registering each deskewed scan only against its immediate predecessor. It never fed back into the primary estimate, and was intended to test whether scan-to-map self-locking—not the scan geometry itself—caused the UrbanNav failure. Both trajectories completed all 3,982 poses. The unchanged primary trajectory measured 792.83613 m RMSE and a 1,263.53 m path; the independent scan-to-scan trajectory measured 790.30935 m RMSE (only 0.32% better) and an even shorter 1,126.17 m path. The tunnel-axis ambiguity therefore persists without the accumulated map. This pre-accuracy diagnostic was removed, and no NavINST, Oxford, or fresh blind input was opened.

Translational inertial-state screen

A bounded 9-state translation/velocity/body-accelerometer-bias filter was screened next. It used IMU propagation and point-to-plane translation information, projecting its correction only into weak LiDAR axes while leaving the established ICP orientation and strong translation axes untouched. The first UrbanNav run completed all 3,982 poses at RTF 0.151 and 212,180 KiB peak RSS. RMSE improved from 792.83613 m to 459.94448 m, and estimated path length recovered from 1,263.53 m toward the 3,152.15 m reference at 3,215.53 m. However, the result depended on a nonphysical final residual accelerometer bias of approximately [-2.89, -0.56, 0.34] m/s^2 and accumulated a 288.58 m z-span versus 31.79 m in the reference.

Gravity-axis rejection improved RMSE to 398.52652 m, but retained a 270.49 m z-span and a 3.88 m/s^2 horizontal bias magnitude. Projecting each accepted correction exactly onto the world horizontal plane measured 378.43759 m, but map feedback still produced a 320.21 m z-span and the estimated bias grew to 5.38 m/s^2. Finally, enforcing a physically bounded 0.5 m/s^2 residual-bias norm caused runaway (909.67430 m RMSE, 1,371.89 m z-span). The apparent accuracy gain was therefore inseparable from an invalid state estimate. The implementation, tests, parameters, and development configuration were removed; NavINST, Oxford, and fresh blind data were not opened.

This result closes the lightweight RKO-LIO post-ICP fusion route. The next frontend candidate should start from a tightly coupled estimator that already meets the exposed NavINST/Oxford accuracy regime, then add a globally applicable Urban tunnel observability treatment and the existing verified-loop backend.

Voxel-SLAM degeneracy-reset diagnostic

The fixed Voxel-SLAM revision 70fc8a28d63823d5989ff184daeea0787b672398 already contains a tightly coupled iterated error-state Kalman filter. Its point-to-plane normal matrix is also used for a degeneracy test: the odometry thread increments a counter when the smallest normal-information eigenvalue is below 14 and calls system_reset() after the configured bound. On all three SOTA-v5 UrbanNav repetitions, the first reset occurred at the same sensor time (1621322564.65, within milliseconds), followed by exactly 156 resets while the moving-platform initialization repeatedly failed.

A bounded, single-repetition development diagnostic changed only the global odometry degrade_bound from 100 to 100000, leaving the pinned image and all sensor parameters unchanged. It reproduced the fixed run exactly through the original reset instant (pose [-65.629, 355.006, -41.457] m) and confirmed that reset suppression did not recover the weak direction. The estimate then ran away to [-12236.830, 34774.239, -28525.708] m, accumulated a 47,638.646 m path, and terminated by SIGSEGV after 3,010/3,980 poses. Peak RSS was 1,705.19 MB; the incomplete-overlap APE diagnostic was 11,603.25 m RMSE with 96/395 reference poses necessarily rejected. The original fixed run's full online trajectory path was 1,976.977 m.

This establishes that the reset is a downstream safety response, not the primary accuracy fault. The diagnostic configuration is removed rather than retained as a candidate. A viable Voxel-SLAM derivative must constrain or project the LiDAR update in weakly observable eigendirections before state and map corruption, while preserving IMU propagation; simply disabling recovery is rejected. No NavINST, Oxford, or fresh blind data was opened.

The first such bounded derivative projected the six-dimensional LiDAR normal equations onto translation eigenvectors whose normal-information eigenvalue met the same upstream 14 threshold. It retained rotation measurement terms, the full propagated IMU prior, and the original reset path whenever fewer than two translation axes remained observable. The exact development source and image were SHA-256-bound as cd3af167...897212 and fc6be7bf...fc1bc1, respectively, against the same fixed base revision.

This instantaneous hard projection also failed its UrbanNav pre-accuracy screen. It crossed the original reset time with zero resets, but by sensor time 1621322613.47 the estimate had already reached approximately [-585, 2640, -844] m. The run was deliberately stopped once failure was unambiguous: after 2,291 poses its endpoint was [-1046.171, 10885.765, -5631.733] m and its accumulated path was 12,424.977 m, still with zero resets. Because the run was intentionally terminated, no completion, APE, or resource result is claimed. The failure shows that a per-scan eigenbasis rotates with the already-corrupted map and does not provide a stable world-frame weak-axis constraint. This derivative is rejected before cross-dataset evaluation; no fresh blind input was opened.

A subsequent code audit found that LocalBA optimizes all 15 state dimensions (R/p/v/bg/ba) but upstream copies only R/p from the terminal window state back to online x_curr. A one-change diagnostic replaced that partial copy with the complete state assignment. The exact patched source and development image were bound as 1ca86c50...e3 and 063b6a27...160d. Although this closes a real state-handoff inconsistency, it did not prevent the UrbanNav failure: the first reset moved only from sensor time 1621322564.65 to 1621322565.45. Immediately before the original reset time, the diagnostic had a 366.765 m path and endpoint [-65.880, 356.027, -41.393] m, versus 365.718 m and [-65.629, 355.006, -41.457] m for the fixed run. The run was stopped after reset onset, so no completion or APE result is claimed. Complete LocalBA state handoff is insufficient by itself and is not promoted or tested on another dataset.

Voxel-SLAM weak-axis bridge screens

Four follow-up derivatives used Voxel-SLAM's own normal-information spectrum as a dataset-independent applicability signal. After ten consecutive scans with exactly one translation eigenvalue below 14, the bridge retained the two observable LiDAR directions and propagated the weak direction from a causal entry state. Fewer than two observable translation axes retained the upstream reset response. These are single-repetition UrbanNav development screens, not promotion evidence.

The first body-stable form fixed the weak axis in the sensor body at entry and therefore followed subsequent IMU rotation. It completed all 3,980 poses with zero resets and improved RMSE from fixed Voxel-SLAM's 818.80010 m to 496.55993 m; it also beat fixed Point-LIO's 638.93870 m on this sequence. If Voxel-SLAM's exposed NavINST/Oxford results were preserved, the three- dataset geometric mean would be 2.29897 m, about 15.35% better than the fixed Voxel-SLAM geometric mean. This is the first material accuracy signal from the recovery work, but it is not promotable: body-axis gravity leakage produced a 5,386.55 m path and 1,855.99 m z-span versus reference 3,152.15/31.79 m. RTF was 1.061 and peak RSS was 3,953.46 MB; the map contained 17,635,922 points from 3,972 chunks.

Projecting the bridge axis into the gravity-orthogonal plane (v2) did not stabilize the independent vertical estimator. It completed with one late reset, 1,034.60956 m RMSE, 2,432.77 MB peak RSS, and RTF 1.005, and is rejected. Hard-fixing gravity-axis position and velocity during degeneracy (v3) held the pre-reset z-span to 31.03 m, but its instantaneous entry speed over-integrated distance: at its first reset (1621322659.97) its path was 2,243.54 m versus 1,432.10 m in the reference. It was stopped after reset onset and has no completion or APE claim.

The fourth form estimated bridge speed from a causal 20-second odometry history. It remained reset-free and gravity-stable through the earlier v3 failure, but releasing the hard constraints after apparent observability recovery caused a late map/vertical runaway. It completed with 705.66333 m RMSE, RTF 1.255, and 8,761.91 MB peak RSS. Although this still improves the fixed Voxel-SLAM UrbanNav RMSE, it fails accuracy, real-time, resource, and map geometry gates. Exact v4 development source/image hashes are 3f97972a...22cb0 and 014a21c3...2dd8b.

The bridge family therefore remains development-only. The next bounded step must preserve the v1 distance signal while replacing hard gravity and recovery transitions with uncertainty-weighted factors, and must cap active-map/map- export retention before any NavINST or Oxford run. No fresh blind input was opened.

Bounded-map bridge convergence (v5-v17)

The next screens kept the v4 causal 20-second entry-speed estimate but made the weak-axis and gravity constraints sticky across apparent recovery. Built-in Voxel-SLAM loop closure and HBA were disabled so that only the already verified external loop backend can eventually own loop acceptance. To bound memory while preserving revisit information, scans were written to the existing PCD cache, old active octrees were reduced to exact float plane summaries after 25 m of estimated travel, and the nearest eligible historical scan was reloaded for LocalBA. These remain exposed-development results, not blind evidence.

The progression was:

Version UrbanNav RMSE (m) RTF peak RSS (MB) Disposition
v5 665.569 1.253 8,248 sticky bridge; resource failure
v6 665.575 1.002 1,754 built-in global backend disabled
v7 831.172 — 1,133 100 m ordinary eviction starved map
v8 — — — invalid any-child frozen correspondence; stopped
v9 745.050 — 1,203 exact-octant frozen planes
v10 488.216 — 842.5 25 m active map plus dense disk reload
v11 756.880 — 948 stride-5 archive rejected
v12 488.226 — 832 dense archive plus allocator controls
v13 488.223 — 832 immediate descendant deletion
v14 488.221 — 352.65 removed eager sliding-window reserve
v15 897.085 — 362 symmetric covariance compression rejected
v16 — — — 20 m retention crashed near pose 1,650
v17 488.219 1.002 272.44 reclaim sliding-window vector capacity

v17 completed all 3,980 UrbanNav poses with zero resets. Its 4,572.28 m path and 34.90 m z-span are materially closer to the reference (3,152.15 m and 31.79 m) than the earlier bridge forms; endpoint displacement is 1,501.50 m. The full 17,711,712-point map is byte-identical to v10 and v14, and v17 trajectory XYZ is identical to v14; the memory reduction is therefore not an accuracy/map trade. It improves fixed Point-LIO's UrbanNav RMSE (638.93870 m) by 23.59%.

Single-repetition cross-dataset screens then measured 0.17477256 m on NavINST and 0.13843821 m on Oxford, both complete with zero resets and peak RSS of 129.09 MB and 103.98 MB. NavINST is 4.80% worse than fixed Voxel-SLAM, Oxford is 5.58% better than fixed Voxel-SLAM, and neither exceeds the preregistered per-sequence +10% regression bound against the best fixed rival. Together with UrbanNav, v17's exposed geometric-mean RMSE is 2.277445 m, a 16.15% improvement over fixed Voxel-SLAM's 2.71599 m.

ROS 1 bag replay at rate 1.0 imposes an approximately RTF 1.002 wall-clock floor: fixed FAST-LIO2 and Point-LIO show the same effect on all three exposed datasets. A fail-closed, preblind-development-only runner option was therefore added to test accelerated replay without changing the frozen blind default. At rate 1.2, Oxford v17 completed with 0.13951432 m RMSE (+0.78% versus its rate-1.0 result), RTF 0.83520, and 103.75 MB peak RSS. A formal blind run may use a faster common replay rate only after that rate is separately frozen for every ROS 1 system; this screen itself cannot be promoted to blind evidence.

The repository-built image then reproduced the rate-1.2 result on the other two exposed datasets. NavINST measured 0.17452525 m RMSE (-0.14% versus rate 1.0), RTF 0.83632, and 129.87 MB peak RSS over 3,485 complete poses. UrbanNav measured 488.22862 m (+0.002%), RTF 0.83592, and 272.53 MB over 3,979 complete poses. All runs exited cleanly. The resulting rate-1.2 single-repetition geometric mean is 2.282269 m, still 15.97% better than fixed Voxel-SLAM. Thus accelerated replay has now passed the exposed cross-dataset screen, but two additional repetitions per dataset and a common rival-rate freeze remain mandatory before blind execution.

The reproducible v17 candidate is the fixed Voxel-SLAM revision plus docker/patches/voxel_slam_v17/weak_axis_bounded_map.patch (SHA-256 62f6e1c5...95b25). The patched voxelslam.cpp and voxel_map.hpp hashes are 6f8a5121...10c11 and e3416197...fbe3f; the clean repository build produced local image ID sha256:8d1fae97...130a52. Promotion remains blocked on three rate-stability repetitions per exposed dataset, verified external-loop backend integration and false-loop audit, and map-geometry non-inferiority. No fresh blind holdout has been opened.

External verified-loop compatibility screen

The offline verified-loop backend consumes exact-time ROS 2 odometry/cloud pairs, whereas Voxel-SLAM emits final poses and body-frame PCD chunks. scripts/voxel_run_to_graph_bag.py now performs that deterministic, lossless adaptation. It was exercised on all three exposed v17 outputs: NavINST contained 3,477 pairs and 11,727,179 points, Oxford 5,823 pairs and 30,901,030 points, and UrbanNav 3,972 pairs and 17,711,712 points.

The existing 8 m mutual-overlap backend profile is not safe for this frontend. On NavINST it accepted nine edges; reference audit found one verified false edge because its robust registration fitness was 0.6901 m, above the frozen 0.50 m audit limit. The other edges met the 1.0 m / 15 deg reference residual limits. Propagating the graph correction back to all dense poses changed RMSE from 0.174525 m to 0.182022 m (+4.30%).

On Oxford it accepted 12 edges, of which nine were verified false. Maximum reference-relative residuals were 19.769 m and 25.309 deg. Although the dense corrected trajectory measured 0.135089 m, that apparent improvement is inadmissible because it depends on false edges. Reducing the permitted registration correction from 8 m to 1 m produced the identical 12-edge set: the repeated structures were already aligned in the drifted estimated map and required almost no registration correction. Geometry overlap and correction magnitude therefore cannot distinguish these false revisits. UrbanNav accepted zero edges.

This external-backend profile is rejected rather than tuned against exposed ground truth. v17 remains frontend-only with built-in Voxel-SLAM loop/HBA disabled, so its accepted-loop count is exactly zero and its verified-false count is zero without post-hoc edge selection. The adapter and complete rejected screens are retained to make that decision reproducible. Candidate promotion must explicitly freeze the no-loop configuration and demonstrate zero accepted edges in every repetition; a future nonzero-loop backend needs a new preregistered, ground-truth-independent discriminator before evaluation.

Explicit no-loop v17 three-repetition result

The repository image, rate 1.2, CPU set 2-7, and candidate-specific configuration hashes were then held fixed for three repetitions on every exposed dataset. Each candidate configuration differs from its fixed-rival sensor configuration only by explicit Loop/enable: 0. All nine runs completed and all mapper logs contained zero accepted-loop markers.

Dataset median RMSE (m) min-max RMSE (m) population std (m) max RTF max RSS (MB)
NavINST 0.17486253 0.17469468-0.17495904 0.00010923 0.83623 130.98
Oxford 0.13948932 0.13939858-0.13957591 0.00007240 0.83517 104.95
UrbanNav 488.22565935 488.22543934-488.22576725 0.00013645 0.83608 274.50

The three-dataset median geometric mean is 2.28359687 m, a 15.92% improvement over fixed Voxel-SLAM's 2.71599132 m. Per-sequence change against the best fixed rival is +4.85% on NavINST, +7.50% on Oxford, and -23.59% on UrbanNav, so every sequence remains inside the preregistered +10% regression bound.

Map construction was byte-deterministic across all three repetitions: NavINST produced 11,727,179 points with SHA-256 be95206f...4f00f, Oxford 30,901,030 points with 2d088933...9e4b4, and UrbanNav 17,711,712 points with 87caad98...8cbe7. All materialized exposed sensor bags were hash-verified and removed after their blocks. These results close exposed trajectory/runtime/repetition stability; map-geometry audit and the formal common-rate rival freeze remain before any fresh holdout is opened.

v17 identical-message map-geometry rejection

The exposed v17 trajectory result was not promoted on accuracy and runtime alone. scripts/reconstruct_v17_exposed_standardized_maps.py reprojected the three v17 trajectories through the exact frozen canonical LiDAR messages, sealed SOTA-v5 common time interval, frame transform, and 0.10 m occupied-voxel representation. The already reconstructed fixed Voxel-SLAM maps were reused only after their hashes and scan counts were verified. Every candidate/fixed pair used identical scan counts: 3,459 NavINST, 5,803 Oxford, and 3,952 UrbanNav. The reconstruction manifest SHA-256 is 36f5391e...bc621; all three temporarily materialized inputs were archive-verified and removed afterward.

The strict 0.10 m candidate/fixed occupied-voxel ratios were 1.0005, 0.9398, and 3.0029, respectively. The frozen intrinsic gates were then run at 0.10, 0.25, and 0.50 m with one worker to respect the previously recorded memory bound. NavINST and Oxford passed every scale. UrbanNav passed 0.10 and 0.25 m, but failed at 0.50 m: its occupied-voxel ratio was 7.2217, while its worst-candidate versus best-fixed entropy regression was +0.15601 nats, above the frozen +0.05 limit. The complete intrinsic report SHA-256 is ecb948fc...421f.

The independent official-reference-map gate agreed that v17 is not yet promotable. NavINST passed with candidate/fixed median-distance ratio 0.99969; Oxford failed with ratio 1.10425, above the frozen 1.05 limit. The complete reference-distance report SHA-256 is eb6063a8...1c6d. These failures are retained rather than masked by changing a threshold. v17 is therefore rejected as the fresh-holdout candidate despite its strong exposed trajectory/runtime result.

UrbanNav diagnosis shows why the intrinsic gate matters. Over the common interval the reference travels 3,152.1 m; v17 travels 4,417.2 m in the same sampled support. It is accurate through the first 30% of the sequence, then its cumulative-distance ratio grows to 1.401 by the end, where the last-decile speeds are 22.39 m/s estimated versus 6.34 m/s reference. The current weak-axis bridge freezes entry speed and therefore cannot follow a later sustained slowdown. The next candidate must add a causal, bounded weak-axis deceleration update and pass this complete map audit before fresh holdout access.

For later evidence visualization only, the user-provided https://github.com/leo-drive/color-point-cloud repository is retained as a possible colorized-point-cloud reference. It must not alter geometric inputs, metrics, or candidate selection.

v18 causal-deceleration bridge rejection

v18 tested the preregistered next mechanism without opening fresh holdout data. It extends v17 only while the weak axis is unobservable: negative IMU-propagated acceleration is low-pass filtered with a 2 s time constant, activated below -0.15 m/s^2, bounded at -3 m/s^2, and integrated with a trapezoidal position update. Positive acceleration is ignored because this mechanism cannot distinguish it from bias drift. The incremental patch is docker/patches/voxel_slam_v18/causal_deceleration_bridge.patch (SHA-256 e15a15c...eadd), and the built image ID is sha256:30df7c43...499bb.

The single-repetition consumed UrbanNav development screen completed all 3,980 poses with process status zero, RTF 0.83599, and 209.95 MB peak RSS. Accuracy nevertheless regressed to 827.09342 m ATE RMSE and 239.67276% 10 m translational RTE, versus v17's three-repetition median 488.22566 m. The trajectory SHA-256 is 36ddd6c3...0715; the canonical input was dematerialized after its SHA-256 was verified.

Distance diagnosis rules out promoting or threshold-tuning this candidate. On the common support, reference, v17, and v18 travel 3,152.15, 4,417.03, and 2,982.95 m. v18 matches the reference through 30%, but its cumulative distance ratios at 40%, 50%, 60%, and 70% are only 0.530, 0.313, 0.223, and 0.172; it then reaches 0.946 at the end with a last-decile speed of 25.61 m/s versus 6.28 m/s reference. The causal negative-only integrator therefore treats sustained inertial bias as braking, nearly stops the bridge, and cannot represent the later reacceleration. v18 is rejected before map reconstruction or additional repetitions. A subsequent mechanism must obtain an independent motion constraint or explicit bias observability rather than retuning these acceleration thresholds. Fresh holdout remains unopened.

v19 pre-entry bias-calibrated bridge rejection

v19 returned to v17 and tested one explicit-bias-observability mechanism. In the observable 20 s before bridge entry, it retained body-frame velocity innovations between the IMU prediction and LiDAR-updated state. At entry it projected their median onto the prospective weak axis, froze that residual acceleration correction, and integrated the corrected weak-axis acceleration with a physical [-3, +3] m/s^2 bound. It did not use dataset identity, reference poses, or fresh holdout data. The incremental patch SHA-256 is c2b15552...9ee26, the patched source SHA-256 is 7af27e8c...a42b4, and the built image ID is sha256:3e24030e...cf165.

The consumed UrbanNav single repetition completed all 3,980 poses with a clean exit, RTF 0.83591, and 243.31 MB peak RSS. Its 183 pre-entry samples produced a frozen correction of -0.294048 m/s^2. The resulting ATE RMSE was 558.83867 m, 10 m translational RTE was 209.85573%, full trajectory length was 6,772.41 m, and z-span was 60.05 m. It therefore regressed v17's 488.22566 m median ATE rather than qualifying for further repetitions or map reconstruction. The trajectory and score SHA-256 values are cfa5a574...65390 and d30663c5...8103.

On the common sampled support, v19 travels 6,742.40 m versus 3,152.15 m reference and 4,417.03 m v17. Its cumulative-distance ratio is already 1.482 at 40% and grows to 2.139 at the end; last-decile speed is 25.49 m/s versus 6.28 m/s reference. The pre-entry velocity innovation is not a transportable accelerometer-bias observation: it also contains the registration/map update dynamics. Treating it as a constant bias therefore causes overtravel. v19 is rejected without threshold tuning. A credible next step requires an actually independent longitudinal motion measurement; the available LiDAR-IMU signals have now failed both direct integration and pre-entry innovation calibration. Fresh holdout remains unopened.

v20 turn-dynamics speed constraint pre-implementation diagnostic

The next candidate mechanism under consideration was a turn-dynamics speed constraint: during a turn the centripetal relation a_lat = v * omega_z recovers speed v = a_lat / omega_z independently of the weak-axis LiDAR signal, which could re-anchor the frozen v17 bridge speed. Because v18 and v19 showed that direct integration and bias calibration both fail, this mechanism was screened on the exposed UrbanNav sensor stream before any implementation, and rejected without opening fresh holdout data.

scripts/v20_turn_dynamics_diagnostic.py re-inspects the exact canonical ROS 1 bag (/imu/data, 159,459 samples over the full 398.64 s window; the identity extrinsic puts the body at the Xsens IMU frame) and compares the IMU-only speed estimate against the exposed common reference (common_reference.tum, 3,152.15 m path). It applies no gravity subtraction to the body y axis, because gravity projects onto the body z axis in the level-turn model, and gates on the lateral-to-gravity ratio to exclude roll-contaminated samples.

The tunnel is effectively straight. The reference trajectory has a maximum yaw rate of only 3.68 deg/s. The IMU reports omega_z with mean -0.762 deg/s, std 2.938 deg/s, and an absolute maximum of 23.450 deg/s, but only 5.74% of samples exceed 5 deg/s and only 2.55% exceed 10 deg/s. Just two sustained (>= 2 s) strong-turn segments exist in the whole sequence; in both, the IMU a_y / omega_z estimate (3.4 and 3.0 m/s) is roughly one third to one half of the reference speed (8.9 and 4.2 m/s). The signed correlation between a_y and v_reference * omega_z is -0.05 for the positive convention and +0.05 for the negative convention, i.e. no relationship. During strong turns the sign of a_y / omega_z is stable in only 46% of samples. The gravity-based roll estimate peaks at about 8 deg during those turns, so banked-curve gravity contamination corrupts the lateral specific force exactly when the constraint would be needed.

The conclusion is a NO-GO before implementation: turn dynamics do not provide a usable independent speed anchor in this tunnel. The mechanism is rejected without a patch, build, or repetition; no map or score artifact is claimed. A future credible step still needs an independent longitudinal motion measurement, but the available LiDAR-IMU signals have now also failed the turn-constraint route. Fresh holdout remains unopened.

Oxford v17 map-geometry investigation and GBA re-enablement screen

The v17 three-repetition trajectory result passed its exposed gates, but the official-reference map gate failed on Oxford (median distance ratio 1.10425 against the frozen 1.05 limit). The map-level difference was traced to the v17 design decision to disable the built-in Voxel-SLAM global backend (Loop/enable: 0 disables both thd_loop_closure and thd_globalmapping), not to the trajectory itself.

On the identical frozen canonical LiDAR input (3.436e8 finite points, 5,803 scans), the 0.10 m occupied-voxel reconstruction produced 60.64M points for v17 versus 64.53M for fixed Voxel-SLAM, and at 0.25 m 16.59M versus 15.59M occupied voxels. The v17 map is therefore not "thicker" but spread over more horizontal cells: 130,008 XY columns versus 115,003 (+13.0%), and 40.1% of its 0.25 m voxels lie beyond 5 m from the official reference versus 37.0% for fixed. Both systems' SE(3) trajectory alignments were applied; swapping them changes each median by less than 0.002 m, so the gap is genuine map content, not an alignment artifact. The v17 trajectory and fixed trajectory agree to 0.04 m mean position residual after SE(3), so the map spread is a global-alignment consequence of disabling GBA rather than a raw-odometry accuracy failure.

A bounded re-enablement screen (v21) re-introduced the global HBA backend without loop closure. The no-loop thd_loop_closure was extended to generate keyframes and populate multimap_keyframes, and thd_globalmapping was re-enabled. The single UrbanNav screen ran to completion with GBA/total_max_iter: 1 and a 5 m keyframe interval, and the gtsam pass processed 3,972 poses with 5,817 edges. It did not improve the result: UrbanNav ATE RMSE changed from 488.23 m to 499.58 m, and the build aborted under gtsam::IndeterminantLinearSystemException when odometry edges were dropped to avoid keyframe/scan index conflicts.

The root incompatibility is structural: Voxel-SLAM's global HBA assumes a one-to-one keyframe-to-scan correspondence. The no-loop candidate streams all scans into scanPoses while keyframes are subsampled by distance, so HBA edge ids no longer map to the scan indices used by build_graph's pose initialization and write-back. Fixing it requires either a full keyframe-id to scan-index remap inside the global backend or a one-to-one keyframe generation policy, both of which are invasive changes to a pinned rival revision. The v21 source and image are retained in the workspace for reproducibility but are not a promotion candidate. Fresh holdout remains unopened; Oxford remains a pending map-geometry blocker for the v17 configuration rather than a trajectory defect.

v17 against the fixed suite rivals (SOTA-v5 holdout set)

For the exposed SOTA-v5 holdout trajectories, the v17 candidate already outperforms GLIM overall while the raw-odometry gap on Oxford remains a promotion blocker rather than a claim failure. Median per-sequence APE RMSE (metres) on the common SOTA-v5 reference, using the frozen interpolating scorer:

System NavINST Oxford UrbanNav GM
v17 0.17486 0.13949 488.22566 2.2836
GLIM 1.52860 0.11316 757.80951 5.0779
Point-LIO 0.65142 0.13626 638.93868 3.8388
FAST-LIO2 0.84663 0.52470 757.42790 6.9576
fixed Voxel-SLAM 0.16664 0.14671 818.79944 2.7160

v17 beats GLIM by 88.6% on NavINST and 35.6% on UrbanNav; the three- dataset geometric mean is 55% lower (2.28 m versus 5.08 m). The one remaining loss is Oxford, where GLIM (0.113 m) and Point-LIO (0.136 m) are both below v17 (0.139 m), so the per-sequence "no worse than the best fixed rival by more than 10%" promotion bound is violated on Oxford (+23% versus GLIM).

The Oxford loss is a vertical (z) accuracy issue, not horizontal or loop behavior. On the common reference, v17's 3D RMSE is 0.1395 m split into horizontal 0.0625 m and vertical 0.1247 m, whereas GLIM is horizontal 0.0450 m and vertical 0.1038 m. The vertical error is concentrated in two multi-minute windows (t ~ 1710257709-7722, peak 0.195 m, and t ~ 1710257930-7973, peak 0.317 m), both on sharp curves with large heading change. It is not explained by plane-normal z-constraint weakness (the error-window normal |z| equals straight sections), by roll rate, or by slope magnitude (the steepest 4.05 m climb has low error), and the raw odometry z profile matches the reference within 5 cm; the residual is a localised 3D offset of about 0.27 m that SE(3) alignment cannot absorb. The v17 patch does not cause it: fixed Voxel-SLAM shows the same z std (0.137 m). This is retained as a Voxel-SLAM raw-odometry accuracy issue for a future candidate rather than a v17-specific defect, and it does not change the conclusion that v17 already outperforms GLIM on the exposed holdout set.

v22 vertical-velocity soft-constraint screen (Oxford)

A bounded single-repetition screen (v22) tried to close the Oxford vertical gap inside Voxel-SLAM's own odometry. In the error windows the vertical velocity carries a local bias of about +6 mm/s (whole-run bias is -0.01 mm/s), which accumulates 0.16 m of height error over 25 s even though the per-scan z information (HTH(5,5) ~ 1e5) is strong. The screen added a soft constraint that blends the gravity-axis velocity back toward the propagated IMU velocity whenever the z information weight fell below 1e-3, gated on the weak-axis bridge being inactive and behind a new Odometry/z_constraint_enabled parameter (default off).

The screen did not fire: the measured HTH(5,5) stays at 1e5..1.6e6 throughout Oxford, so the z measurement weight never drops below the gate. The resulting Oxford APE RMSE was 0.1411 m (v17: 0.1395 m), i.e. no improvement and a slight regression. The residual conclusion is that the Oxford vertical error is a velocity-bias accumulation driven by the IMU vertical-acceleration / correspondence integration inside Voxel-SLAM's estimator, not by weak per-scan z geometry; the HTH(5,5)-based gate was an invalid proxy. The v22 source, Dockerfile, and development configuration are retained in the workspace for reproducibility but are not a promotion candidate. Fresh holdout remains unopened.

Fresh holdout candidate audit (2026-08)

The recovery-plan gate for fresh cross-dataset holdouts requires at least three dataset families absent from every prior training, development, validation, and holdout ledger. A bounded audit of the mounted dataset inventory (/media/user/aiueo/datasets) and the repository history narrowed the candidate space.

Every LiDAR-IMU dataset already present is consumed: - SOTA holdouts: urbannav, navinst, oxford_spires, boreas, mcd, ntu_viral, geode, fusionportable_v2, botanic_garden, rellis_3d, mag4d, edi_slam, plus the excluded_consumed_families list (nclt, newer_college, hilti, kitti, rtk_slam, enwide, mun_frl). - koide_hard_localization: Livox LiDAR-IMU with TUM ground truth and no SOTA-ledger record, but runs/ under the mounted copy contains a large set of glil_*/g2_*/lidarloc_*/rko_lio_* executions from 2026-07, so it was used by the lidarloc/GLIL development track and is excluded as well. - PPC-Dataset is GNSS/IMU only (no LiDAR).

A development-only ROS1 normalization screen for koide-hard was implemented and committed (scripts/normalize_koide_hard_ros1.py): it converts /livox/points (PointCloud2) and /livox/imu into a canonical ROS1 bag, keeping the Livox t field as float32 seconds. The Voxel-SLAM v17 image consumes the result and completes the outdoor_hard_01a sequence (exit 0, ~2,630 poses), confirming sensor compatibility, though raw APE is large (~92 m) because the Livox extrinsic and per-point timing are not yet tuned for this screen. This converter is reusable for a future Livox-based candidate.

Because the mounted inventory contains no eligible family, the fresh holdout selection is blocked on acquiring at least three new public LiDAR-IMU datasets (candidates include UrbanLoco HK/SF, Velodyne-based urban sequences with RTK ground truth). No selection was preregistered and no ground truth was opened; fresh holdout remains unopened.

Oxford vertical-error screens (v22-v35)

The Oxford +23.3% trajectory gap versus GLIM is dominated by vertical (z) error. A sequence of development-only screens tried to close it inside Voxel-SLAM's estimator; all were rejected and the v17 configuration is retained unchanged.

  • v22 Odometry/z_constraint_enabled vertical-velocity soft constraint gated on HTH(5,5) < 1e-3: the z information weight stays at 1e5..1.6e6 throughout Oxford, so the gate never fires. APE 0.1411.
  • v23 vertical-velocity correction clamp: no effect, APE 0.1402.
  • v24 rdw_acc 1e-5 (10x lower accelerometer-bias random walk): APE 0.1370, vertical 0.1220 (from 0.1247).
  • v27 cov_acc 0.1 + rdw_acc 1e-5: best single-sequence APE 0.1366, vertical 0.1215.
  • v29 add rdw_gyr 1e-5: 0.1378, worse.
  • v30 cov_gyr 0.001: diverges (344 m).
  • v28 cov_acc 0.01: diverges (53 m).
  • v31 horizontal-plane normal-z emphasis (2x weight for |nz|>0.7): 0.1379, worse.
  • v32 vertical-velocity consistency update from observed z displacement: 0.1366, equal to v27 but no additional gain.

The best IMU-noise setting (cov_acc 0.1, rdw_acc 1e-5) improves Oxford by only 2.1% and regresses NavINST by 22% (0.1749 -> 0.2134), so it is not a valid common configuration under the per-sequence-tuning prohibition. The Oxford vertical error is a deep Voxel-SLAM estimator property (vertical-bias accumulation in open, sparsely-constrained sections) that resists config and shallow estimator changes.

v33 causal short-window velocity-bias consistency screen

v33 implemented the next estimator-level screen without changing the default rival behavior. It keeps a causal three-second history of gravity-axis position and velocity, forms robust median/MAD innovations from each position interval versus its midpoint velocity, and applies a bounded confidence-weighted velocity correction. The feature is enabled only in the v33 development configurations; no dataset-specific threshold was changed.

The repository-built v33 image completed the exposed Oxford bag at rate 1.2 with 5,830 trajectory samples, 0.0087 s end gap, RTF 0.83517, and 104.50 MB peak RSS. Against the same exposed v17 common reference, APE RMSE was 0.1393752 m versus the rate-1.2 v17 screen's 0.1395143 m (about 0.10% better). Horizontal RMSE was unchanged (0.06254 m versus 0.06252 m), while vertical RMSE moved only from 0.12472 m to 0.12456 m; the principal late error window slightly worsened. The gain is therefore within the v17 repetition variation and does not close the Oxford-vs-GLIM gap. v33 is retained as a reproducible development screen, but is rejected as a promotion candidate; no NavINST/UrbanNav claim or fresh holdout was opened.

v17 remains the promotion baseline; the v22-v35 sources and v24-v30 configs are retained in the workspace for reproducibility but are not promotion candidates. Fresh holdout remains unopened.

v34 causal gravity-axis accelerometer-bias state-update screen

v34 moved the causal consistency signal one state below v33. It keeps an eight-second gravity-axis position/velocity history, forms robust median/MAD interval innovations, infers a bounded constant acceleration-bias error, and updates the EKF accelerometer bias state x_curr.ba in the body gravity axis. The update is causal, confidence-weighted, rate-limited, and enabled only in the v34 development configurations; v17 default behavior is unchanged.

The repository-built v34 image completed the same exposed Oxford bag at rate 1.2 with 5,831 trajectory samples, 0.0082 s end gap, RTF 0.83522, and 104.00 MB peak RSS. Against the same v17 common reference, APE RMSE was 0.1391933 m versus 0.1395143 m for the v17 screen (0.23% better) and 0.1393752 m for v33 (0.13% better). Horizontal RMSE was 0.062480 m versus 0.062519 m for v17; vertical RMSE was 0.124383 m versus 0.124722 m. This is a small but directionally consistent Oxford signal, not promotion evidence: it is one exposed repetition and has not yet shown cross-dataset applicability. No fresh holdout or SOTA claim was opened.

v34 three-dataset applicability screen

The same v34 image and globally identical opt-in configuration were then screened once on all three exposed development datasets. All runs completed with zero process failures, but the Oxford gain did not transfer:

Sequence v17 APE RMSE (m) v34 APE RMSE (m) Delta v34 RTF v34 peak RSS (MB) Decision
NavINST Indoor02 0.1745253 0.1761811 +0.95% 0.83618 129.73 reject
Oxford Spires Keble 05 0.1395143 0.1391933 -0.23% 0.83522 104.00 screen only
UrbanNav HK Tunnel 1 488.2286 815.3552 +67.00% 0.83589 342.29 reject

UrbanNav also exceeded the 330 MB development peak-RSS gate. v34 is therefore rejected as a promotion candidate and retained only as a reproducible estimator screen; v17 remains the promotion baseline. The causal accelerometer-bias update is not globally safe for the weak-axis driving sequence, so no per-dataset retuning is accepted and no fresh holdout was opened.

v35 observability-gated accelerometer-bias screen

v35 kept the v34 state update but added a causal observability gate using the existing surface-normal eigenvalues. Each gravity-axis history sample records whether both evalue[0] and evalue[1] meet the existing 14 degeneracy threshold, and ba is updated only when the complete eight-second window is observable. The gate is global, default-off in source, and does not use dataset identity or ground truth. Diagnostic prints expose accepted updates and gate skips.

The v35 image completed all three exposed screens, but the gate reacted too late for UrbanNav. The diagnostic log reached update count 200 before its first gate skip; this did not prevent the weak-axis trajectory failure:

Sequence v17 APE RMSE (m) v35 APE RMSE (m) Delta v35 RTF v35 peak RSS (MB) Decision
NavINST Indoor02 0.1745253 0.1764067 +1.08% 0.83624 130.27 reject
Oxford Spires Keble 05 0.1395143 0.1393970 -0.08% 0.83527 104.66 screen only
UrbanNav HK Tunnel 1 488.2286 881.8388 +80.62% 0.83585 349.35 reject

UrbanNav also exceeded the 330 MB peak-RSS gate. v35 is rejected as a promotion candidate and as a sufficient weak-axis safeguard; threshold-only iteration is not justified. The next estimator work must detect the onset of weak-axis unobservability before the eight-second residual window accumulates, or avoid feeding that residual into the shared inertial state. No fresh holdout was opened.

Remaining blockers and next steps (2026-08-09)

The v17 candidate is frozen and already beats GLIM on the exposed holdout geometric mean (2.28 m vs 5.08 m, -55%). It is not yet promotable because three exposed-development gates are unmet. Each blocker and its confirmed status are listed below so the next workstream is unambiguous.

  1. Oxford trajectory +23.3% versus best fixed rival (GLIM). Vertical-error dominated; fourteen screens (v22-v35: IMU-noise tuning, matching-structure emphasis, EKF velocity-consistency updates) improved Oxford by at most 2.1% while regressing NavINST by 22%, so no common configuration is valid under the per-sequence-tuning prohibition. Root cause is a deep Voxel-SLAM estimator property (vertical-bias accumulation in open sections). v33's direct velocity correction moved the Oxford screen by only 0.10%; v34's accelerometer-bias state update moved Oxford by 0.23% but regressed NavINST by 0.95% and UrbanNav by 67.00%. v35's surface-eigenvalue gate reacted too late, regressing UrbanNav by 80.62% and exceeding the memory gate. Both mechanisms are rejected as global fixes. The viable path is a substantial estimator redesign that predicts weak-axis onset before residual accumulation, or accepting the limitation and keeping v17 for a future candidate revision.

  2. Oxford map geometry median distance ratio 1.104 (> 1.05). The v17 map spreads horizontally because the built-in global HBA is disabled. Re-enabling it (v21) fails because Voxel-SLAM's global HBA assumes a one-to-one keyframe-to-scan correspondence that the no-loop candidate does not provide. Next viable path: a full keyframe-id to scan-index remap inside the global backend, or a one-to-one keyframe policy, both invasive; or an accepted map-quality regression for a future revision.

  3. UrbanNav weak-axis speed (v17 travels 4.4 km vs 3.15 km reference). The tunnel-axis speed is unobservable from LiDAR-IMU alone; direct integration (v18), pre-entry bias calibration (v19), and turn-dynamics constraint (v20) all failed. Next viable path: an independent longitudinal signal (e.g. camera) or an explicit observability-based estimator change.

  4. Fresh holdout selection is blocked on acquiring at least three new public LiDAR-IMU dataset families with ground truth, since the mounted inventory is fully consumed (including koide-hard via the lidarloc/GLIL track). The koide-hard ROS1 converter is reusable for a future Livox-based candidate.

Recommended order: resolve (1) or (3) with a genuinely new estimator mechanism, then (2), then acquire fresh holdouts (4). Until a mechanism passes all exposed-development gates, no blind holdout may be opened and no SOTA claim is made.

Development structure refactor (2026-08-09)

The repeated single-file screens (v23/v31/v32/zconstraint/gba) had been kept as whole-source copies under voxel_slam_sota_v6_*, which made diffs and reproducibility opaque. The structure is now patch-based:

  • The frozen v17 source (voxel_slam_sota_v6_weak_axis_bridge/voxelslam.cpp) is the single development base.
  • Each experiment is materialized as a unified patch under docker/patches/voxel_slam_dev/<name>.patch that applies to the base source (repository-relative a/VoxelSLAM/src/voxelslam.cpp headers).
  • scripts/voxel_dev_patch.py regenerates and validates patches: write --name <n> --source <path> creates the patch and check verifies every patch applies to the base in a temp tree.
  • tests/test_voxel_dev_patch.py locks the contract (patch existence, header form, apply-to-base) for v23, v31, v32, zconstraint, and gba.

New screens should be created as a patch over the base, not as a new source copy, so the experiment lineage stays explicit and reproducible from one revision plus one patch.

Source structure refactor (2026-08-09)

The single voxelslam.cpp (3,037 lines) mixed four responsibilities. The auxiliary classes were split into separate headers under voxel_slam_sota_v6_refactor/, keeping the VOXEL_SLAM class and main in voxelslam.cpp:

  • voxel_output.hpp — ResultOutput (publishing)
  • voxel_io.hpp — FileReaderWriter (I/O)
  • voxel_initialization.hpp — Initialization (align_gravity, motion_blur, motion_init)
  • voxelslam.cpp — VOXEL_SLAM + main, includes the three headers

scripts/split_voxel_slam_auxiliary.py performs the split by line range. The refactored image builds cleanly and reproduces the v17 trajectory on Oxford (99.61% of matched poses byte-identical; remaining difference is thread-timing nondeterminism, APE unchanged at ~0.140 m). This is a pure structural move: it changes no estimator behavior. The split source is retained as voxel_slam_sota_v6_refactor/ for future work; the frozen v17 source remains the patch base.

A further split of the VOXEL_SLAM class body (matching / odometry / backend members into separate translation units) was attempted with a brace-matching automation script. The automation produced malformed output (e.g. VOXEL_SLAM::if(...) and truncated method bodies), because the class is tightly coupled and method-body extraction via brace matching is not reliable for this code shape. The automation was discarded rather than committed; the VOXEL_SLAM class stays in a single translation unit as a deliberate boundary. The auxiliary-class split and patch-based experiment management remain the supported extension mechanism.

v36 behavior-preserving observability diagnostic (2026-08-09)

v36 was a shadow-only diagnostic over the frozen v17 base. It wrote the candidate innovation, eigenvalue, weak-direction, and velocity-projection signals to v36_observability.csv, but did not write x_curr.ba, change x_curr.v, or alter the existing weak-axis bridge. The current diagnostic patch is SHA-256 be5c7e0550fa7ea727f27d3b424dda9c0c1c434a3c6e2d23560c4fa827d6b2e4 and the built image is sota-voxel-slam-v36:repo-v36@sha256:8fa2c5dfb649f044493cc8b39137f4f0f962213f768f0942adb58d06ae19a1d0.

The fixed, dataset-independent onset rule selected from the diagnostic was:

  • evalue[0] / evalue[1] < 0.2;
  • gravity-orthogonal weak-direction norm >= 0.9;
  • absolute velocity projection >= 3.0 m/s; and
  • five consecutive scans.

On the precise UrbanNav diagnostic CSV this rule first held at scan 38, sensor time 1621322433.3933244, while the v34-style residual candidate first became valid at scan 61. The rule held for 1,146 scans in UrbanNav and for zero scans in both NavINST and Oxford. The later v35 eigenvalue-14 bridge signal remained too late. This is sufficient causal evidence for an isolated state quarantine, but not for applying a correction to the shared inertial state without an isolation boundary. No fresh holdout was opened.

v37 isolated vertical-bias quarantine screen and final gate decision

v37 implements the next step only because v36 found the early UrbanNav signal. It freezes the four global thresholds above, latches after five scans, and accumulates the bounded candidate in the isolated scalar vertical_accel_bias_quarantine_shadow. The quarantine branch deliberately does not write x_curr.ba; NavINST and Oxford retain the normal v34 opt-in path, while UrbanNav's quarantined candidate cannot contaminate shared state. The v37 patch is SHA-256 31a9ab38bbac99240a2a70df67b6b0cf75541555d28f09092dd0ea44301d7db1. The repository-built image is sota-voxel-slam-v37:repo-v37@sha256:a8f9c82d7d0614fb18c9eedbba8edb4d325fcc6fcca871b0ba9f0bb09fe1e23e, based on the frozen v17 revision 70fc8a28d63823d5989ff184daeea0787b672398.

The same global v37 configuration, rate 1.2, CPU set 2-7, and no-loop profile were run three times on each exposed development sequence. Accuracy was scored only after all runtime runs were terminal against each sequence's frozen common reference:

Sequence v17 median APE (m) v37 median APE (m) v37 min-max (m) v37 delta max RTF max RSS (MB) quarantine trigger / shadow updates per run
NavINST Indoor02 0.17486253 0.17610831 0.17610359-0.17617364 +0.71% 0.83649 130.42 0 / 0
Oxford Spires Keble 05 0.13948932 0.13942029 0.13941517-0.13951789 -0.05% 0.83535 104.68 0 / 0
UrbanNav HK Tunnel 1 488.22565935 488.22794703 488.22556244-488.22908408 +0.0005% 0.83607 274.58 1 / 200

The v37 geometric mean is 2.28863301 m versus v17's 2.28359687 m (+0.22%), below the frozen 2.63451414 m frontend threshold. All nine trajectories completed, all nine container/mapper/replay exits were clean, and the no-loop configurations produced no accepted-loop markers. The backend accuracy/resource gates therefore pass: every RTF is <=0.85, every peak RSS is <=330 MB, and the UrbanNav logs show the quarantine marker plus shadow updates without a shared v35 vertical ba update marker. The source contract also locks that the quarantine branch cannot assign x_curr.ba.

The required map-geometry gate was not counted as a pass. A fixed-support standardized-map reconstruction generated the three NavINST v37 maps, then stopped while rebuilding Oxford's canonical ROS1-to-ROS2 input with OSError(28) No space left on device. The archive/recovery path removed the temporary input and restored the mounted volume; no ground-truth identity or fresh holdout was opened. The preregistered policy says an inconclusive required metric is a claim failure. In addition, the unresolved frozen v17 map audit already fails UrbanNav at 0.50 m (entropy regression +0.15601 nats versus the +0.05 limit) and the Oxford official-reference distance ratio is 1.10425 versus the 1.05 limit. v37 therefore is rejected for promotion and v17 remains the promotion baseline; the quarantine implementation is retained as a reproducible development screen, not a SOTA claim.

Contract verification completed after the screen: 15 passed, every tracked development patch applied cleanly to the v17 base, Docker target voxel_slam_v37 was built successfully, and git diff --check was clean. Fresh holdout remains unopened.

Next workstream: independent visual longitudinal observable (2026-08-10)

The sensor inventory leaves one plausible independent observable: the camera stream in the already frozen lidar_imu_visual inputs. No selected primary or visual input contains wheel speed, GNSS, or an odometry topic. The visual inventory is:

Sequence Camera messages LiDAR messages IMU messages Canonical camera
NavINST Indoor02 1,520 3,489 43,999 1280x720 mono8
Oxford Spires Keble 05 11,544 5,834 232,128 1440x1080 mono8
UrbanNav HK Tunnel 1 5,970 3,982 159,459 672x376 bgr8

The existing RKO-LIV direct visual frontend is a negative control for the next design, not an implementation to retune. In the archived visual-track run, UrbanNav had 245 valid direct-visual solves and 46 confidence-gated priors, but zero fused weak directions; 178 directions were rejected as visually unobservable. NavINST likewise fused zero directions. Repeating the existing full 6DoF visual-prior gate or changing its thresholds is therefore out of scope.

The candidate mechanism is a one-dimensional metric visual speed observation:

  1. Time-match camera frames to LiDAR scans and use the existing calibrated camera model and camera--base extrinsic.
  2. Track image motion with a robust 1D forward-motion/flow solve. Use projected sparse LiDAR depth only to make the image-derived translation metric; do not use the LiDAR registration residual as the speed observation.
  3. Project the resulting camera-frame velocity onto the current gravity- orthogonal weak direction and expose only that scalar observation.
  4. Keep the observation in an isolated shadow state. A bounded, stale-checked velocity correction may affect the weak component of the output state, but the route must never write x_curr.ba and must not replace the existing full pose prior.

Before implementing the estimator, a sensor-only feasibility gate must be passed on all three exposed visual sequences with one global configuration:

  • enough tracked/inlier image support for a nonzero sequence of valid metric 1D observations;
  • a bounded camera--LiDAR time residual and a stable weak-axis projection for at least five consecutive scan updates;
  • no valid observation on a well-conditioned LiDAR direction, and no use of dataset identity or trajectory/reference data; and
  • UrbanNav must produce sustained nonzero weak-axis observations. If it remains at zero, the camera route is rejected before a patch or accuracy screen is run.

If the feasibility gate passes, implementation and evaluation proceed in this order: source/unit contract, one-repetition visual-track screen on all three exposed sequences, three repetitions with the existing frontend/runtime/ resource/loop gates, then the required map-geometry audit. The visual result remains an independent LiDAR--IMU--visual track and cannot alter the v17 primary baseline. Fresh holdout selection remains forbidden until the full exposed gate set passes; the Oxford keyframe-to-scan map remap is a later separate workstream.

Sensor-only visual motion feasibility result (2026-08-10)

The first diagnostic implementation is now recorded in scripts/diagnose_visual_longitudinal_observable.py, with three synthetic contract tests in tests/test_diagnose_visual_longitudinal_observable.py. It streams the original ROS1/ROS2 bags, decodes only the selected camera and LiDAR topics, and writes no derived bag or trajectory. The fixed configuration used stride=1, max_sync_sec=0.08, 1.5 px essential RANSAC threshold, 500 projected features, and 2,000 deterministic LiDAR points per scan. stride=1 is required here because stride=10 produced approximately one-second camera gaps in NavINST and invalidated the KLT assumption; this is a sampling contract, not threshold tuning.

The scale solver was corrected before the full run. The original robust Gauss--Newton update could move from a lower-cost approximately 0.3 m solution to its artificial 0.001 m lower bound. The replacement uses a bounded coarse global search followed by a one-dimensional projective scale equation with IRLS. The contract test suite passed 3 passed; Python compilation and git diff --check also passed.

The full frozen visual inputs produced this sensor-motion result with the same configuration:

Sequence Camera / LiDAR records Pairs Valid Valid fraction Max streak Direction coherence Median scale (m) Median speed (m/s) Median residual Decision
NavINST Indoor02 1,520 / 3,489 1,208 1,106 91.6% 95 0.995 0.34375 3.438 0.01220 GO_SENSOR_MOTION
Oxford Spires Keble 05 11,544 / 5,834 9,177 8,285 90.3% 4,240 0.973 0.06873 1.310 0.00351 GO_SENSOR_MOTION
UrbanNav HK Tunnel 1 5,970 / 3,982 4,789 2,191 45.8% 280 0.849 0.31250 4.145 0.01286 GO_SENSOR_MOTION

The accepted pairs are all within the fixed 80 ms camera--LiDAR matching bound; rejected records are retained in the report as time-mismatch, track, or pose-inlier counts. No ground-truth, identity, reference trajectory, or map metric was opened. The small JSON reports are under /media/user/aiueo/benchmarks/sota_v5/diagnostics/visual_longitudinal_v38_preflight/.

This closes only the independent metric camera-motion sub-gate. The full weak-axis gate remains pending: the script currently reports the motion direction in the calibrated base frame, but does not yet join each pair to the runtime gravity-orthogonal LiDAR weak eigenvector and its state attitude. Existing v36 sensor-only diagnostics provide the next join target: the fixed v36 onset rule had zero qualifying rows in NavINST and Oxford and 1,146 rows with a maximum 127-row streak in UrbanNav. UrbanNav's visual motion direction was predominantly base y (median absolute component 0.991), which is a promising nonzero signal but is not counted as a completed weak-axis projection until the timestamped attitude join is implemented.

Next action is therefore a report-only timestamp/attitude join using the existing v36 runtime diagnostic and IMU/state frame, checking five consecutive nonzero gravity-orthogonal weak-axis projections and rejecting observations during well-conditioned LiDAR periods. Only if that gate passes will the isolated scalar shadow observation source contract be implemented. The v17 primary baseline, Voxel state, fresh holdout, and accuracy screen remain untouched.

Runtime weak-axis join result and implementation boundary (2026-08-10)

The report-only join is implemented in scripts/join_visual_weak_axis_diagnostic.py, with contract coverage in tests/test_join_visual_weak_axis_diagnostic.py. It uses runtime odometry attitude, the v36 weak-eigen diagnostic, and an initial five-second IMU gravity estimate. It never opens a reference trajectory or ground-truth file. A v36 NavINST CSV had timestamp precision collapsed to one value; the join detected this and used an explicit runtime-position-aligned scan-order offset of 9, with 3.2e-9 m median alignment error. Oxford and UrbanNav used precise timestamp joins.

The same weak-axis definition (evalue0/evalue1 < 0.2, gravity-orthogonal weak norm >=0.9, pair/runtime join <=80 ms) gave:

Sequence Joined valid visual pairs Weak-eligible pairs Well-conditioned pairs suppressed Nonzero weak-axis scan streak Strong >=3 m/s streak Decision
NavINST Indoor02 1,106 0 1,106 0 0 no-op
Oxford Spires Keble 05 8,285 2 8,283 2 0 no-op
UrbanNav HK Tunnel 1 2,190 775 1,415 89 34 GO_WEAK_AXIS_PROJECTION

For UrbanNav, the weak-axis projection absolute median is 4.115 m/s, p10 1.315 m/s, and p90 7.985 m/s; the maximum join residual is 75.6 ms and p95 is 29.0 ms. The nonzero threshold was 0.1 m/s; the separate 3.0 m/s threshold is the frozen v36 strong-motion reference. The result supports a weak-axis-only observation: NavINST and Oxford remain unchanged because their runtime diagnostics are well-conditioned or too short-lived, while UrbanNav has the required sustained nonzero signal. These are sensor-only results, not accuracy claims.

The full sensor feasibility boundary is therefore passed for a conditional weak-axis route: metric visual motion is available on all three tracks, the visual observation is suppressed on well-conditioned directions, and UrbanNav supplies the sustained weak-axis case. The next authorized implementation step is only the source/unit contract for an isolated scalar shadow observation with bounded/stale checks. It must not replace the full visual pose prior, write x_curr.ba, or affect the v17 primary path until a separate one-repetition screen is explicitly completed. The join reports are under /media/user/aiueo/benchmarks/sota_v5/diagnostics/visual_longitudinal_v38_preflight/.

Isolated scalar shadow source/unit contract (2026-08-10)

The conditional route now has a source-side contract in docker/patches/voxel_slam_dev/v38_visual_longitudinal_shadow.patch. The patch is derived from the frozen v17 source and is disabled by default. It adds only an isolated scalar shadow observation record containing stamp_sec, signed metric velocity_mps, and normalized confidence; it does not alter the inertial pose, velocity, position, or bias state.

The receiver accepts an observation only when the feature is enabled, all values are finite, confidence is in [0,1], speed is within the configured bound (20 m/s default), and the measurement timestamp is finite, not in the future, and no older than the configured 0.2 s maximum age. Reset clears the shadow validity and counter. No ROS topic, camera producer, estimator correction, or output consumer is wired yet; this is intentionally a source/unit boundary, not a candidate runtime or accuracy experiment.

The patch checker and initial contract tests pass (12 passed across the v38 candidate, patch, visual diagnostic, and weak-axis join suites); compilation and git diff --check pass. The source-contract patch SHA at this checkpoint was 8a2b1cf6cc46ea5a9d853ac77504ca6f020b63be700b7827c126a4d1657f2ee4; the later opt-in receiver/consumer revision is recorded below. The next gate is an isolated report-only producer adapter that converts the already validated weak-axis visual pair into this scalar contract, followed by one authorized one-repetition screen. Until that screen is approved and passes, v17 remains the primary path and no fresh holdout is opened.

Report-only scalar producer result (2026-08-10)

The producer adapter is implemented in scripts/emit_visual_weak_axis_shadow.py, and the join now exposes deterministic per-scan observations in scripts/join_visual_weak_axis_diagnostic.py. The adapter consumes only the ground-truth-free pair report and the passed weak-axis join report. It emits the exact receiver fields stamp_sec (seconds), signed velocity_mps (m/s), and normalized confidence ([0,1]), plus provenance metadata; it does not publish a ROS topic or call the estimator.

Using the existing UrbanNav reports, the detailed join contained 675 weak-eligible per-scan observations. With the receiver speed bound 20 m/s and producer confidence floor 0.2, 534 observations were emitted and 141 were rejected at the confidence bound. The emitted confidence range was 0.2000--0.7039 with median 0.3850. The generated report is /media/user/aiueo/benchmarks/sota_v5/diagnostics/visual_longitudinal_v38_preflight/urbannav_visual_weak_axis_shadow_v38.json; the detailed join is /media/user/aiueo/benchmarks/sota_v5/diagnostics/visual_longitudinal_v38_preflight/urbannav_weak_join_v3.json.

The adapter contract tests, join tests, visual diagnostic tests, v38 patch tests, and patch checker pass (14 passed in the combined Python suite); compilation and git diff --check pass. This completes the report-only producer boundary. A runtime camera/ROS bridge, Voxel correction consumer, and one-repetition accuracy screen are still intentionally not executed because the v38 source patch remains a disabled shadow receiver with no estimator-state mutation. v17 remains primary and no fresh holdout or ground-truth input has been opened.

Visual transform correction and final v38 candidate screen (2026-08-10)

Before the candidate screen, a source-contract audit found two issues in the initial report chain. The visual diagnostic had reused the LiDAR-to-camera rotation when expressing camera-center displacement in the base frame; it now keeps LiDAR-to-camera projection and camera-to-base motion rotation separate. The weak-axis join also now resolves arbitrary eigenvector sign using the runtime world velocity, matching the frozen Voxel bridge convention that orients its weak direction toward the current velocity. The original v38 preflight/join/payload files and the first two candidate runs are retained as superseded artifacts, not final evidence.

The corrected fixed-configuration sensor-only results are:

Sequence Valid visual pairs Direction coherence Median scale (m) Median speed (m/s) Corrected weak-eligible pairs Nonzero streak Strong >=3 m/s streak Decision
NavINST Indoor02 1,106 0.995 0.34375 3.438 0 0 0 no-op
Oxford Spires Keble 05 8,285 0.973 0.06873 1.310 2 2 0 no-go
UrbanNav HK Tunnel 1 2,191 0.849 0.31250 4.145 775 89 40 GO_WEAK_AXIS_PROJECTION

The corrected UrbanNav join has projection absolute median 4.137 m/s, p10 1.315 m/s, p90 8.035 m/s, join maximum 75.6 ms, p95 29.0 ms, and positive weak-axis/runtime-velocity alignment median 6.127 m/s. The corrected producer again emits 534 observations from 675, with 141 rejected below confidence 0.2. The final report is /media/user/aiueo/benchmarks/sota_v5/diagnostics/visual_longitudinal_v38_preflight/urbannav_visual_weak_axis_shadow_v41.json and the join is /media/user/aiueo/benchmarks/sota_v5/diagnostics/visual_longitudinal_v38_preflight/urbannav_weak_join_v41_signed.json.

The candidate-only source patch was expanded to an opt-in geometry_msgs/Vector3Stamped receiver and a bounded weak-axis-speed consumer: source timestamp in the header, signed m/s in vector.x, confidence in vector.y, gain 0.25, and maximum weak-speed change 0.5 m/s per update. The receiver remains disabled by default; only the v38 UrbanNav config enables it. The image built successfully as sota-voxel-slam-v38:repo-v38@sha256:3b8fb3c48bbc3e1d3eca0dbbc9a1cfa2da6097aa0939815da7af242c382b5d42, with patch SHA fe9c313a9cacaffce333f9518f4c0c26cf955741dc9c44961ade9995b6a6e3b8.

The final corrected-payload runtime was one UrbanNav replay at rate 1.2 using the frozen dev_v34 canonical ROS1 bag. It completed cleanly (replay=0, mapper=0); the replay bridge validated/published 534/534 observations with zero late skips, and the mapper logged 600 applied shadow updates. The runtime RTF was approximately 0.844, but peak RSS was 342,604 kB = 334.57 MB, exceeding the frozen 330 MB limit. After the run completed, the existing UrbanNav common reference was used for the single accuracy screen: APE RMSE was 834.3353 m, versus the retained v17/v37 one-run reference 488.2263 m (+70.89%). No new reference or holdout was opened.

The corrected v38 route therefore fails the one-repetition accuracy and resource gates and is rejected for promotion. The map-geometry gate was not run after this failure. v17 remains the primary baseline; the v38 source/config/replay artifacts are retained for diagnosis only. The combined contract suite now passes 20 tests, the full patch checker passes, C++ compilation passes, and git diff --check is clean.

v38 causal audit and v39 output-only visual rejection (2026-08-10)

The v38 failure was reproduced without opening any reference trajectory. The ground-truth-free audit in scripts/audit_visual_shadow_runtime.py compares only the behavior-preserving v37 output, the rejected v38 output, the frozen visual payload, and the v38 source patch. It found all four prohibited runtime properties: wall-clock rather than estimator-state timing, repeated consumption of one observation, a scalar projected onto a producer-side frozen axis, and correction feedback into mapper state. The two trajectories were identical until the first correction application, 122.802 s after the first payload observation. Their position difference then crossed 1/10/100/500 m after 124.002/126.708/132.211/220.987 s, reached a maximum of 1,823.210 m, and increased path length from 4,572.247 m to 5,113.177 m. The audit decision is FAIL_VISUAL_SHADOW_RUNTIME_CONTRACT; its retained report is /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_v39_output_shadow_20260810/v38_runtime_audit.json.

v39 tests the narrowest safe alternative: a deterministic output-only shadow trajectory. The mapper, map, inertial state, and baseline orientation are immutable. The producer in scripts/emit_visual_velocity_vector_shadow.py keeps the complete calibrated base-frame velocity vector. The consumer in scripts/compose_visual_longitudinal_shadow_trajectory.py projects each observation against the timestamp-matched behavior-preserving state, consumes it at most once, and changes only the weak component of a new output trajectory. The same world-frame translation is applied to the reference-point output while its orientation is preserved. The fixed global configuration is configs/voxel_slam_v39/output_only_visual_shadow.yaml: eigenvalue ratio <0.2, horizontal weak norm >=0.9, baseline weak speed >=3 m/s, five consecutive scans, 80 ms join, 200 ms maximum age, gain 0.25, and maximum change 0.5 m/s per accepted observation. No dataset identity or ground truth is an input to either stage.

The hash-sealed pre-accuracy result was:

Sequence Vector observations Applied once Rejected join / gate Output decision Baseline preservation
NavINST Indoor02 0 0 0 / 0 NO_OP_OUTPUT_SHADOW raw and reference files byte-identical
Oxford Spires Keble 05 0 0 0 / 0 NO_OP_OUTPUT_SHADOW raw and reference files byte-identical
UrbanNav HK Tunnel 1 533 346 6 / 181 GO_OUTPUT_ONLY_SHADOW_SCREEN orientation unchanged; map and mapper untouched

UrbanNav path length changed from 4,572.2474 m to 4,493.5887 m; maximum and final position differences from the baseline were 158.1319 m and 157.4931 m. A second independent composition produced byte-identical raw and reference-point trajectories (SHA-256 d125a42a9a7d3d1ff5e9c2f4bbde3d71cc93974ba95650e4dc3c95c581d351d9 and bbff9328e2f650d919ccbc82b1f39e0a1e8f9ceb21c9c7b0a33a86a8f3f71671). The measured offline composition cost was 0.36 s wall time and 57,984 kB peak RSS; it does not run concurrently with or feed back into the mapper.

Only after those outputs were sealed was the existing UrbanNav common reference opened for the one authorized accuracy screen. With the same 395 interpolated pairs and zero rejected reference points, the behavior-preserving baseline measured 488.2306509 m APE RMSE and v39 measured 520.8089392 m: +32.5782883 m, or +6.6727%. v39 therefore fails the one-repetition accuracy gate and is rejected. No gain, threshold, sign, or timing parameter is retuned after this result; no three-repetition or map-geometry gate is run, and no fresh holdout is opened. v17 remains the primary baseline. All v39 evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_v39_output_shadow_20260810/.

The visual weak-axis route is now a retired negative control on these consumed development sequences. Its remaining value is architectural: output isolation prevents the catastrophic nonlinear map feedback seen in v38, but the camera velocity itself is not accurate enough to improve the frozen trajectory.

v40 full-scan GBA graph audit and sealed Oxford rejection (2026-08-10)

The initial remap premise was disproved before implementation. In the frozen source, Keyframe::id is already the source scan index, and HBA edges preserve that namespace. The v21 failure came from initializing GTSAM with only sparse keyframe IDs while declaring and writing back a dense range through the last keyframe. It also omitted the dense odometry chain and anchor prior. Changing the ID namespace would therefore have hidden the defect rather than fixed it.

The ground-truth-free audit in scripts/audit_v40_gba_graph_contract.py replayed the fixed window=10, stride=1, minimum distance=5 m selection against behavior-preserving trajectories. All IDs were unique, strictly increasing, and within the source-scan range:

Sequence Scans Sparse keyframes First-last scan ID v21 missing initialized keys Trailing scans outside v21 range Full-graph odometry edges Decision
NavINST Indoor02 3,485 141 9-3,289 3,149 195 3,484 GO_FULL_SCAN_GRAPH_RESTORATION_NO_ID_REMAP
Oxford Spires Keble 05 5,831 123 9-5,759 5,637 71 5,830 GO_FULL_SCAN_GRAPH_RESTORATION_NO_ID_REMAP
UrbanNav HK Tunnel 1 3,980 313 9-3,979 3,667 0 3,979 GO_FULL_SCAN_GRAPH_RESTORATION_NO_ID_REMAP

The reports are retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_v40_gba_graph_audit_20260810/. Their SHA-256 values are 33237d60... (NavINST), 6ad4352d... (Oxford), and 267929cb... (UrbanNav). Synthetic contracts cover contiguous, dropped/noncontiguous, duplicate, and out-of-range IDs without mutating a trajectory or map.

The opt-in/default-off docker/patches/voxel_slam_dev/v40.patch (SHA-256 890c351d05938921c427c577ce89f348c5ee39f393bf6a9e890724c1a6b3b4e6) restores every scan as a graph variable, adds N-1 odometry factors and a prior on scan zero, and fails closed on invalid keyframe IDs or variances. The three configurations differ from v17 only in the global-GBA contract. The patched voxelslam.cpp SHA-256 is 185be253dcf42a2d1917739fdc643b20b02a84358cdc4c31cef3171a3a48f86f; the source compiles and links in the voxel_slam_v40 Docker target.

One Oxford runtime-only screen was then sealed before any accuracy or map reference was opened. It used canonical bag SHA-256 8e7660e600f1f52758e0301f861a2174500ef38f60df5ac9f9d7198df9c54217, configuration SHA-256 6844604a99781fe0e86d7665227f4be6495022a272274012b33e30ee358e7800, rate 1.2, CPUs 2-7, and execution image ID sha256:48d5f6fe57e2bb69f16f3e1562d91d7a06e59ae1823983d3dbbc9f23cd65871a. Bag replay exited zero and the official /tf stream produced 5,831 valid samples through the input endpoint. The global backend did not finish within the 600 s finalization allowance, however: mapper exit was 143, container exit was 1, trajectory_complete=false, no alidarState.txt or map output was finalized, and no GBA completion marker was flushed. The 5,823 frontend PCD chunks were retained, but without the state file the runner could not assemble a world map. Total processing wall time was 1,085.6503 s for a 580.9839 s bag (RTF 1.86864). A live process sample reached 374,840 kB = 366.05 MiB, already above the 330 MiB gate; because the forced shutdown left mapper_process_time.txt empty, the formal run record correctly leaves peak_rss_mb=null rather than presenting the sample as a persisted peak.

The retained runtime evidence is under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v40/oxford_spires_keble_05/runtime_v40_full_scan_gba/. The sealed SHA-256 values are bc96b13b... for summary.json, 23a924e7... for run.json, 470597c3... for the raw trajectory, and 52bbea4e... for mapper.log. accuracy_ground_truth_accessed=false and accuracy_metrics_present=false; neither an accuracy scorer nor a map geometry scorer was run.

The execution image inherited a stale v17 voxelslam.cpp label even though its patch label was correct. This provenance-only defect was fixed after the failed run and locked by a test. The same compiled source was rebuilt as image ID sha256:5d07f7b2edd05c1e1eb38e5c30399f7f6dfb07b4792f069ce49bbe1fbb19f8d7; its source file and label both report 185be253.... The failed execution is still attributed to its original image ID and was not rerun or reclassified.

v40 therefore fails completion, runtime, and resource gates before accuracy and is rejected. No parameter was tuned from this result, no global three-sequence gate is authorized, and the fresh holdout remains unopened. v17 remains the primary baseline.

v41 cancellable/resource-bounded HBA rejection (2026-08-10)

The v40 source audit found a more precise completion cause than raw optimizer cost. Its standalone worker could observe is_finish && gba_flag == 0 and exit before the producer set the request. The producer would subsequently set gba_flag=1 and enter while(gba_flag);, with no worker left to acknowledge it and no sleep or ROS shutdown check. This request-before-exit race explains the silent one-core busy-wait and forced v40 termination.

v41 replaces that handshake with atomic request and producer-done state. A standalone worker may exit normally only after producer-done and a cleared request; the producer wait yields and checks ROS shutdown. The default-off runtime guard adds flushed stage/RSS/HWM markers, a global 330 MiB ceiling, and a 30 s backend deadline. Cancellation skips GTSAM writeback and map-topic publication, then saves the unchanged state so the runtime harness can close normally. The patch is docker/patches/voxel_slam_dev/v41.patch (SHA-256 69dcfede4d80fcabc4bc04d8846d4297f116798eb68ed8f2fc9d098a0d776c77), the patched source SHA-256 is e6516064abe6a16876af2b9b3e7cfb61519d992562b96f64bec484a5f963d66e, and the Release image is sha256:a72cacc930adf576e6572e1063e54a4948620c352ea10a947bf8aba30d645427. The image label and in-image source digest agree.

The ground-truth-free lifecycle audit covers success, RSS cancellation before a request, and deadline cancellation after a request. It requires worker acknowledgement before writeback and proves that both cancellation paths exit without writeback or map publication. Its decision is GO_V41_RUNTIME_ONLY_DIAGNOSTIC; the retained report SHA-256 is c3afbaf79bab3350426d7a657a29d40dbd1a9a5534421afef23d44c1cd81d366.

The single Oxford runtime-only execution used the same bag, rate 1.2, and CPUs 2-7, with config SHA-256 e5055d4c4ce47cbe6ac44237bceba3f68d78b21ae592048a9dc5c1464d6e1f10. It completed cleanly: replay, mapper, and container all exited zero; all 5,831 trajectory samples were present with a 0.00264 s end gap. Processing took 485.2646 s for RTF 0.835246, inside the 0.85 runtime limit.

The resource result rejects the backend. After 90 keyframes and 17 completed bottom-up windows, the seventeenth local cloud merge raised HWM to 352,172 kB = 343.91797 MiB. The guard emitted cancel_rss_limit at source scan 4,360, followed by worker_exit_cancelled; current RSS fell from 337,508 kB to 253,956 kB. The frontend then completed all 5,823 saved states/chunks and emitted backend_no_writeback and state_saved_unmodified. No worker_request, ISAM, writeback-complete, map-publish-complete, or optimized-state marker exists. Thus the fixed RSS gate is exceeded during bottom-up submap construction, before the restored full-scan graph or top-down optimizer is even requested.

The cancellation fallback is exactly behavior-preserving. Its alidarState.txt SHA-256 7f531a4121fc76afda19e3ac3b7cb7d075cc4d45884b64fc3ccf6c53ca92b522 is byte-identical to all three retained v17 Oxford repetitions. The assembled 30,901,030-point map SHA-256 2d0889336140638d3b6c87f2e4d0cf26dfe4b951d8a1ec8c7138032b55b9e4b4 is also byte-identical to v17. It is not an optimized v41 map and is not sent to an accuracy or geometry scorer.

The sealed v41 summary/run/log SHA-256 values are 7c591c52..., ab7ebe51..., and 3f72a8ed.... The independent runtime audit verifies clean completion, cancellation/no-writeback, resource failure, config binding, and v17 fallback identity; its decision is REJECT_V41_RESOURCE_GATE_RETIRE_BUILTIN_HBA and its report SHA-256 is 34a1e57c9af83fdb748fd0e05689b7268075488ca3f662951f94f166e54cf90c. All evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v41/ and /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_v41_gba_lifecycle_audit_20260810/.

v41 therefore fixes shutdown, observability, and fail-closed behavior, but fails the RSS gate and produces no map improvement. The built-in HBA route is retired for this candidate family; its keyframe distance, stride, voxel size, and memory threshold will not be tuned from Oxford. v17 remains primary, and no accuracy reference, reference map, or fresh holdout was opened.

v42 preregistered streaming external pose-graph feasibility

The remaining Oxford map blocker must now be addressed outside the retired in-process HBA. The next stage is report-only and fixed in this order:

  1. Inventory the retained v17 scan chunks and poses on all three consumed sequences, then identify temporally separated revisit candidates from the candidate trajectory only. Ground truth and reference maps are prohibited.
  2. Build fixed-budget submaps by streaming chunks from disk. Geometry-verify revisit pairs with one global set of distance, overlap, and residual gates; retain rejected pairs and peak working-set measurements. No trajectory or map may be changed in this stage.
  3. Add synthetic contracts for transform direction, frame identity, duplicate edges, disconnected graphs, invalid covariance, deterministic ordering, and fail-closed memory exhaustion.
  4. Only a report with repeatable, nonconflicting verified constraints on all applicable sequences authorizes an opt-in external sparse pose graph over the frozen v17 odometry chain. The online estimator remains untouched.
  5. A built candidate must first pass one Oxford runtime/RSS/completion screen. Accuracy and map geometry remain closed until that screen passes; one scored failure ends v42 without threshold tuning.

v42 streaming external pose-graph feasibility result (2026-08-10)

The five-stage order above was executed through its report-only authorization gate. The global contract is configs/sota_v6/development/v42_streaming_pose_graph_audit.json (SHA-256 c43475b61f87d621efd5528e034149b66e1c473f93e1206f4b3b1d7e21e446fb), and the implementation is scripts/audit_v42_streaming_pose_graph_feasibility.py (SHA-256 c266e10a7334cafdd6487519a2fc11a51fd6f5d5d58063db08299d75a3ec4253). The contract contains no dataset-specific branch. It reconstructs the exact 1.5 m v17 anchor policy, requires 100 m travel separation, searches a 20 m candidate radius, and tests a five-submap sequence at offsets [-6,-3,0,3,6]. Each anchor-local submap streams at most three chunks, is 0.5 m voxelized and capped at 6,000 points, and uses a three-entry LRU. The process RSS ceiling is 256 MiB.

The new discriminator combines a yaw-invariant 20-ring/60-sector place descriptor with candidate-episode uniqueness, independent constrained ICP at all five offsets, mutual overlap/residual limits, and a requirement that one world-frame correction explain the full sequence within 0.5 m and 2 deg. The complete edge set from the previously rejected 8 m profile is retained as a fail-closed challenge set: surviving its eight-anchor dedup neighborhood is an automatic rejection, not an accuracy label. The script has no ground-truth, accuracy-reference, reference-map, or optimized-trajectory input surface.

Six real-data audits completed, two per consumed sequence:

Sequence Anchors Episode queries Descriptor-qualified Geometry pass before dedup Verified constraints Legacy survivors Peak RSS range Decision hash
NavINST indoor02 575 302 36 0 0 0 79.90–80.16 MiB 1336b00a...
Oxford spires_keble_05 443 193 21 15 3 3 81.86–81.97 MiB ce111e88...
UrbanNav HK tunnel 1 2,005 204 65 0 0 0 95.98–95.99 MiB c2c731ef...

Each sequence's two deterministic payload SHA-256 values are byte-identical; wall times were 18.46/19.87 s, 16.80/17.51 s, and 50.07/47.04 s, respectively. The 16 synthetic contracts pass and cover transform direction, map <- base_link identity, duplicate constraints, disconnected graphs, invalid covariance, deterministic ordering, strict PCD/state parsing, and fail-closed memory exhaustion.

The Oxford result is decisive. Its three deduplicated constraints were 6 -> 303, 70 -> 365, and 164 -> 425, with minimum five-sequence mutual overlap 0.8613, 0.8869, and 0.8698. Despite that strong local geometry, all three match clusters from the already rejected edge set: respectively {3 -> 300, 14 -> 311}, {66 -> 362, 76 -> 371}, and {165 -> 427}. Long-sequence geometric consistency therefore still cannot identify the Oxford repeated structures. There are zero new verified constraints across all three sequences.

The report-only invariant also passed. State/map SHA-256 pairs remained 4162962d.../be95206f..., 7f531a41.../2d088933..., and d91e54f0.../87caad98... for NavINST, Oxford, and UrbanNav. No trajectory, pose graph, or map output was written. The aggregate report is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v42_streaming_pose_graph_audit_20260810/; its file SHA-256 is b00ccca9d286d9239ff76b0a67e6ab8d1fdfa508b65b9c0a04764a0647d156cd and its deterministic payload SHA-256 is 63c7176f74d3963936b2be3bfe2a954140e58d9261fc2f7363d840ffa9fd865a.

The sealed decision is REJECT_V42_LEGACY_AMBIGUITY_SURVIVED. External sparse pose-graph implementation is not authorized, so the Oxford runtime/RSS candidate gate and all accuracy/map scoring remain unopened. v42 thresholds are retired without tuning, the online estimator remains untouched, and v17 remains the primary behavior.

v43 preregistered independent place-identity feasibility

Another geometry-only loop gate is not justified. Any v43 revisit proposal must first add an independent place-identity signal, such as temporally separated camera appearance or calibrated LiDAR reflectivity, and demonstrate that it rejects every retained v42 Oxford ambiguity before producing an edge. The first stage remains report-only, globally configured, fixed-memory, and forbidden from opening accuracy/reference-map inputs. If the retained sensor inputs cannot supply an independent signal on all applicable sequences, the Oxford global-correction route closes and development returns to a new local frontend estimator rather than tuning loop thresholds.

v43 independent place-identity availability result (2026-08-10)

The retained-signal gate was implemented before any new matcher or edge screen. The global contract is configs/sota_v6/development/v43_place_identity_audit.json (SHA-256 39fa10345406b3a1019c5acfe6540705c11fdee3cc47f40c755b25720f427eca), and the bounded streaming implementation is scripts/audit_v43_place_identity_feasibility.py (SHA-256 88475a67bac7cbf51ae42d7c2ff22bb83789961ca6ee0ffa35ed77c44b348087). The contract has no dataset branch. It requires finite, nonconstant XYZI intensity with at least 1% nonzero support, 16 distinct values, and 90% of chunks varying, or an explicit sequence-bound camera manifest. It reads one PCD at a time under a 128 MiB RSS ceiling. Geometry, loop edges, trajectories, accuracy references, and reference maps are not inputs.

The complete retained chunk set was scanned twice for each sequence:

Sequence Chunks Intensity points Nonzero Dynamic range Distinct values Varying chunks Camera retained Peak RSS range Decision hash
NavINST indoor02 3,477 11,727,179 0 0 1 0 no 38.05–38.07 MiB 63f72511...
Oxford spires_keble_05 5,823 30,901,030 0 0 1 0 no 40.31–40.34 MiB f5d72be1...
UrbanNav HK tunnel 1 3,972 17,711,712 0 0 1 0 no 38.93–38.94 MiB 65f89f37...

All 60,339,921 retained intensity values are finite but exactly 0.0. The per-sequence intensity-payload SHA-256 values are 38c4b1d3..., 6634efba..., and 710dd1a3...; each decision payload is byte-identical across the two runs. No camera manifest or image payload is retained in any v17 run, and the original canonical LiDAR/IMU bag paths recorded by the runs are not currently mounted. The 12 synthetic contracts pass for zero and valid signals, nonfinite values, malformed/oversized PCDs, state ordering, camera sequence binding, memory exhaustion, protected-input identity, and repeatability.

State/map SHA-256 pairs remained 4162962d.../be95206f..., 7f531a41.../2d088933..., and d91e54f0.../87caad98.... No trajectory, map, matcher, or pose-graph output was produced. Evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v43_place_identity_audit_20260810/. The aggregate file SHA-256 is a43885ffa7ab8d2ca8859e76a0cbcb67895a44433b690856a4ade521ebb92596, and its deterministic payload SHA-256 is 246beb418b62a3dbe57739f25fe940d0eac3ddb0a712dc900393987e341e7b2d.

The v43 decision recorded at that stage was (superseded by v43b below) CLOSE_OXFORD_GLOBAL_CORRECTION_ROUTE_NO_INDEPENDENT_PLACE_IDENTITY. A reflectivity/appearance challenge matcher is not authorized because there is no independent signal to consume. This closes the v17 Oxford global-loop, external pose-graph, and map-refinement branch without threshold tuning. v17 remains primary; accuracy, map scoring, and fresh holdout stay closed.

v43b exact-raw correction and place-identity challenge (2026-08-10)

The v43 retained-PCD measurement remains valid, but its source-availability inference is superseded. A later read-only inventory found exact-byte copies of all three canonical ROS1 bags under /media/user/aiueo/benchmarks/quarantine/opencode_inputs_20260810/. Their SHA-256 values exactly match the original run bindings: b8afd9649a310669... for NavINST, 8e7660e600f1f527... for Oxford, and 95524232d9a4c278... for UrbanNav. Thus the saved v17 PCD chunks do contain only zero intensity, but they are not the only retained source from which an independent signal can be recovered. No v43 evidence was rewritten.

The corrective gate was frozen before reading any Oxford identity score. Its global contract is configs/sota_v6/development/v43b_raw_intensity_identity_audit.json (SHA-256 28e5a6b8216f33e1d8a38ca2773b41cbb647696c3283be4e7c295269031e0c1f), the exact source binding is configs/sota_v6/development/v43b_raw_intensity_sources_20260810.json (SHA-256 c24f4443367e32507ed2af81f88a5dc9b0c95d775fc0b2fddcfed80f3d3b6f2a), and the report-only implementation is scripts/audit_v43b_raw_intensity_identity.py (SHA-256 59377380bb245185d297294c25fd4d25bace2f472fa8964d36d54edee2bdf717). The contract has no dataset branch. It requires the exact common 48-byte PointCloud2 layout, header-to-state matching within 1 ms, at least 90% nonzero intensity, a 16-unit dynamic range, 32 distinct values, and 90% of scans varying. Raw points are transformed by the frozen v17 convention p_body = R_body_lidar p_lidar + t_body_lidar, range filtered to 1–50 m, 0.5 m voxelized, and bounded to 12,000 points per selected scan and 6,000 points per three-scan submap under a 256 MiB RSS ceiling.

Every retained v42 geometry pass is then challenged before edge creation. Each of its five registered submap pairs must have at least 128 reciprocal 0.5 m correspondences, 5% support, nonconstant overlap intensity, overlap-local Pearson correlation at least 0.6, and a correlation peak margin of at least 0.08 over four fixed 1 m spatial decoys. Ground truth, reference maps, accuracy trajectories, optimized maps, and graph writeback are absent from the interface.

The exact bags were scanned twice per sequence:

Sequence Raw scans Raw intensity points Nonzero fraction Dynamic range Distinct sampled values Varying scans Peak RSS range Decision hash
NavINST indoor02 3,489 89,292,327 99.0613% 255 240 3,489/3,489 67.39–67.41 MiB ba6f0d73...
Oxford spires_keble_05 5,834 345,360,040 99.999993% 255 118 5,834/5,834 151.14–151.29 MiB 9a934fbb...
UrbanNav HK tunnel 1 3,982 246,291,863 97.9559% 255 254 3,982/3,982 95.08–95.14 MiB 4e0a4eb3...

All 680,944,230 values are finite, and each complete intensity payload and decision payload is byte-identical across repetitions. Oxford required 115 raw scans for its 15 geometry-qualified candidates; all 115 matched the v17 state clock, with maximum absolute error 0.307016 ms. The selected-cloud payload SHA-256 is f6f18abe....

Oxford's 75 pair scores explain why raw intensity does not safely rescue the global route. Selected-alignments have high Pearson correlation (0.8687 minimum, 0.9071 median, 0.9483 maximum) and strong reciprocal support (2,619 minimum), but the same intensity texture remains locally ambiguous. Only 39/75 pairs meet the frozen 0.08 spatial-peak margin; the margin range is 0.0240–0.1257. Every one of the 15 five-pair candidates therefore fails at least one independent identity pair. This rejects all three old ambiguity clusters before an edge exists: 6 -> 303, the 68--75 -> 363--370 cluster, and the 158--164 -> 419--425 cluster. There are zero legacy survivors and zero new verified constraints.

The 18 v43b synthetic contracts and the complete v40–v43b regression set pass (70 passed). The aggregate evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v43b_raw_intensity_identity_audit_20260810/. Its file SHA-256 is f6bd29ec58b78079751bcd19fd676ddafc28a1fc04eb80aef5ef939654900c65, and its deterministic aggregate payload SHA-256 is 9cb9e874f879b337439a9b761235898ace128751cb6f86401a63a62e05f9496b.

The corrected sealed decision is CLOSE_V43B_GLOBAL_CORRECTION_ROUTE_NO_NEW_UNAMBIGUOUS_CONSTRAINT_SET. Independent raw intensity is available, so v43's availability conclusion is retired. The global-correction conclusion remains closed for the stronger reason that the preregistered identity challenge removes every ambiguous candidate and leaves no replacement constraint. External pose-graph and map refinement implementation remain unauthorized; v17 remains primary and no accuracy/reference-map input was opened.

Next workstream: v44 local-estimator architecture boundary

The next candidate must be local and causal rather than another global correction. The exact-hash raw-input prerequisite is now satisfied by the v43b source manifest. v44 proceeds in this fixed order:

  1. Perform a report-only LiDAR/IMU clock, rate, gap, covariance, unit, frame, and startup inventory on all three exact bags. Freeze one common source, timestamp, and memory contract without accuracy or map references.
  2. Define one fixed-lag shadow estimator with explicit IMU preintegration, gyro/accelerometer bias state, gravity handling, per-factor observability, bounded marginalization, and no loop closure or global map correction.
  3. Add synthetic contracts for preintegration direction, bias Jacobians, timestamp gaps, marginalization consistency, weak-axis observability, deterministic ordering, and fail-closed resource exhaustion.
  4. Replay in shadow mode first: consume the exact raw stream and export only diagnostics while v17 remains the published trajectory. Require bounded RSS, completion, deterministic state payloads, and no protected-output writeback on all consumed sequences.
  5. Only after that runtime gate passes may one frozen v44 candidate open the three-sequence accuracy and map screens. One scored failure ends the candidate without threshold tuning; fresh holdout remains last.

The model must address Oxford vertical-bias accumulation and UrbanNav weak-axis speed in one global architecture rather than reusing the retired direct bias/velocity threshold variants from v22–v38. The immediate next implementation is therefore the stage-1 raw LiDAR/IMU contract and inventory, not another loop matcher.

v44a stage-1 raw LiDAR/IMU readiness result (2026-08-10)

The report-only stage-1 inventory is complete. The active global contract is configs/sota_v6/development/v44_raw_lidar_imu_readiness_audit.json (contract ID v44a-raw-lidar-imu-readiness-20260810, SHA-256 a71d9271d200e86537cb6f8b25e21662ffe120258e749e6dff17a77e6d1a162b), the exact bag/topic/frame/provenance binding is configs/sota_v6/development/v44_raw_lidar_imu_sources_20260810.json (SHA-256 002cc82fcd86a70fb774e1e3f576b41cbe86b849eabcfc6406e98cb26232ff1d), and the bounded one-pass implementation is scripts/audit_v44_raw_lidar_imu_readiness.py (SHA-256 db32ce9b65c4ab54825585e801ebbb26984b56fad86a430d7b27c84dacae7500). The contract has no dataset branch, consumes no accuracy trajectory or reference map, writes no trajectory or map, and enforces a 128 MiB RSS ceiling.

One preregistration correction is retained explicitly. The first NavINST preflight used contract SHA-256 2336c29dce06e38f4be3dd4f6d9495ca42e2bd4bef68015f05540dbfae8771ea and rejected readiness solely because only 94.898252% of scans contained a physical return at exactly t=0. All its other LiDAR, IMU, provenance, normalization-digest, and synchronization checks passed. A point timestamp defined as unsigned nanoseconds from scan start need only be a non-negative offset; it does not imply that the sensor observed a return at the exact scan origin. The rejected contract is preserved byte-for-byte as configs/sota_v6/development/v44_raw_lidar_imu_readiness_rejected_preflight.json, and the original report remains under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44_raw_lidar_imu_readiness_20260810/. Before reading any Oxford or UrbanNav score, v44a replaced that invalid invariant with the exact uint32 non-negative-offset schema plus mandatory equality to each sealed normalization timing digest. No observed performance threshold was relaxed.

Each exact bag was then scanned twice under v44a:

Sequence LiDAR rate / max gap IMU rate / max gap Fully bracketed Max boundary distance Orientation / covariance Startup Peak RSS range
NavINST indoor02 9.9147 Hz / 101.121 ms 125.0091 Hz / 7.999 ms 99.9713% 7.999 ms absent / absent stationary 67.66–67.76 MiB
Oxford spires_keble_05 10.0403 Hz / 101.636 ms 399.5339 Hz / 2.534 ms 99.9829% 2.520 ms absent / absent dynamic 118.08–118.14 MiB
UrbanNav HK tunnel 1 9.9914 Hz / 105.746 ms 400.0048 Hz / 9.227 ms 99.9749% 4.332 ms provided / provided stationary 97.44–97.55 MiB

All six runs pass. Every LiDAR payload has the exact required 48-byte schema, all point-time normalization digests match their sealed source reports, all IMU measurements are finite and within the frozen norm bounds, and each sequence has only one unbracketed prefix scan with zero interior or suffix coverage holes. The complete deterministic payload SHA-256 is identical between repetitions for each sequence: 6c611320..., 90b1cba5..., and 6604d0c5.... The 18 v44a synthetic contracts and the complete v40–v44 regression set pass (88 passed).

Aggregate evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44a_raw_lidar_imu_readiness_20260810/. The aggregate file SHA-256 is 532b62ad534cb533c1ecfaa735a844f68dfb38a0a4c5184ef598ab98f3b959fb, and its deterministic aggregate payload SHA-256 is 57c14d126beeacc741471864bc6828869e2f7f3980edfba2b3e62f93c5b437c0.

The sealed decision is AUTHORIZE_V44_FIXED_LAG_ARCHITECTURE_DEFINITION. This authorizes only the stage-2 architecture contract. Shadow-estimator implementation remains unauthorized. Because orientation and covariance are not universal and the Oxford startup window is dynamic, that architecture must use one explicit noise model, orientation-independent initialization, and dynamic-startup initialization across all datasets. Dataset-specific algorithm thresholds, loop closure, and global map correction remain forbidden. Accuracy and map screens remain closed until the architecture, synthetic contracts, and report-only shadow runtime gates have passed in order.

v44b stage-2 fixed-lag architecture result (2026-08-10)

The architecture-definition stage is complete. The machine-readable global contract is configs/sota_v6/development/v44b_fixed_lag_shadow_architecture.json (contract ID v44b-fixed-lag-shadow-architecture-20260810, SHA-256 74a80cca0e3dc7de9a10950d50c4fea7518dfe3d4fed8c5e1aeec77c3efee8e7). Its readable design and failure analysis are in docs/research/v44-fixed-lag-shadow-architecture-2026-08.md (SHA-256 3a2267b5ae0a1d3115b4900e55b593de28278e904762ff39d6e712ff64811edc). The static validator is scripts/validate_v44b_fixed_lag_architecture.py (SHA-256 43576fb83f6011f0724b672f6671a849302b1862ff9c74f1ad9a6d1e1b37a232). It verifies the exact v44a contract/source/aggregate hashes before accepting the architecture and has no bag, estimator, accuracy, trajectory, or map execution interface.

The single architecture uses one 15-DoF state at every accepted scan end: R_WB, p_WB, v_WB, gyroscope bias, and accelerometer bias. Dynamic initialization jointly estimates all bootstrap states, both biases, and a fixed-magnitude S2 gravity direction over a 2–5 second window. Message orientation, stationary assumptions, zero-velocity factors, and message covariance are not used. After deterministic gravity/yaw rebasing, streaming uses g_W = [0, 0, -9.80665] m/s² and the exact fixed sensor extrinsics.

Adjacent states use bias-linearized SO(3) midpoint IMU preintegration with all five required bias Jacobians and an explicit common continuous-time noise model. LiDAR contributes only causal binary point-to-plane factors between active knots. Each whitened factor is projected by deterministic SVD onto its six-DoF relative-pose observable subspace. Rejected modes preserve process uncertainty; they do not freeze an axis or directly overwrite velocity or bias. Thus the architecture does not claim to invent UrbanNav tunnel-axis speed, but it prevents weak geometry from becoming false state information.

The smoother has a three-second, 64-knot fixed lag. This covers the v44a maximum permitted 20 Hz LiDAR rate (62 required slots), bounds the active state dimension at 960, and bounds active LiDAR correspondences at 768,000. Optimization is single-threaded, fixed-order, and limited to four Gauss–Newton iterations with streaming square-root QR. Marginalization uses a square-root separator prior with first-estimate Jacobians; prior reset and ad-hoc covariance inflation are forbidden. The limits remain 330 MiB RSS and 0.85 processing RTF.

The exact v44a startup coverage policy is inherited rather than retuned: at most two unbracketed prefix scans may be dropped before the first state, while any interior boundary gap after startup fails closed. Sensor time is integer nanoseconds, every observation is consumed once, and wall clock is never an estimator input. v17 files are hash-checked before and after; ROS publication, primary state/map/bias writeback, loop closure, global correction, camera, and visual-speed feedback are all prohibited.

The architecture explicitly removes the failed v22–v39 mechanisms: raw coordinate HTH gates, direct velocity or accelerometer-bias corrections, weak-axis speed/history latches, producer-frozen axes, wall-clock age, repeated observations, and visual feedback. It also keeps the v40–v43 global correction routes closed.

The contract pre-registers 20 stage-3 synthetic contracts covering SO(3) direction, gravity sign, bias Jacobians, covariance propagation, timestamp gaps, deskew and extrinsic direction, binary LiDAR Jacobians, weak-subspace projection, dynamic initialization, square-root/FEJ marginalization, fixed-lag eviction, deterministic payloads, resource exhaustion, and protected outputs. The 27 architecture-boundary tests pass, and the complete v40–v44b regression set passes (115 passed).

Static validation was executed twice. Both reports bind the same contract and validator hashes and have the identical deterministic payload SHA-256 2ba8e5d7f5191658d4ced36aecf56514f10210ec78f2b2c380eca22585d3db4f. Evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44b_fixed_lag_architecture_20260810/. The aggregate file SHA-256 is 9bc723886d3faa067de979e203eb9e49d534b2252fb733e05f08de386fb64407, and its deterministic aggregate payload SHA-256 is 7794446259eb915e967294e3c515495b23abef57a19398cc2f1beac8b8ec6d52.

The sealed decision is AUTHORIZE_V44_STAGE3_SYNTHETIC_CONTRACT_IMPLEMENTATION. Only the numerical synthetic harness is now authorized. Fixed-lag estimator implementation, raw shadow replay, accuracy/reference-map inputs, and primary trajectory or map mutation remain unauthorized. The immediate next work is to implement the preintegration, LiDAR Jacobian/observability, initialization, and square-root marginalization primitives only inside synthetic tests, then aggregate two deterministic stage-3 validation runs before considering a shadow estimator implementation.

v44c1 stage-3 fixed-lag synthetic-contract result (2026-08-10)

The synthetic numerical-contract stage is complete. The original 20-case scenario contract is configs/sota_v6/development/v44c_fixed_lag_synthetic_contracts.json (contract ID v44c-fixed-lag-synthetic-contracts-20260810, SHA-256 0f6c57303f9889b0fd1a21b92ef80bf7d14a632010ad68222a7b6efd816e326f). The active resource-scope correction is configs/sota_v6/development/v44c1_fixed_lag_synthetic_contracts.json (contract ID v44c1-fixed-lag-synthetic-contracts-20260810, SHA-256 e5e8a8a3ae8d7e6b17ae8c7c6580751ac3bee8ff914ba80274222617aedb6c92). It binds the original scenarios, the v44b architecture, and prior aggregate evidence by hash. The validator is scripts/validate_v44c_fixed_lag_synthetic_contracts.py (SHA-256 ce5e5c5c69849838c7c0a7b6167d3bb67738b78ff106ffa63dbfd4ec1847ad66), and the readable contract report is docs/research/v44-synthetic-contracts-2026-08.md (SHA-256 65f3b43bfa8bbc0fd3ed6952623acb5eaba92625c8fe788a17340bb20b2bd6da). The implementation has no bag, ROS, trajectory, map, accuracy, or reference input interface and cannot mutate the primary estimator.

The first standalone v44c run passed all 20 numerical cases at 42.34 MiB RSS. However, the full pytest host already occupied 131--143 MiB before the synthetic harness began, so the original 128 MiB absolute ceiling rejected embedded tests before any v44c allocation. The preliminary evidence remains preserved under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44c_fixed_lag_synthetic_contracts_20260810/; its aggregate file SHA-256 is f3cb78a04efad5615bc37d9ee32242b64238375c71bb21b7a52f1401d6e0a55f and its aggregate payload SHA-256 is c89f1cf5fcece0799afd7d897423bf17ee850b85156706f2fc34cd68c00e4c6f. This was a resource-measurement modeling defect, not a numerical failure. v44c1 therefore changes only measurement scope: a standalone process retains the absolute 128 MiB ceiling, while an already-loaded host receives a 64 MiB incremental ceiling. No scenario, seed, ordering rule, motion, geometry, numerical threshold, or tolerance changed.

The reference implementation now covers SO(3) exponential/logarithm and right perturbations, integer-nanosecond IMU bracketing, midpoint preintegration with covariance and bias Jacobian oracles, scan-end deskew, fixed extrinsics, analytic binary point-to-plane Jacobians, deterministic SVD observable projection, dynamic initialization, square-root marginalization, immutable first-estimate Jacobian priors, fixed-lag eviction, deterministic serialization, resource rejection, and protected-output hashing. These are bounded synthetic oracles, not a streaming estimator.

All 20 frozen challenges pass. Representative results are 9.94e-17 rad SO(3) direction error, 7.11e-15 maximum constant-motion residual, 1.11e-10 rad gyro-bias reintegration error, 3.34e-12 m/s accelerometer bias velocity error, positive-semidefinite covariance with a 2.46475 trace growth ratio, 4.69e-11 maximum binary LiDAR Jacobian error, and 1.55e-15 SVD projector rotation-invariance error. The weak-axis case retains rank five and exactly zero weak-mode update/information while preserving the 0.75 process increment. Dynamic startup recovers nonzero initial velocity without an orientation, stationary, or ZUPT branch. Square-root marginalization matches the batch solution to 6.11e-16; FEJ gauge error is zero; a four-second 20 Hz stream leaves 61 active and evicts the 20 oldest states. Timestamp-gap, ordering, duplicate, resource, and write failures all reject with zero state output and unchanged protected outputs.

The first numerical execution also exposed a real implementation defect: the SO(3) logarithm returned twice the intended rotation vector because both the skew-vector extraction and scale supplied a factor of two. The implementation was corrected without relaxing the frozen tolerance. The 46 focused tests pass, including finite-difference, malformed-input, prerequisite-hash, authority, allocation, repeatability, and incomplete-report negatives. The complete v40--v44c1 regression passes (161 passed in 2.60s).

Two final standalone validations both produced report payload SHA-256 710564cc07242d98e9bddd0170362a20c2741e4318ddb5602f8964699800b86c and combined case payload SHA-256 7c6fd8e106885f90c1fc1627939212bfaf454c6692ae9be15aa1abb06ea593d2. They began at 37.10--37.14 MiB RSS, peaked at 42.47--42.55 MiB, and added 5.34--5.45 MiB. Evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44c1_fixed_lag_synthetic_contracts_20260810/. The run file SHA-256 values are 33825a159c396bf8a62d6001e1afdd07044add65b2f51204d7da5c96f1a4c6db and 036ad5d4dc636681b94eb6155bee76a2fff0a12407dbb7f1721faec37be555d9. The aggregate JSON SHA-256 is 18dc483ae4a924608d16ed3f47b7dd5bb812f6baea027d931d7f2010287bf663, and its deterministic aggregate payload SHA-256 is c89f1cf5fcece0799afd7d897423bf17ee850b85156706f2fc34cd68c00e4c6f.

The sealed decision is AUTHORIZE_V44_STAGE4_REPORT_ONLY_SHADOW_IMPLEMENTATION. This authorizes only implementation of an isolated, diagnostic-only fixed-lag shadow estimator. Raw shadow replay is still unauthorized. Before replay can be considered, the source must be complete, pass a static audit against every v44b boundary, and receive a separate execution contract. Accuracy or reference-map inputs, primary trajectory/map/bias writeback, loop closure, global correction, and any production-path mutation remain forbidden. The immediate next work is stage 4 source implementation plus static source audit; it is not raw-data execution.

v44d stage-4 fixed-lag shadow-source gate (2026-08-10)

The isolated estimator-core implementation and pre-replay source audit are complete. The machine-readable gate is configs/sota_v6/development/v44d_fixed_lag_shadow_source_audit.json (contract ID v44d-fixed-lag-shadow-source-audit-20260810, SHA-256 462d34fb984a45f4a15a5482d2d0e9f61e45c68491e2ddefa9a7fdaa3348df5e). The report-only estimator core is scripts/v44_fixed_lag_shadow_estimator.py (SHA-256 c353cce1cde17e45ea1244871da8a7077f10b28055f8c8cdfae49a5bddaf8ee9), the AST/static validator is scripts/audit_v44d_fixed_lag_shadow_source.py (SHA-256 8d2644e9b05102029743f6cff7701913c88bacaa61fb4cf231b9da026e14bd06), and the readable source-gate report is docs/research/v44-shadow-source-audit-2026-08.md (SHA-256 7bf2bae05d9f96aa152826525bd77f5f76098fd3eabb4435e53bfb10336c5eb4).

The core has no CLI, raw-bag decoder, ROS subscriber/publisher, filesystem, network, subprocess, wall-clock estimator input, dataset identity, accuracy input, reference map, or primary-state callback. Its public sensor boundary is limited to immutable body-frame IMU records, LiDAR points/scans with integer nanosecond time, one fixed T_BL, and externally supplied resource/protected- output observations. It returns in-memory diagnostic records only. No raw bag was opened and no runtime adapter exists in this stage.

The implementation follows the single v44b path: exact point-time IMU integration and scan-end deskew; deterministic range/voxel/surfel handling; one 15-DoF state per accepted scan; midpoint SO(3) preintegration with 9x9 covariance and all five bias Jacobians; binary point-to-plane factors; Huber whitening; per-factor relative-pose SVD projection; chronological/type factor ordering; bounded block Householder QR and rank solve; fixed line-search work; three-second/64-knot eviction; square-root FEJ separator priors; source-surfel removal; and diagnostic-only full-lag bias Schur information. Every capacity is checked before the relevant allocation or output, and a terminal failure has zero valid state count and no state payload hash.

Bootstrap uses the same dynamic path for moving and stationary inputs. The first LiDAR translation divided by sensor time is only the velocity seed. All bootstrap knots, both biases, and a fixed-magnitude two-DoF S2 gravity direction enter one joint solve under the single broad prior set. The result is rebased deterministically to negative-z gravity, first position at the origin, and first yaw zero. Message orientation, stationary detection, ZUPT, and direct state-component correction are absent.

Semantic review before sealing found that the first draft had kept gravity fixed during bootstrap even though knot and bias states were jointly solved. That incomplete implementation was corrected by adding the S2 gravity variable and broad priors. The same review replaced a linear within-scan pose proxy with integration at every exact point timestamp, changed the velocity seed to the specified first translation/time pair, and added the full-lag bias-information diagnostic. No threshold, input authority, or replay authority was relaxed; all final evidence was regenerated after the corrections.

The static validator parses 21,393 AST nodes and checks 21 classes, 29 module functions, 33 estimator methods, and the exact eight-root import allowlist. All 36 source-boundary checks pass. Ten independent in-memory smoke probes also pass: zero-error SO(3) round trip; all five finite bias Jacobians; two-DoF S2 gravity with zero magnitude error; PSD covariance with minimum eigenvalue 9.99972e-6; binary LiDAR Jacobian error 4.68738e-11; Householder dense- solution error 9.45728e-16; separator-solution error 8.94472e-16; immutable FEJ payload; protected-output mutation rejection; and pre-allocation resource rejection.

The full frozen-configuration synthetic stream runs 11 planar scans with 100 Hz IMU twice. Both runs produce state payload SHA-256 928df5580492ecd6471427bcdc2e6e474ce5921753c2f782826b29d46b6353de and diagnostic payload SHA-256 1512d9971a1853ed3dac6acc53003a2d19abe379d48a0e9f7d2abf1723fd5257. It reaches LiDAR observable rank three, reports 62 retained modes in the 66- dimensional full-lag bias-information system, and finishes with exact gravity, origin, and yaw gauges. A shortened-lag integration test evicts the three oldest of seven states, leaves knot IDs 3,4,5,6, removes old surfels, and retains an immutable separator prior. The 28 core tests and 23 audit tests pass, and the complete v40--v44d regression passes (212 passed in 18.39s).

Two final static validations have identical deterministic report payload SHA-256 9a7bf365e5a90e0f0a1b1296b358305277091059d0ea8bb3beece45fdea2b7d2. They began at 34.38--34.40 MiB RSS, peaked at 50.45--50.46 MiB, and added 16.06--16.07 MiB. Evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44d_fixed_lag_shadow_source_audit_20260810/. The two report file SHA-256 values are 1b6aeba274d75329ac69d58bba0f1991309aee93966d10dcc7cb089718be333d and 8f6ee2eb85209f27c3155778478de7a11bdd4139f09a04f5510f63b28b4aa06e. The aggregate JSON SHA-256 is b2494f42b07d9fd844f14112da5b6df6fd050007326018967f9f794b0cb5e997, and its deterministic aggregate payload SHA-256 is 2344a9e9b31861fff44c1b457199a2863615a2c06834a6a9c39a710fc53b97b2.

The sealed decision is AUTHORIZE_V44_STAGE4_RAW_SHADOW_REPLAY_CONTRACT_DEFINITION. This does not authorize raw replay. It authorizes only definition of a separately hash-bound execution contract and a read-only decoder adapter outside the estimator core. That next contract must bind the exact v44a source manifest and all three bags, write only into a new bounded evidence directory, enforce 330 MiB RSS and 0.85 RTF after every scan, compare protected v17 hashes before and after, and require completion plus identical state payloads across two repetitions. Accuracy, ground truth, reference maps, ROS publication, primary trajectory/ map/bias writeback, loop closure, and global correction remain forbidden.

v44e raw-shadow replay execution-contract gate (2026-08-10)

The separately hash-bound raw replay contract and read-only decoder adapter are complete. The machine-readable contract is configs/sota_v6/development/v44e_raw_shadow_replay_execution_contract.json (contract ID v44e-raw-shadow-replay-execution-20260810, SHA-256 08d7f8850f5081fd18d58a860cae3d9566846605460dadb37660b3c3111d1692). The adapter is scripts/v44_raw_shadow_replay_adapter.py (SHA-256 6fbf16d811a5bcf26c2cac9190f0557844cd3596cf7704cfdc016c7080ddf0ee), the static auditor is scripts/audit_v44e_raw_shadow_replay_contract.py (SHA-256 8fa736d6c59b17eb3364c50cad0a9ec88cf01ffe867864e9558387db37909389), and the readable contract report is docs/research/v44-raw-shadow-replay-contract-2026-08.md (SHA-256 990a80dff5adafd086ae132050772696c25823c46132efc4093e57d7b8f0db68).

The contract explicitly repeats and verifies all three v44a bag SHA-256 values, sizes, topics, frames, message/point counts, and serialized LiDAR/IMU stream digests. The three sensor-adapter hashes and exact T_BL matrices are sealed as source calibration rather than algorithm branches. The estimator still sees one global v44b configuration and receives no sequence ID, dataset name, bag path, orientation, covariance, protected path, accuracy input, or output callback.

Bag receive order is converted into the architecture's sensor-time order by a bounded watermark heap. The exact observed maximum receive-minus-header delays are 3,573,549 ns for NavINST, 135,957,579 ns for Oxford, and 136,013,359 ns for UrbanNav. The heap is limited to 128 messages and 32 MiB, emits by header time with IMU before LiDAR and stable source index, and terminates on a larger delay, receive-time reversal, duplicate/out-of-order output, or capacity overflow. PointCloud2 remains the exact 48-byte little-endian v44a schema; IMU decoding uses only header time, angular velocity, and linear acceleration.

The retained v17 run_01/trajectory.tum and run_01/pcd/voxel_global.pcd files are individually path/size/hash-bound for each sequence. Before and after a raw shadow run, the adapter rehashes both and passes only the digest dictionary to the core protection guard. The core cannot open those paths. A changed v17 state or map is terminal and cannot yield a valid shadow result.

The runtime CLI exposes only contract, sequence ID, and repetition. Bag, output, trajectory, map, and calibration paths cannot be supplied. Output is restricted to a new nonsymlinked raw_replay/<sequence>/run_NN/{diagnostics.jsonl,run.json} directory under the v44e evidence root, opened without overwrite and capped at 256 MiB. RSS and cumulative processing RTF are passed to the core after every completed scan; the frozen ceilings remain 330 MiB and 0.85. A complete run must also match the bound message, point, and serialized-stream identities.

The activation chain is fail closed. Before importing rosbags or opening a bag, the adapter verifies every prerequisite/source hash, its own hash, the contract hash, and the v44e aggregate decision. It then reopens and rehashes both static source reports listed by that aggregate, requiring repetitions one and two, identical deterministic payloads, and explicit proof that neither static pass opened or replayed raw data. Thus an aggregate filename alone is not sufficient authorization.

All 29 static checks and 10 synthetic adapter probes pass. The focused v44e suite passes (49 passed in 4.89s), and the complete v40--v44e regression passes (261 passed in 24.69s). Two final standalone static validations have identical deterministic report payload SHA-256 b98739d779a6e79b2027f100172230f48700b4045dc4b8610b3b54e5fe503d3f. Their report SHA-256 values are 27d98bb4fe182efa5c38ce97f5ee8f876e588ca353e0a393be4339d8cf26bae9 and 30a9ba4a02399278e2682e9acdd45daeab82f038eb8b1cebb30763d5491082a4. They began at 34.63--34.73 MiB RSS, peaked at 43.16--43.23 MiB, and added 8.50--8.53 MiB. Evidence is retained under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44e_raw_shadow_replay_contract_20260810/. The aggregate JSON SHA-256 is 5d3516edcadea924c5481688666edddec2db7b2be61eaa5cf42f7c52cc278181, and its deterministic aggregate payload SHA-256 is 7550b7303021d8440b749dc2b19279f3cdb66c8c6dfade6f3d2a27a64c91c10e.

The sealed decision is AUTHORIZE_V44_STAGE4_RAW_SHADOW_REPLAY_EXECUTION. No raw bag was opened in this gate. The next action is exactly six report-only runs: two repetitions of each bound sequence. Every run must complete with a valid non-null state payload, preserve both v17 hashes, remain under RSS/RTF after every scan, and match its exact stream identities; each sequence's two state payload SHA-256 values must be identical. Any one failure rejects the route. Accuracy, ground truth, reference maps, fresh holdout, ROS publication, primary writeback, and promotion remain closed even if all six raw runs pass; opening an accuracy screen requires another separately sealed contract.

v44f fail-fast raw-shadow execution result (2026-08-10)

The first authorized v44e raw-shadow attempt was executed and failed the frozen resource gate. The ordered first run was navinst_indoor02, repetition 1. The exact v44e authorization chain passed, the NavINST bag SHA-256 was verified, and scan index 0 was recorded as the permitted unbracketed prefix drop. The mandatory post-scan runtime observation then caused the estimator to terminate with processing RTF capacity exceeded after scan, meaning the attempted cumulative RTF was strictly above the unchanged 0.85 ceiling.

The failed adapter report is retained at sota_v6_dev_v44e_raw_shadow_replay_contract_20260810/raw_replay/navinst_indoor02/run_01/run.json (SHA-256 1913305fe4b3a8ded99c62aef3026428443643ecaa281a41c7fe0de6a1bd0b4a), and its two-record canonical diagnostic is beside it (SHA-256 32a5764f1db2abdd4b46d2eefb48ce4e69a6481ec93433b035f7cf6ced247639). The adapter report payload SHA-256 is 1194b8b69b1a35183ba8f33046ec494df9660f6dacc3b76823040b4680a8396a. The core returned FAIL, zero active states, valid_shadow_result=false, and a null state payload hash. NavINST's protected v17 state and map hashes are identical before and after. Accuracy/reference-map inputs, ROS publication, primary writeback, and every valid-state output remained closed.

The v44e execution contract defines any single run failure as rejection. The remaining NavINST repetition and both Oxford/UrbanNav repetitions were therefore not started. The failed directory was retained without overwrite, and no threshold was relaxed. Since execution stopped before stream completion, the report has no completed full-stream inventory and makes no claim that the whole bag was consumed.

The exact offending RTF number was not persisted. The core checks RTF before replacing the scan diagnostic's placeholder runtime fields, so the dropped scan retains processing_rtf=0.0 even though the terminal reason proves an over-limit attempted value. This is an observability limitation, not a pass or an unknown gate result. It prevents phase attribution from this run and must not be interpreted as measured zero RTF.

The post-execution audit contract is configs/sota_v6/development/v44f_raw_shadow_replay_execution_audit.json (contract ID v44f-raw-shadow-replay-execution-audit-20260810, SHA-256 3da6877b93a386fe875bfefc3844d9105e1cb90ec5820ac86ae7364d97f330b7). The auditor is scripts/audit_v44f_raw_shadow_replay_execution.py (SHA-256 9e0292e0055b06663314c5f7fe04cd82af17e2ca203e56265fda913c21882675), and the readable result is docs/research/v44-raw-shadow-replay-execution-2026-08.md (SHA-256 9b2586b1bdcde9bf3bcb0ea52f99be8f89bdbff5e0882003592a81a50af14567).

All 18 evidence checks pass, the focused v44f suite passes (16 passed in 0.88s), and the complete v40--v44f regression passes (277 passed in 25.38s). Two standalone audits have identical deterministic report payload SHA-256 29ef8cef8237544110e776662fd3e73ddc07f7aedd751c7a581a9c35149711bd. Their report SHA-256 values are b0b96e016143d5f42d2661deb21994dea14889be21f489b87d643b20842dd323 and 3a5fe3b2f1bda2c45d9973782c7aa8d301cf2b2b95ddc5efea657710e040861f. The aggregate JSON SHA-256 is 9b4dbbaede2239ae619594b6c84ee0f33b7fc89e360bfac777a5c37c37a2e2eb, and its deterministic aggregate payload SHA-256 is f9a9b25e746f49f5b2dc9665dba1a8019c59bfb2b044897f0561750a1da31073. Evidence is under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44f_raw_shadow_replay_execution_audit_20260810/.

The sealed route decision is REJECT_V44_STAGE4_RAW_SHADOW_REPLAY_RESOURCE_GATE. Raw replay continuation, accuracy, ground truth, reference maps, fresh holdout, primary mutation, and promotion are unauthorized. The only next authorized work is to define a bounded failure-profile contract that records the attempted RSS/RTF and fixed decoder/reorder/core phase timings before invoking the same unchanged 0.85 gate. That diagnostic contract cannot run another full replay, tune from accuracy, or relax the resource threshold.

v44g definition-only failure-profile contract (2026-08-13)

The bounded failure-profile boundary is now defined and statically audited. The machine-readable contract is configs/sota_v6/development/v44g_raw_shadow_failure_profile_contract.json (SHA-256 b985b4454908619215d0f41c1c26845d244790f23b6a40d2dda9677e1935c281), the auditor is scripts/audit_v44g_raw_shadow_failure_profile.py (SHA-256 3d93a4ef49b568071bcac2df56458400dd9cf2cb2dce0d0569553227f04cb755), and the readable contract report is docs/research/v44-raw-shadow-failure-profile-contract-2026-08.md (SHA-256 ed7f82fe4444b807977cc30f204117a346624b42cdfa1f14008b524631d203ab).

The definition binds the completed v44f reject audit, the v44e execution contract, and the exact adapter/core hashes. Its target is only the already attempted navinst_indoor02, repetition 1, scan 0 failure. The required fixed diagnostic phases are decoder -> reorder -> core; all count toward the same v44e cumulative processing RTF. The profile must capture attempted RSS and processing RTF, sensor interval, phase timings, gate order, and terminal reason before comparing against the unchanged 330 MiB / 0.85 limits. A placeholder 0.0 RTF is explicitly invalid.

Two definition audits passed all 15 checks with identical deterministic report payload SHA-256 3a057ba8ad418a64a0639af04f046b98ec51fb8e12821cbcf28668113169d8a7. The retained aggregate is under /home/user/workspace/old_~2026/lidarslam_ws/sota_v6_dev_v44g_failure_profile_contract_20260813/, has file SHA-256 6964094a85c746bee7f4081ee3173d6db21d604fe6ea9ed518d6955f01405610, and deterministic aggregate payload SHA-256 83d4d0b742ce44c7149c9e703389b6b206a0affa075ac31b9b60b121372c92ac. The focused v44g suite passes (12 passed). The sealed decision is AUTHORIZE_V44G_FAILURE_PROFILE_CONTRACT_DEFINITION_ONLY: profile execution, raw replay continuation, accuracy, primary mutation, and threshold relaxation remain unauthorized.