Validation of the full USCIS snapshot
Primary artifact: uscis.duckdb, 391,131,136 bytes, SHA256 e0957fb5726b990f694b08c86c0bf8f1d785a5b41d5ad5079ecaac2dfe5a2f9f. This is the metadata-corrected full owner-format build, not the preserved pilot or earlier two-table minimal candidate.
The unchanged importer at commit 2293a11b56aa3b1e66d03f7a8a0f9ceea9673ea8 ran against both complete verified Parquets. It compared every original typed value and added ordinal for all 11,206,545 rows in 4,096-row batches, matched 31/30 original fields and all NULL baselines, and retained every duplicate flag and source anomaly. The manifest records both application tables and their source-value checks. Independent SQL/client checks establish 2 metadata records, 63 field definitions and 2 full-source provenance rows with is_pilot=false; those metadata counts are not separate manifest fields.
Nine importer/acquisition regressions passed locally and in the cloud. A separate Codex audit verified complete source/database hashes, schemas, all metadata definitions and NULL counts, full ordinal consistency and duplicate-flag combinations, 104 targeted complete source tuples, and six positive/negative comparator regressions. It reviewed the full comparison receipt; its 104 targeted rows are not described as a second independent full-row scan. The original independent minimal candidate also passed complete bidirectional source multiset equality and remains separate.
Fresh desktop client checks on November 3 UTC passed on the corrected database SHA256 e0957fb5726b990f694b08c86c0bf8f1d785a5b41d5ad5079ecaac2dfe5a2f9f through Python Datapond 0.1.5/DuckDB 1.5.6 and R Datapond 0.1.0/DuckDB 1.5.5, using a private loopback registry and no pilot flag. The updated Python smoke SHA256 is 1b6b25deb9c2d95a3c1c24245838ff44822b5ec863a0dc0cdb6256bc8bd24453; the R smoke SHA256 is fed52c3839b63417eab7de27e3d97ed9367bf6bc0535ba696a7b4e7f817e6e6c. Both compare all 63 descriptions and join hints with the corrected definitions; R also requires exactly two provenance rows. The nine updated importer regressions passed in 3.136 seconds from a fresh export copy. Both clients returned 4,526,373 DACA and 6,680,172 TPS rows, correct metadata and scope, and the README aggregate results. All 113 audit-packet file hashes and the full database hash were verified before this check. The original pilot evidence was preserved. These are checks of the full candidate; public URL/catalog checks follow publication.
Full measured anomalies are documented in README/FIELD-GUIDE and the manifest. No filtering, deduplication, missing-value coercion or date correction was performed by this importer. Upstream publisher transformations are documented separately. Raw input hashes are checked before the build, and output DB/WAL/manifest collisions fail closed. Tests use explicit exceptions under optimized Python and owned tiny fixtures; the full production resource limits remain unchanged.
The measured full build used one thread and the documented 2 GiB memory / 4 GiB spill settings under an additional external cloud resource guard. It completed in about 99 seconds with sampled peak RSS 1,441,271,808 bytes; no guard tripped. These are observed samples, not hard kernel limits. Independent full-data Codex clearance and released-client checks are complete. The first actual Claude source review found description issues in the DACA filled-form rule, TPS country fallback and manifest duplicate-count wording; the corrected source review confirms those three issues resolved. The updated full client checks and documentation clarifications address its remaining execution/status notes; release receipts record the separate final review outcome. Public-URL checks follow publication.
The historical manifest key null_duplicate_flags measures NULL duplicate_drop_row only. Unknown duplicate_likely is recorded separately in each table's null_counts (19,135/612), and is not zero. The source lock is retained byte-for-byte for provenance. Its pre-acquisition schema_evidence phrase "full hash remains to be checked on acquisition" is superseded by the manifest's completed full_source_hashes_verified: true.
A metadata-only correction after the first Claude review updated exactly 11 _columns rows (seven descriptions and four join hints) across six field names. The original full artifact SHA256 7431fdcb3be8ba87eee249ff563c70d0c683fd8cd1314a8b6b409fa90730e641 remains preserved in the build archive. Before/after canonical typed-row SHA256 scans covered all 11,206,545 application records and matched exactly, including schemas and contiguous source ordinals. _metadata, _provenance and every other _columns cell are unchanged. The revised manifest records both artifact identities and the correction; source data and source-lock bytes are unchanged. This is a corrected metadata release, not a new USCIS observation period.