Nason/ice-287g-database

ICE 287(g) agreements and agencies in DuckDB

This frozen Datapond snapshot packages the Deportation Data Project's agreement history and agency reference table. It adds policy participation, timing and jurisdiction information to the existing enforcement-event database. The catalog identifier requires the separate registry integration described below.

Table	Unit	Rows
agreements	One agreement, including renewed and superseded agreements	3,201
agencies	One participating agency, with agreement summaries	2,491

The 15,478,784-byte database contains _metadata, _columns, _provenance and _geometry_metadata. The total of 5,692 table rows mixes units; it is not a count of people or unique participating places. There are 2,725 active, 273 removed and 203 superseded agreements in this snapshot.

Version and provenance

Both inputs are pinned to publisher commit 5f4b3630a0eedc5ecaff8e8c065d1ce1438e44b7 (November 1, 2026). The latest observed listing is October 28; signing dates span July 2, 2002 through October 27, 2026. These source observation dates are distinct from the November 2 local build date. Complete file sizes, SHA-256 values and immutable URLs are recorded in source-manifest.json. Both source hashes were verified.

Sources: publisher catalog, codebook, and immutable source repository. The repository's CC0-1.0 data license is retained. Cite the Deportation Data Project, its source commit and Labia and Blair's 2026 codebook.

Use and interpretation

Join the tables using agency_id within this snapshot. agreement_id is unique in agreements; agency_id is unique in agencies. An agency may have several models or renewals, with overlapping times and jurisdictions. No person linkage to EOIR, USCIS or the ICE enforcement records is asserted.

The dates last_appeared_date and removed_by_date bound a termination interval; neither is an exact termination date. Historical intervals may be wide. Current status describes this source snapshot, not today's status. Agency geography is taken from its active or most recently listed agreement, not a historical panel. All 2,725 Active agreements have no removal bound. One Superseded agreement also has no removed_by_date; an unknown bound must not be converted into an exact termination date or interpreted as still Active.

The records represent formal agreement scope, not enforcement intensity, arrests or informal cooperation. Facility locations are approximations, and historical facility coverage is incomplete. Geography is missing for some agreements. In this release, 406 geometry values are empty geometries, even though none are SQL NULL; testing only geometry_wkb IS NULL will miss them. Another 125 agreement rows have no county FIPS and 141 have no ORI9. Some county/ORI fields contain semicolon-separated associations in county_fips (96 rows) and ORI9 (5 rows): split them into a bridge before joining, preserving leading zeros. Check geographic vintage and jurisdiction before joins. Agency last_removed_by_date is deliberately NULL while any agreement is Active; other missing aggregate dates also do not establish that an agency was never listed.

All source fields and rows are retained. The sole storage conversion is upstream geometry to geometry_wkb, a standard WKB BLOB produced by ST_AsWKB; no coordinates are transformed. This source uses WGS84/EPSG:4326; geometry coordinates follow the publisher's WKB, and original Arrow/CRS metadata is preserved in _geometry_metadata. DICTIONARY.md lists every field, its type, exact missing count and definition.

Local queries

Python (DuckDB 1.5 or later):

import duckdb
con = duckdb.connect('ice-287g.duckdb', read_only=True,
                     config={'memory_limit': '128MiB', 'threads': '1'})
try:
    print(con.sql('SELECT status, count(*) AS agreements FROM agreements GROUP BY status ORDER BY status').fetchall())
finally:
    con.close()

R:

library(DBI)
con <- dbConnect(duckdb::duckdb(), dbdir='ice-287g.duckdb', read_only=TRUE)
tryCatch({
  dbExecute(con, "SET memory_limit='128MiB'")
  dbExecute(con, 'SET threads=1')
  print(dbGetQuery(con, 'SELECT status, count(*) AS agreements FROM agreements GROUP BY status ORDER BY status'))
}, finally=dbDisconnect(con, shutdown=TRUE))

Public datapond.connect('ice-287g') and dp_connect('ice-287g') require a future catalog release. The separate smoke runner tests the released clients against a temporary localhost registry and this exact artifact.

Rebuild and checks

The build uses Python 3.11+, DuckDB 1.5.5 and the exact datapond-build dependency revision in pyproject.toml. With those dependencies installed:

python -m unittest -v test_build
python build_database.py --output fresh-ice-287g.duckdb

Missing inputs download from pinned URLs, with a 16 MiB per-file cap and SHA-256 verification before parsing. Existing files are verified and reused. The builder requires 1 GiB free on the actual filesystem, uses one thread, limits DuckDB memory to 128 MiB and disables spill. It builds in a private directory, then publishes locally with an atomic hard link that refuses an existing target. The filesystem must support hard links; private intermediates remain for inspection and share storage with the final file. Failed files are also preserved. Peak process memory is not identical to DuckDB's limit.

Validation covers exact source equality, unique nonmissing keys, agency links, summary counts, status values, termination interval ordering and WKB round trips. Source updates may be frequent, but this is a manual frozen snapshot: refresh by reviewing a new publisher commit, updating both source pins, auditing schema and definitions, and repeating Python/R tests. No automatic update schedule is implied.

Snapshot files

Download the database, browse its files, or inspect the reproducible source. Verify the snapshot SHA-256 78ce2f63df3c0f37877085d74eee52fdc167b5eec5ccd39f48a75eb1381426f0 (15,478,784 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-287g")
try:
    print(con.sql("SELECT status,count(*) AS agreements FROM agreements GROUP BY status ORDER BY status").fetchall())
finally:
    con.close()

library(datapond)
con <- dp_connect("ice-287g")
tryCatch(print(DBI::dbGetQuery(con, "SELECT status,count(*) AS agreements FROM agreements GROUP BY status ORDER BY status")), 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.