Nason/ice-reference-database

ICE facility, office and jurisdiction references

Frozen Datapond snapshot: 4 data tables, 4,410 stored source rows, 26,226,688 bytes. The catalog identifier requires separate registry integration. Row totals mix table units and are not unique people, events or places.

Table	Rows	Unit	Time semantics
facilities	878	One detention-facility directory record; not a detention event or historical facility panel	reference snapshot, not historical panel
offices	275	One office directory record across ERO, HSI and other listed agencies; not an ERO-only jurisdiction	reference snapshot, not historical panel
aor	26	One ERO area-of-responsibility record; 25 geographic areas plus one nongeographic HQ placeholder	reference snapshot, not historical panel
county_aor	3,231	One county-to-ERO-area assignment record; 3,230 county records plus one nongeographic HQ placeholder	reference snapshot, not historical panel

Source, scope and limitations

Facility scope starts with locations used since January 2025 plus newly listed sites; it is not a complete historical directory. Source observation vintages vary. The facility repository commit is October 20, 2026, and the office repository commit is April 21, 2026. Neither is a universal observation date.

The facility table preserves one NULL detention_facility_code and 75 missing coordinate pairs. Use _source_row as a pinned-file row reference, not a cross-release entity key. Preserve leading zeros in ZIP/FIPS/GEOID. Office fields are office_type, latitude and longitude; six obsolete population-summary fields from older facility documentation are absent from this payload.

AOR and county-AOR geometry is retained as WKB with NAD83 EPSG:4269 metadata, despite older codebook prose naming EPSG:4326. No coordinates are transformed or relabelled. Each geography table retains its nongeographic HQ placeholder; exclude it explicitly when analyzing geographic coverage. Do not silently drop source rows merely because a county or office name is NULL.

The office directory includes ERO, HSI and other agencies. Its AOR geography tables describe ERO responsibility only, not HSI territories or a historical boundary panel. The source office lists are described as September 2025 and county geometry derives from Census 2024 boundaries. Harlingen's FY2022 split and other source boundary caveats matter for time-specific joins. Facility codes need release/year coverage checks before LEFT JOINs to detention events. Office-name variants require an explicit reviewed crosswalk; no fuzzy link or legal-venue inference is supplied.

Every source byte hash, immutable URL and repository commit is in source-lock.json. Every row is retained; _source_row is added as an ordinal within the pinned member. Geometry, when present, is stored as WKB in geometry_wkb, with all original metadata in _geometry_metadata. _provenance, _metadata and _columns document the build. DICTIONARY.md records types, exact NULL counts, units and time semantics. Original source flags and text are not coerced into new numeric categories.

Publisher: Deportation Data Project. Data license: CC0-1.0, as labelled by the publisher; see the retained license notes. Cite government data collated by the Deportation Data Project and the pinned source commit. Repository commit dates do not promise current observations or automatic refreshes.

Local queries

import duckdb
con = duckdb.connect('ice-reference.duckdb', read_only=True,
                     config={'memory_limit':'128MiB','threads':'1'})
try:
    print(con.sql('SELECT state, count(*) AS facilities FROM facilities GROUP BY state ORDER BY facilities DESC, state LIMIT 5').fetchall())
finally:
    con.close()

library(DBI)
con <- dbConnect(duckdb::duckdb(), dbdir='ice-reference.duckdb', read_only=TRUE)
tryCatch({
  dbExecute(con, "SET memory_limit='128MiB'")
  dbExecute(con, 'SET threads=1')
  print(dbGetQuery(con, "SELECT state, count(*) AS facilities FROM facilities GROUP BY state ORDER BY facilities DESC, state LIMIT 5"))
}, finally=dbDisconnect(con, shutdown=TRUE))

Public datapond identifier ice-reference requires a future approved catalog release. The smoke scripts use an isolated localhost registry, never a global registry edit.

Reproduce and refresh

Use Python 3.11 or newer in a dedicated environment. Install the exact tested dependencies with python -m pip install -r requirements.txt (or use the pinned uv source in pyproject.toml), then:

python acquire_sources.py
python build_database.py --validate-only --report fresh-validation.json
python build_database.py --output fresh-ice-reference.duckdb

Downloads are pinned, bounded and SHA-verified; existing source files are checked and reused. The importer preserves all source fields and compares every stored row with the input using symmetric EXCEPT ALL. It uses one thread, 128 MiB DuckDB memory, no spill, a 128 MiB working-file cap and a 512 MiB physical free-space reserve. A watchdog interrupts work when a budget is crossed. Builds are private until complete; hard-link publication refuses an existing final database, manifest or orphan WAL path. Synchronous checks enforce the working-file cap and physical reserve after close and before publication; hash checks remain enabled under optimized Python. Failed/private files remain for inspection, with no cleanup of unrelated work. Peak process memory can exceed DuckDB's buffer limit.

To refresh, review the publisher's changed files and definitions, update pins and expected schemas/counts, run full source checks and Python/R smoke tests, and audit both data and documentation. No scheduled or unattended refresh is configured.

Source documentation: facilities and offices/AORs. Actual pinned schemas and CRS override the older codebook discrepancies described above. Office names alone are not unique; the current (agency, office_type, office_name, address_full) combination is unique. Each geography table has one empty MULTIPOLYGON for HQ and no SQL-NULL geometry.

Validation commands: python -m unittest -v test_safety.py and python run_smoke.py --python /path/to/python --rscript /path/to/Rscript --r-library /path/to/R/library. The latter needs the released Python and R datapond packages and an already installed HTTPFS extension. It uses only localhost and this database. See VALIDATION.md for the executed versions and results.

Snapshot files

Download the database, browse its files, or inspect the reproducible source. Verify the snapshot SHA-256 c0c6f79bea8ad54e84c29103a7d854b68a381d931e104e2a3ccc03c1c80d70a5 (26,226,688 bytes). The registry PR pins immutable Hugging Face revisions; a later main may refer to a newer snapshot.

Once the separate catalog entry is merged, use the released clients:

import datapond
con = datapond.connect("ice-reference")
try:
    print(con.sql("SELECT state, count(*) AS facilities FROM facilities GROUP BY state ORDER BY facilities DESC, state LIMIT 5").fetchall())
finally:
    con.close()

library(datapond)
con <- dp_connect("ice-reference")
tryCatch(print(DBI::dbGetQuery(con, "SELECT state, count(*) AS facilities FROM facilities GROUP BY state ORDER BY facilities DESC, state LIMIT 5")), finally=DBI::dbDisconnect(con, shutdown=TRUE))

Until catalog integration, the local Python/R examples above work with the downloaded snapshot. The smoke scripts exercise both released clients with an isolated registry and this same database.