EPA FRS (ECHO Exporter) - landscape for Facility Foundry v1 onboarding

Phase 1 desk landscape, written 2026-05-06. Reference document for the EPA FRS fetcher built against the v1 multi-dimensional Classification schema. Cross-references: docs/adr/007-multi-dimensional-classification.md, docs/DATASOURCES.md, src/strata/fetchers/epa_frs/, ECHO Exporter schema reference at https://echo.epa.gov/tools/data-downloads/echo-exporter-data-files.

Live-data observations (verified 2026-05-06 against the 2026-05-03 ECHO Exporter):

1. Bulk URL is https://echo.epa.gov/files/echodownloads/echo_exporter.zip (~442 MB compressed, ~2.0 GB CSV; weekly refresh). The national_combined.zip from ordsext.epa.gov is the alternative raw FRS dump but is 3× larger and lacks ECHO's pre-joined cross-program identifier columns. We consume ECHO Exporter.
2. 3.14 M rows total. Aggressive filtering required. Most are drinking water systems, retail gas stations, dry cleaners, and similar non-FF-domain entries. The fetcher admits rows passing four cumulative filters: (a) FAC_ACTIVE_FLAG = 'Y' (1.37 M), (b) lat + lon present (3.08 M), (c) FAC_ACCURACY_METERS ≤ 1000 drops state/zip/county centroids (887 K), (d) industrial NAICS prefix in {21 mining, 22 utilities, 31-33 manufacturing} (~148 K). Final admitted ≈ 148,000 facilities.
3. Coordinate quality bimodal. 50th-percentile accuracy is 150 m (geocoded street address); 75th percentile is 17,477 m (state/zip/county centroid). The accuracy filter is load-bearing. Facilities with FAC_COLLECTION_METHOD ∈ {ADDRESS MATCHING-HOUSE NUMBER, GDT-ADDRESS MATCHING, GPS} and accuracy ≤ 1 km are HIGH-confidence; ≤ 100 m additionally; coarser drops.
4. Cross-program identifiers are the key value-add. Each FRS facility carries up to six program-tier identifiers: AIR_IDS (Clean Air Act / AIRS), NPDES_IDS (NPDES water permits), RCRA_IDS (hazardous waste handlers), TRI_IDS (Toxic Release Inventory), GHG_IDS (Greenhouse Gas Reporting), SDWA_IDS (Safe Drinking Water Act). FF emits one FacilityIdentifier per program ID, anchoring future cross-source matching against program-specific datasets.

This is the reference for the v1 EPA FRS fetcher: slim Facility + multi-row Classification + FacilityIdentifier (per-program) + FacilityAttribute (compliance metadata).

---

1. Source overview

Publisher. US Environmental Protection Agency, Office of Enforcement and Compliance Assurance (OECA), via the ECHO program. ECHO ("Enforcement and Compliance History Online") is the consumer surface over the Facility Registry Service (FRS) - EPA's master registry of ~3 M facilities subject to environmental regulation.

License. US federal-government work - public domain (17 U.S.C. § 105). No registration, attribution, or fees. Cite per the canonical citation in release.py.

Coverage. All US facilities regulated by one or more EPA programs: Clean Air Act stationary sources, NPDES water dischargers, RCRA hazardous waste handlers, TRI reporters, GHGRP reporters, SDWA public water systems, federal facilities. Includes Indian Country, US-Mexico border zones, Chesapeake Bay watershed.

Vintage. The ZIP Last-Modified HTTP header dates the snapshot; ECHO Exporter refreshes weekly.

---

2. The artefacts

Two artefact families. The first ships every facility row with pre-joined cross-program identifiers; the second carries the affiliation block ECHO Exporter omits.

• ECHO_EXPORTER.csv (inside echo_exporter.zip, ~442 MB) - 133 columns, ~3.14 M rows, ~2 GB. Header row + double-quoted comma-separated values (RFC 4180 compliant - unlike MSHA, EPA escapes internal quotes properly). The fetcher's first parse pass.
• <STATE>_ORGANIZATION_FILE.CSV (inside per-state state_combined_<state>.zip at ordsext.epa.gov, ~233 MB / state)
• 11 columns, ~1.07 M rows for CA. Header row, comma-separated. Carries REGISTRY_ID, AFFILIATION_TYPE (OWNER / OPERATOR / PARENT COMPANY / etc.), ORG_NAME, ORG_TYPE (PRIVATE / STATE / MUNICIPAL / FEDERAL / TRIBE / OTHER / UNKNOWN). Fetched per-state best-effort (single-state failure logs a warning and skips); the fetcher walks all 51 codes.

The ECHO schema is documented at https://echo.epa.gov/tools/data-downloads/echo-exporter-data-files. The state-combined index is at https://www.epa.gov/frs/epa-state-combined-csv-download-files.

---

3. Records emitted per row (admitted)

For each admitted row the parser emits:

• 1 Facility - canonical_name = FAC_NAME, coords from FAC_LAT / FAC_LONG, country_iso2 = "US", region = FAC_STATE, locality = FAC_COUNTY + FAC_CITY, status mapped from FAC_ACTIVE_FLAG. coordinate_confidence is HIGH when FAC_ACCURACY_METERS ≤ 100, MEDIUM ≤ 1000, else the row is dropped.
• 1+ FacilityIdentifier rows:
  • epa_registry_id (always present; the REGISTRY_ID column).
  • epa_air_id (one per ID in the comma-separated AIR_IDS).
  • epa_npdes_id (one per ID in NPDES_IDS).
  • epa_rcra_id (one per ID in RCRA_IDS).
  • epa_tri_id (one per ID in TRI_IDS).
  • epa_ghg_id (one per ID in GHG_IDS).
  • epa_sdwa_id (one per ID in SDWA_IDS).
• 1+ Classification rows:
  • industry: naics (rev 2022) - one row per code in FAC_NAICS_CODES, primary on the first code.
  • industry: isic (rev 5) - one row per the ff_naics_to_isic derivation.
  • process_role: ff_process_role - derived from NAICS sector (extraction for 21, beneficiation for some 21+331, fabrication/assembly for 33, etc.).
• 0+ FacilityAttribute rows:
  • coordinate_collection_method (FAC_COLLECTION_METHOD)
  • coordinate_accuracy_meters (FAC_ACCURACY_METERS)
  • coordinate_reference_point (FAC_REFERENCE_POINT)
  • is_federal_facility (FAC_FEDERAL_FLG)
  • is_indian_country (FAC_INDIAN_CNTRY_FLG)

After the ECHO row stream completes, the parser walks each <STATE>_ORGANIZATION_FILE.CSV and emits affiliation records for admitted facilities (the second pass keys on the REGISTRY_ID -> facility_id map built during the ECHO pass). For each admitted ORGANIZATION row:

• 0-1 Entity - entity_id = epa_frs_organization:<sha1_first16> keyed on the normalised ORG_NAME. The same corp operating five sites collapses to one Entity. Confidence is MEDIUM for structural-bucket rows (government / utility / tribal) and LOW otherwise. entity_type = state_vehicle for structural-bucket, corporation otherwise. jurisdiction = US-<state> from the per- state file the row came from.
• 0-1 OperatorReference - role mapped from AFFILIATION_TYPE: OWNER → owner, OPERATOR → operator, OWNER/OPERATOR → operator, PARENT COMPANY{,1,2} / PARENT OWNER → parent_guarantor, TRANSMISSION OR DISTRIBUTION SYSTEM OWNER / PART PARENT OWNER → owner. Confidence MEDIUM (EPA's regulatory intake is stronger than name-block synthesis but weaker than a CIK / LEI hard anchor, which EPA doesn't carry).

The admit filter excludes sole-prop / fictitious-name rows (~70 % of the 1.07 M CA OWNER / OPERATOR / PARENT rows). Keep iff ORG_NAME carries a corporate suffix (Inc / LLC / Corp / Co. / Ltd / LP / Holdings / Trust / Group / Industries / Partners / GmbH / SA / NV etc.) or ORG_TYPE is a structural government / utility / tribal bucket. ORG_TYPE = OTHER (762 K of CA's 1.07 M) is the unhelpful catch-all; the corp-suffix gate carries the load there.

Not emitted (out of FF Phase 1 scope):

• Compliance history columns (FAC_3YR_COMPLIANCE_HISTORY, per-program *_COMPLIANCE_STATUS, penalty amounts, inspection counts, etc.). These belong in a future EHS-attribute submodule.
• Person-affiliation rows where ORG_NAME is a natural person (sole-props, owner-operators, fictitious-business-name registrations without a corp suffix). The admit filter drops these because they don't bridge to any FF lineage source (BQ / GLEIF / CorpWatch all index legal entities, not natural persons). False-positive risk: fictitious-business-name LLPs without a suffix in the FRS row (e.g. "PALACE PLATING") get dropped too.
• LifecycleEvent - FRS doesn't track commissioning dates reliably. FAC_DATE_LAST_INSPECTION is enforcement metadata, not facility history.
• Contact / mailing-address / regulatory-contact affiliations from <STATE>_CONTACT_FILE.CSV. That file carries individual contacts (names + phones + emails) which don't fit the OperatorReference contract; would need a separate Contact record kind.

---

4. Filtering rationale

Why filter at all? 3.14 M rows × ~7 record types per row = ~22 M records - most of them irrelevant to FF's "industrial production facilities" remit.

The four filters:

1. FAC_ACTIVE_FLAG = 'Y'. Drops historical / decommissioned facilities. Keeps 1.37 M of 3.14 M.
2. Coords present. Both FAC_LAT and FAC_LONG non-null. Keeps 3.08 M.
3. FAC_ACCURACY_METERS ≤ 1000. Drops state-centroid (≈100 km), zip-centroid (≈10 km), county-centroid (≈30 km), and sentinel 999999999 rows. Keeps 887 K - the geocoded subset.
4. Industrial NAICS prefix. First two digits ∈ {21 mining/oil-gas, 22 utilities, 31/32/33 manufacturing}. Keeps ~148 K.

NAICS prefix filtering is the most opinionated choice. Specifically excludes:

• 11 Agriculture / Forestry / Fishing - out of FF's "industrial production" remit.
• 42-49 Wholesale / Retail / Transport - gas stations and warehouses dominate; not facility-foundational.
• 51-81 Information / FIRE / Education / Health / Other - almost entirely non-industrial.
• 92 Public Administration - government offices.

Mining, utilities, and manufacturing capture the FF-relevant industrial production footprint cleanly.

---

5. Coordinate confidence policy

This is stricter than MRDS (which admits everything regardless of coord quality) and stricter than MSHA (which has no accuracy field). The strictness is intentional: EPA FRS is exactly where coordinate quality matters most for cross-source matching, since synthesis uses coord distance as 40 % of the composite score.

---

6. Cross-source value

The expected wins from adding EPA FRS to the synthesis layer:

1. MSHA coord-drift triangulation. Both Mountain Pass and Bingham showed MSHA records with coords drifted 4-6 km from the actual ore body. EPA's geocoded coords are typically much closer to ground truth. If an EPA-MRDS pair high-matches AND an MSHA-EPA pair high-matches, the resulting connected component binds all three correctly even when MSHA-MRDS alone fails.
2. Manufacturing coverage. MSHA + MRDS are mining-only. EPA FRS adds ~47 K manufacturing (NAICS 31-33) and ~12 K utilities (22) facilities - a step toward FF's broader scope per docs/SCOPE.md.
3. Cross-program identifier anchors. A facility might be msha_mine_id:0402542 AND epa_registry_id:110000123 AND epa_tri_id:CAxxxx. Phase 2 cross-source matching against TRI data (in a future fetcher) will pivot on these IDs without re-resolving identity.

---

7. Affiliation onboarding (added 2026-05-10)

The per-state <STATE>_ORGANIZATION_FILE.CSV adds owner / operator / parent-company affiliations the ECHO Exporter row stream cannot. National 51-state verification on synth_884450a21531; the CA-only intermediate run was synth_299ca740a801 (kept here as a sanity checkpoint).

Facility coverage from 20.7 % → 33.0 % (+12.3 percentage points; +70,892 facilities now have an operator reference). EPA FRS dominates the OperatorReference universe: 297,213 of 366,607 total = 81 % of all FF operator references. The matcher's name-block pass synthesises 2,897 HIGH bridges from EPA FRS organisations, with the counterpart distribution:

Sample HIGH bridges (verbatim from the resolved tier):

• STAR CUTTER COMPANY (epa_frs, US-MI) ↔ STAR CUTTER CO. (GLEIF LEI 549300ELKX5LQ6SKLC44).
• WOLF FAMILY LLC (epa_frs, US-NY) ↔ WOLF FAMILY LLC (GLEIF LEI 549300A4LQ8CUSJHBS20).
• DYNOMAX INC (epa_frs, US-IL) ↔ DYNOMAX, INC. (GLEIF LEI 549300RYR9SD5E3XFL95) - comma normalisation handles the punct.

Largest CA operators by facility count (post-resolve):

† Name-fragmentation visible: ECHO ingests the same legal entity under multiple casings ("AND" vs "&", "COMPANY" vs "CO"). The synthesis name-block pass collapses these via same_entity_as at matcher time; the underlying Entity store keeps both as distinct nodes anchored on the verbatim ORG_NAME.

Filter rejection rate (national 5.81 M OWNER / OPERATOR / PARENT rows):

• ~3.81 M dropped at admit time - sole-prop / fictitious-business- name layer (no corporate suffix, ORG_TYPE = OTHER / UNKNOWN / PRIVATE). STEEL SERVICES CO. of LA's affiliations all fall here (BILL / CAROLINE / WILLIAM SALLENBACH).
• 1,700,945 admitted but the REGISTRY_ID didn't match an industrial- NAICS facility we admitted - gas stations, dry cleaners, retail outside the FF mining / utilities / manufacturing remit.
• 297,213 admitted and matched a facility - emitted as 72,164 Entity records (after dedup on normalised name) + 297,213 OperatorReference records.

Cache contract. State zips + extracted ORGANIZATION_FILE CSVs are cached at data/raw/epa_frs/<vintage>/state_files/. The fetcher's per-state loop short-circuits before download when the CSV is already on disk; re-running strata fetch epa_frs after the initial sweep is a no-op for the network. On-disk footprint of the state files: ~2.5 GB (zips + extracted CSVs).