MID-360 onboarding fixture pilot — 2026-08-10

Result

A provenance-pinned 50-second derivative of the public MID-360 bag is a viable onboarding candidate. Two clean builds from commit 0f91452c505e03fc810c79a0b6e602101ffefce7 produced byte-identical ZIPs and byte-identical manifests. An extracted candidate then completed the installed golden-path map route with a schema-valid seven-of-seven PASS receipt.

This is a local candidate, not a published asset, a cold onboarding trial, or a replacement for the full 277-second real-data gate. The default demo and release workflow are unchanged.

Candidate identity

The source is Driving SLAM Test with Livox MID360 by Kenji Koide, DOI 10.5281/zenodo.14841855, licensed under CC BY 4.0. The candidate ZIP contains ATTRIBUTION.md with the creator, source, DOI, license link, change notice, and no-endorsement notice.

Identity	Full source	50-second candidate
Requested slice window	not applicable	50.000000000 s
Recorded / selected-message duration	277.166836670 s	49.999532958 s
LiDAR messages	2,772	500
IMU messages	55,435	10,001
Archive / ZIP bytes	517,088,133	98,873,952
rosbag2 sqlite bytes	1,468,932,096	264,880,128

The candidate is 418,214,181 bytes, or 80.879%, smaller than the current source download. It passes the explicit 100,000,000-byte gate with a 1,126,048-byte margin.

Exact candidate identities are retained in the schema-valid build manifest:

ZIP SHA-256: 20e5151728522877bff75021a473e91c5ae900448fa9e6977bf88653fa464bd3;
build-manifest SHA-256: 60c37f5c7efa7d61ca20f21803fa11b02add4bad047ae99d277e9e6811fbbb6e;
sqlite SHA-256: 0a38fbccdb9c135b9a8e72141515b87ff64172a3b80a2d27a260ecd30683b5dd;
metadata SHA-256: d866804b49e8208dfcd70c2f91584976df0ca3d918c910d2ae20786cd1fa5b18;
attribution SHA-256: d363a9e9cf529876e38fa4b705e3b323cd40b0d642c02cf9ef5a003c3061aee8.

Reproducibility and provenance gates

The builder verified the exact 517,088,133-byte source archive by MD5 and SHA-256, then verified the original metadata.yaml and 1,468,932,096-byte sqlite file by SHA-256. It re-hashed the source bag after slicing to detect a change during generation.

The original Humble rosbag does not embed message definitions. The slicer now permits that case only when the generated definition and RIHS01 hash are identical in the maintained Humble and Jazzy typestores. An unverifiable custom type, a source-hash mismatch, a source or storage symlink, a root-escaping storage path, an existing output, or an oversized ZIP fails closed.

Both clean builds used:

Component	Value
Git revision	`0f91452c505e03fc810c79a0b6e602101ffefce7`
Git dirty	`false`
Python	`3.12.3`
rosbags	`0.11.0`
sqlite	`3.45.1`
zlib compile / runtime	`1.3` / `1.3`
ZIP recipe	DEFLATE level 9, sorted members, mode `0644`, fixed 1980 timestamp

The two ZIP byte streams and two JSON manifest byte streams matched exactly. The manifest intentionally reports map_validation_status: NOT_RUN: it is the immutable output of the build step. Map proof is retained separately so validation cannot mutate the artifact identity after generation.

Installed product map proof

The ZIP was integrity-checked, safely extracted, and passed to the installed lidarslam-map run route with profile rko_lio_graph_mid360_preset. The installed product build information was clean and identified the same commit as the generator.

Check	Result
Product exit / manifest / lifecycle	`0` / `succeeded` / `complete`
Diagnosis / Autoware verifier	`success` / `PASS`
Receipt	`PASS`, exactly 7 / 7 checks
Product execution time	15.043069 s
Raw / corrected poses	48 / 87
Pointcloud tiles	86
Pointcloud tile bytes	3,990,757
Combined `map.pcd` bytes	3,970,854
Output file / allocated bytes	8,249,858 / 8,585,216
Final `.partial` output	absent

The reviewed, geometry-free first-map receipt contains no local path or exact command. Its SHA-256 is 86d2b5d2aa493cbb6ecc6efd88095a591f247bcea4bc171c68093cf165cc0754. The receipt binds these private run artifacts without publishing them:

run manifest: 3875f246c2379cdec36bffefea365417cd92564678a7526e6f46359f8153aa17;
diagnosis: 890b3be607d0b2f88b03a10e8bba410cc8963ea9e431b5dd96896cdf154efdd5;
verifier log: 63053526d4a3686219461a2c2c7277890940bf8522d44dfd6f2d82748de1dfea.

The build manifest and receipt were validated against their Draft 7 schemas. review_before_publishing: true and review_before_sharing: true are mandatory review requirements, not automatic clearance. The two bounded JSON files were reviewed before being added to Git. The ZIP, sqlite bag, maps, trajectories, and private logs were not added.

Activation impact and limits

Replacing only the 517,088,133-byte dataset transfer in the existing Docker machine probes gives an arithmetic estimate of 1,352,422,163 bytes for Humble and 1,488,595,341 bytes for Jazzy. Those would be reductions of 23.619% and 21.933%, respectively. They are not measurements: image layers, registry compression, cache state, and network conditions were not rerun or controlled.

The estimate also shows that the image side remains larger than the candidate dataset. A smaller bag alone cannot reach the stretch target of less than 1 GB cold workflow RX. The next independent intervention is a multi-stage/runtime image experiment with at least 25% image-size reduction as its promotion gate.

This local run did not measure a clean download, cold wall time, human active time, command count, peak disk, Humble runtime, or an external user's first success. It therefore cannot be promoted into the comparable onboarding matrix. It also does not make an accuracy claim.

The shortened run produces 48 raw poses, 87 corrected poses, 86 tiles, and 3,990,757 tiled pointcloud bytes. The full gate requires 2,500 raw poses, 500 corrected poses, 300 tiles, and 50,000,000 pointcloud bytes. Keeping the full 277-second scheduled proof is therefore a hard requirement, not a preference.

Reproduce the candidate locally

Use an exact clean checkout of the generator commit and the source identities in configs/real_data_e2e/driving_slam_mid360_v1.json. Keep all geometry outside the checkout:

python3 scripts/build_mid360_onboarding_fixture.py \
  --source-archive "$SOURCE_ARCHIVE" \
  --source-bag "$SOURCE_BAG" \
  --output-dir "$BUILD_A"

python3 scripts/build_mid360_onboarding_fixture.py \
  --source-archive "$SOURCE_ARCHIVE" \
  --source-bag "$SOURCE_BAG" \
  --output-dir "$BUILD_B"

cmp "$BUILD_A/mid360_onboarding_50s_v1.zip" \
  "$BUILD_B/mid360_onboarding_50s_v1.zip"
cmp "$BUILD_A/mid360_onboarding_50s_v1.manifest.json" \
  "$BUILD_B/mid360_onboarding_50s_v1.manifest.json"

Do not use --allow-dirty for a publication candidate. That option exists only for explicitly non-publishable local experiments.

Publication review follow-up — 2026-08-11

The separate publication review now revalidates the exact ZIP, all three member byte streams, attribution, map receipt, and two clean rebuilds through a fail-closed machine-readable gate. The local packet reports LOCAL_ARTIFACT_PASS, but the overall decision is AWAITING_PUBLICATION_DECISION: the generator and reviewed gate revisions are not yet publicly resolvable, no host has been selected, and no explicit upload authorization exists. No fixture or source branch was published by the review.

Promotion decision

The candidate passes the local build, size, provenance, attribution, reproducibility, map, verifier, and receipt gates. It is not yet approved for upload or for changing the default demo. Promotion remains gated on:

an explicit source-push, host, and release/upload decision after the local publication review;
hosting the geometry-bearing ZIP outside Git with the exact manifest and checksum;
a REMOTE_ARTIFACT_PASS from the host-aware publication audit;
activation of the reviewed public URL through the tested resumable, checksum-pinned acquisition path, retaining the full proof route as fallback;
clean dedicated-VM Humble/Jazzy Docker and source trials with cold RX, wall time, active time, command count, and isolated peak disk;
the external first-map cohort; and
a separate runtime-image slimming experiment.

Until those gates pass, the full public bag remains the default proof route and broad promotion remains paused.