[corlin]

What

Stage 1 of ADR 0005: the candidate set of US/EP patents that the stage-2 legal-status check (#11, #14) runs over. The design is in ADR 0007, and every scope parameter there is backed by a measurement.

Decision	Basis (measured 2026-10-04)
US B1/B2/E1, EP B1/B2/B3, granted, H01M/H02J by CPC OR IPC	Reissues (E1, 201) replace the original patent. CPC-only filtering misses IPC-only documents (see the spike report)
US: filed ≥ as_of − 20y − 5y	Patent term adjustment is common: 48/59 sampled patents had it, up to 1,220 days. Stage 2 computes the exact expiry
EP: filed ≥ as_of − 20y, no margin	The EPC has no term adjustment. BigQuery's EP filing date matched EPO's for 57/57 samples
Cost of the margin	+11% rows (US 140,344 → 156,306; EP 48,173 → 53,452)

ADR 0007 also defines a calibration rule for the US margin: after the first full legal-status run, check whether any US patent whose term started 20–25 years ago is still presumed in force.

Storage (ADR 0006 pattern)

• Global, append-only tables patent_corpus_loads and patent_corpus_members. Worker: SELECT+INSERT; app: SELECT; no UPDATE/DELETE; no tenant columns.
• Each load records its scope and cutoffs, the exact query and its SHA-256, bytes billed, the source table's last-modified time and the row count, so an FTO analysis can name the corpus it used.
• A load is written atomically (tested: a failure part-way leaves nothing). The row count is known up front because the tables can't be updated.
• Inventors are not loaded. Assignee names are, since they're needed for ownership; some may be individuals' names, which are public bibliographic data.

Loader

scripts/load-patent-corpus.py is run by an operator; it's not part of any image.

• It dry-runs by default (no cost), and the load query dry-runs at 58.88 GB.
• --execute requires --max-gb, and the script refuses if the scan exceeds it (both guards tested: exit 2 and exit 1).
• The BigQuery client is an optional ingest extra. It's absent from the API image (verified on the rebuilt image).
• Credentials come from the operator's Google Cloud ADC; nothing is stored in the repo.

Not done yet

• No data has been loaded. The first full load runs after this is reviewed (≈59 GB scan; this month's usage so far is about 333 GB of the 1 TB free tier).
• Legal-status support for US reissues (RE-prefixed numbers) isn't there yet; they're flagged.
• Next steps: batch legal-status job, then the watch tier (published applications), then margin calibration.

Validation

• ./scripts/test-postgres.sh: 103 passed (99 + 4 new), including the downgrade/upgrade cycle through 0019
• .venv/bin/pytest apps/api/tests/unit/: 382 passed (6 new)
• pnpm test:web: 62 passed; pnpm build:web: ok
• verify-source-lock, verify-scaffold, verify-p0-02, verify-p0-e2e, docker compose config: ok
• Loader dry run against BigQuery: 58.88 GB; guard refusals verified

🤖 Generated with Claude Code