Studies and archives¶
The LUCiD research record lives in versioned release assets, not in the tree.
Each asset is one tarball whose extracted root carries an INDEX.md (what the
material was for, the commit it ran against, how to run it, and every known
defect), a PROVENANCE.txt and a MANIFEST.sha256. Each is built from one
manifest under tools/assets/; see "Rebuilding an asset" below.
The two the installer reads, lucid-data-v1 and lucid-paper-results-v1, are GitHub
releases of their own; every other asset is a file of the
lucid-archives release.
Every asset is under CC BY 4.0 (its INDEX.md says so); the code in this
repository is under the terms of LICENSE.
No path inside an asset may be cited from lucid/, analysis/ or tests/ —
a citation names the asset and this page instead. If you came here from a
code comment or a docstring, the row below is where that material went.
Every row is what tools/make_asset.sh --row <manifest> prints
for one manifest under tools/assets/ (its summary:, size: and commit:),
so a change goes into the manifest and the row is regenerated.
| asset | what it was for | size | ran against |
|---|---|---|---|
| detector-geometry-sources | the WCSim geofile dumps behind the tracked npz | 5.5 MB | eaf2e89f |
| legacy-io-archive | read HDF5 files written before format 6 | 176 KB | eaf2e89f |
| lucid-data-v1 | the first-run bundle: the four SIREN emitters' weights + 5-event 1 GeV and 6-event 1.5 GeV GEANT4 muon slices and the 12 overlap tables the shipped code loads | 142.2 MB (141.7 MB tarball) | 7e3452a5 |
| lucid-paper-results-v1 | the paper figures' inputs: the nrays, energy and geometry tracking scans and the cached data of the other figures | 299 MB (84.2 MB tarball) | 7a913a09 |
| studies-archive-absorption-siren | the student's SIREN-as-absorption-field project and its merged-quality twin | 6.3 MB | 97c9076a |
| studies-archive-adam-optimizer | the retired 5-stage Adam recon pipeline, its 3-D displays and its composite losses | 36 KB + notebooks | eaf2e89f |
| studies-archive-calib-campaign-2026-09 | the 2026-09 calibration estimator campaign + the schur_gn arm it is the only user of |
195 KB | 71f00e49 |
| studies-archive-calib-closure-2026-10 | the calibration-closure figure and its plot module, as committed on paper-rerun-qe | 21 KB | 2b03c27b |
| studies-archive-geometry-v2-2026-10 | the V2/CSG research record -- the gradient ladder, the benchmarks, the NaN audit, the diaries | 4.6 MB in 520 files (1.7 MB gzipped) | 85e98bd1 |
| studies-archive-gifs-2026 | the three event/optimisation animations | 13.2 MB | eaf2e89f |
| studies-archive-icecube | the IceCube/ppc measurement programme + the legacy string engine + the sizing maths |
3.7 MB (plus 144 MB of run outputs on /sdf/data, not in the tarball) | 958807f6 |
| studies-archive-lartpc | the LArTPC optical package and its calibration arc | 2.9 MB | 85545f72 |
| studies-archive-method-ab-2026 | the expected-vs-sampled A/B driver | 6 KB | eaf2e89f |
| studies-archive-notebooks-2026 | the S-class half of good_notebooks/ + the tutorials + s3df_jobs/ |
1.5 MB | eaf2e89f |
| studies-archive-notebooks-legacy-2026 | the 24 pre-rename notebooks/ + wcte_bldg157.geo |
18.3 MB | c799b404 |
| studies-archive-string-telescope-2026 | the string event producers + the standalone three.js viewer | 56 KB | eaf2e89f |
| studies-archive-tracking-diagnostics-2026-07 | the 2026-07 track-fit diagnostics and the TrackingPipeline branches they drove | 225 KB | 252f89b8 |
lucid.fitting.schur_gn itself is still in the library; the calibration-campaign
asset carries the campaign that was its only user.
Large run artefacts do not travel in an asset: 144 MB of IceCube run
outputs, the 282 MB CRB-appendix data and 1.6 MB of pool_bench/out/ run
artefacts are kept outside the repository, and each asset's INDEX.md records
where.
Rebuilding an asset¶
Each asset is described by one manifest under tools/assets/, and
tools/make_asset.sh builds it:
LUCID_ASSET_OUT=<somewhere with room> tools/make_asset.sh tools/assets/legacy-io-archive.manifest
The script refuses to build an asset whose manifest carries no INDEX.md
body, and refuses one whose commit: does not resolve — a study tarball with
no provenance is worse than a deleted directory, because it looks runnable.
The script touches no remote. tools/publish_assets.sh <dir> then publishes
every built asset in <dir> to the release its manifest's release: names (the
asset's own name when there is none): it checks each tarball against its
.sha256 and the two installer assets against data/MANIFEST.sha256 and
analysis/paper/RESULTS.sha256, prints what it would do (a dry run), and with
--go creates the missing releases on main and attaches the missing files.
A published file is never replaced; a changed asset is a new tag.
lucid-data-v1 and lucid-paper-results-v1 carry material that is on no git
ref: the trained SIREN weights, the PhotonSim ROOT slices, the git-ignored
overlap tables and the paper's results. Their manifests name those parts
relative to the directory in $LUCID_ASSET_SRC, so no tracked file carries
anyone's absolute path.