USGS MRDS - landscape for Facility Foundry v1 onboarding

Phase 1 desk landscape, written 2026-05-05. Reference document for the MRDS fetcher rebuild against the v1 multi-dimensional Classification schema landed in commit ebb0962. Cross-references: docs/adr/007-multi-dimensional-classification.md, docs/DATASOURCES.md, src/strata/models/, strata's production fetcher at strata/src/strata/fetchers/usgs_mrds/, rendered Cu deposit-atlas at strata/results/deposit_atlas/copper.md.

Live-data corrections (verified 2026-05-05 against the actual mrds-csv.zip):

1. Bulk URL is https://mrdata.usgs.gov/mrds/mrds-csv.zip (with the /mrds/ segment), NOT https://mrdata.usgs.gov/mrds-csv.zip as the original §1 wrote. The bare URL returns 404. All sibling artefacts also live under /mrds/.
2. The bulk denormalised CSV exposes 46 columns - plant_type, plant_id, and operator are NOT among them. Those columns live in the normalised rdbms-tab.zip dump only. Initially (commit 5768124) the v1 fetcher dropped those paths and emitted no OperatorReference rows + no Plant-specific role fan-out; subsequently (commit 9e523b5, see RDBMS-upgrade callout below) the fetcher gained those paths back by consuming the RDBMS dump alongside the bulk CSV.
3. score is a single-letter grade A-E (D=70 % / C=14 % / E=8 % / B=8 % / A=0.3 %), NOT the integer 0-100 the §3.1 row claims. The implementation maps A/B → HIGH, C → MEDIUM, D → LOW, E → PLACEHOLDER.

Other landscape claims hold up: 304,632 rows, ~80 % US, the dev_stat

• commodity vocabularies, the model-code brace-delimited shape, approximate-year qualifier semantics. Live first-fetch admitted 237,087 facilities (78 %) - Occurrence rows + missing-coordinate rows account for the 67 K dropped.

This is the reference for porting the v0 (strata-shaped) MRDS fetcher onto the v1 schema spine: slim Facility + multi-row Classification + FacilityIdentifier + OperatorReference + LifecycleEvent + FacilityAttribute. The landscape below names every load-bearing MRDS field, maps it to its v1 destination, and surfaces the schema gaps that may need an ADR-grade decision before the rebuild lands.

RDBMS-dump upgrade landed (2026-05-05, commit 9e523b5). The v1 fetcher now downloads BOTH mrds-csv.zip (~24 MB) and rdbms-tab-all.zip (~118 MB). The bulk CSV remains the spine (universe of 237 K admitted facilities + identity + coords + commodity

• lifecycle); the RDBMS dump layers depth on top via per-dep_id joins:

• Ownership.txt → OperatorReference rows (87,032 facilities with operator-chain history; Owner-Operator fans out to OWNER + OPERATOR; deterministic entity_id keyed on company_id or sha1(owner_name)).
• Production_detail.txt + Resources.txt + Resource_detail.txt → FacilityAttribute time series (annual production + resource estimates + grades).
• Materials.txt → tuple-valued ore_minerals / gangue_minerals attributes.
• Workings.txt → workings_type / workings_area / workings_depth physical-site dimensions.
• Names.txt + Reference.txt + Holdings.txt → alternate_names / bibliographic_references / land_holding_types attributes.
• Deposits.txt plant_tp / plant_ident / mill_meth / min_meth → process_role + process_technology Classifications (Smelter / Refiner / Mill / Acid Plant disambiguation; Smelter/Refiner fans out to two roles).
• Deposits.txt disc_yr / deposit_size → discovery_year / deposit_size_class attributes.

Live ingest count after the upgrade: 2,508,581 records across the resolved store (vs. 1,854,724 pre-upgrade - 35 % growth from per-dep_id depth). Sections §3.7 / §4.4 / §4.9 below describe the currently-implemented behaviour for plant rows + operator emission; earlier text below those headings is preserved as historical design rationale.

---

1. Source overview

Publisher. U.S. Geological Survey, Mineral Resources Program (Reston, VA). Owning division historically the National Minerals Information Center; the dataset itself is a 1996 merger of two predecessors:

• Original MRDS (USGS, 1970s - 1990s mainframe system).
• MAS / MILS - Mineral Availability System / Mineral Industry Locator System, originally maintained by the U.S. Bureau of Mines before its 1996 closure. Records migrated wholesale into MRDS.

The migrated harmonisation produced today's denormalised record set; both the legacy mrds_id (pre-merger MRDS records) and mas_id (pre-merger MAS records) are preserved as columns on each row, and a new system-generated dep_id (12-digit numeric) is the canonical primary key.

License. US federal-government work. The formal MRDS metadata record at mrdata.usgs.gov/metadata/mrds.html declares both Access Constraints: none and Use Constraints: none. The applicable statutory authority is 17 U.S.C. § 105 ("works of the United States Government" are not eligible for copyright protection). The USGS open-data conventions (usgs.gov/information-policies-and-instructions/copyrights-and-credits) add the standard agency-attribution courtesy: cite USGS as the source even though redistribution is unrestricted. License code FF should register: public_domain (jurisdiction US).

Citation (canonical). Mason, G.T. and Arndt, R.E., 1996, Mineral Resources Data System (MRDS): U.S. Geological Survey Data Series 20, https://doi.org/10.3133/ds20. Edition: 20160315. Publisher: USGS, Reston, Virginia. A secondary DOI exists for the record-quality scoring tables (doi.org/10.5066/P9DYLWMP).

Public landing page. https://mrdata.usgs.gov/mrds/ - file-format options published below.

File formats published (verified 2026-05-05).

The single denormalised mrds-csv.zip is the canonical bulk-ingest artefact and what the FF fetcher should target. The rdbms-tab-all.zip carries richer normalised relations (per-deposit ore minerals, per-deposit references, etc.) that the denormalised CSV collapses into pipe-separated cells; FF v1 does not currently need those, but flag for v1.1 if geochemistry queries become first-class.

Dataset frozen at 2016-03-15. Verified - the MRDS metadata record publishes Edition: 20160315, and the MRDS landing page carries an explicit notice: "As of 2011, USGS has ceased systematic updates to MRDS." The 2016 date is the last file-system mtime push; no content refresh has occurred since. The bulk URL still resolves to the 2016 snapshot today (2026-05-05) - verified by HEAD request and content-length match. The successor dataset USMIN is US-only and not a replacement; it warrants its own fetcher when needed (out of scope for the MRDS port).

Global coverage statistics. 304,328 entity points (deposit / occurrence / plant records) worldwide per the formal MRDS metadata. Geographic skew (from strata's loaded copper subset, which is representative): ~80 % of records are US; Chile, Peru, Mexico, Canada round out the top 5 non-US jurisdictions. Older records (pre-2000 MAS imports) typically have spotty operator + grade data; modern entries (post- 2005) are richer but cover only the high-value deposits curated for USGS deposit-model studies. Per-status fan-out for Cu (a useful proxy): 3,931 producer + 18,345 past_producer + 7,564 prospect + 12,161 occurrence + 558 plant + 2,374 unknown (see copper.md).

---

2. File-level structure

The bulk ZIP. mrds-csv.zip (23 MB). Contains one CSV file named mrds.csv (~360 MB uncompressed). UTF-8 encoded with a BOM (utf-8-sig at decode) - strata's parser uses encoding="utf-8-sig" to strip it. Comma delimiter, double-quoted fields, RFC-4180 CSV semantics. Header row carries the column names shown in §3 below.

The mrds.csv primary file. ~304,328 data rows. Fully denormalised - many-to-many relationships in the source database (commodities, ore minerals, references, etc.) are collapsed into pipe-separated or comma-separated values within single cells. Fields that USGS publishes as multi-row child tables in the normalised RDBMS dump (rdbms-tab-all.zip) flatten into delimiter-packed cells here:

• Multiple commodities → commod1 / commod2 / commod3 columns, each potentially comma-separated ("Copper, Gold, Silver") and optionally hyphen-qualified ("Copper-primary, Gold-byproduct").
• Multiple ore minerals → ore cell as comma-separated descriptor.
• Multiple deposit models → model cell, brace-delimited and numbered ("{53: Porphyry Cu, 17: Skarn Cu}").
• Multiple host rocks → hrock_unit / hrock_type (single string each).

Sibling files. None inside mrds-csv.zip - that archive is single-CSV-only. The richer rdbms-tab-all.zip archive (118 MB) expands to 36 tables: deposits, ownership, production (+ production_detail), resources (+ resource_detail), materials, workings, names, references, holdings, plus lookup vocabularies. As of commit 9e523b5 (2026-05-05) FF v1 consumes both archives: the denormalised bulk CSV remains the spine for facility identity / coords / lifecycle, and the RDBMS dump feeds eleven domain tables for operator chains, production / resource time series, mineralogy, workings, and plant disambiguation (see top-of-doc upgrade callout). Cell-packed multi-values inside the bulk CSV are still handled via defensive splitting per §7.

Provenance metadata USGS publishes.

• DOI (per-publication): 10.3133/ds20 (canonical); 10.5066/P9DYLWMP (record-quality scoring tables). Both stable.
• Edition tag in the formal metadata: 20160315 - exact byte hash of the bulk CSV is not published, but the edition tag plus the record count (304,328) is the de-facto integrity claim.
• Record-quality score column on every row - USGS-published estimate of information completeness for that record (varies wildly across the 304k records; useful for FF confidence-tiering at parse time).

The FF fetcher should record source_url=https://mrdata.usgs.gov/mrds/, citation per the canonical citation above, license=public_domain, and stamp source_hash from the sha256 of the on-disk mrds-csv.zip at fetch time (matches the BaseFetcher pattern from the v0 MSHA fetcher).

---

3. Column-by-column schema dump

The denormalised CSV publishes the following columns in this order (verified against mrdata.usgs.gov/mrds/about.php). The "v1 destination" column points each MRDS field to where it lands in the FF schema. Drop = not carried (out of scope for v1, or already covered by another column). JM = stored under Facility.jurisdictional_metadata as a free-form key (the v1 escape hatch - promote to a structured field via ADR if it becomes load-bearing). Sample values are drawn from the verified Bingham Open Pit (dep_id=10069296), Bingham Canyon Plant (dep_id=10129805), San Manuel Smelter (dep_id=10234965), and the Cu deposit-atlas tally.

3.1 Identity + metadata

3.2 Site identification

3.3 Geographic location

3.4 Site characteristics

3.5 Temporal data

3.6 Commodities

3.7 Plant-specific fields (when dev_stat=Plant)

When dev_stat=Plant, USGS uses two additional fields that are blank for non-plant rows. They are documented in the WLCI MRDS user manual:

The full plant_id enum (~50 values) is the most disciplined process-technology vocabulary in MRDS - far cleaner than the free-form dep_type. The FF ff_process_technology taxonomy can seed every value as a sub-code under a mrds_plant_id_v1 extension. Critical for the §4 process-role mapping for Plant rows.

3.8 USGS commodity vocabulary (the commod* codelist)

Verified from the WLCI MRDS user manual pp. 10-15: MRDS uses a fixed Commodity lookup table with ~150 distinct commodity strings (ranges from "Abrasive" through "Zirconium"), each tagged with a Commodity_Group (~80 groups - "Copper", "Aluminum", "PGE", etc.) and Commodity_Type (M/N/E single character). Notable conventions:

• Compound entries - "Copper Oxide", "Copper Sulfide", "Iron-Pyrite", "Tin, Tailings" are first-class commodity codes.
• Refined-vs-mined disambiguation - "Gold, Refinery", "Silver, Refinery", "Lead, Refiner", "Lead, Smelter", "Tungsten, Mill Concentrate", "Zinc, Refiner", "Zinc, Smelter" are separate commodities from their unrefined counterparts. Useful at parse time to cross-confirm dev_stat=Plant rows.
• REE / PGE lump-up - "REE", "Cerium", "Yttrium" all map to the Commodity_Group="REE"; "PGE", "Palladium", "Platinum" etc. all map to Commodity_Group="PGE".
• Energy commodities - "Coal" (and its 4 ranks), "Petroleum (Oil)", "Natural Gas", "Uranium", "Geothermal", "Peat" are all present (out of FF mining scope but in MRDS).

The FF usgs_commodity taxonomy registry should seed all ~150 commodities at version mrds_2016 with their Commodity_Group as the parent code. This is distinct from the USGS Mineral Commodity Summaries vocabulary (~88 commodities, current annual version mcs2026) - they overlap but differ. The MRDS-vocabulary seed is for parsing MRDS rows; the MCS-vocabulary seed is for joining against USGS MYB and USGS MCS data products.

---

4. v1 schema mapping (the bridge)

Each MRDS row becomes one Facility (when admissible) plus a fan-out of multi-record output. The dev_stat value is the master switch.

4.1 Facility admission

This represents a behaviour change vs strata: strata emits all 6 statuses as DepositOccurrence (including occurrence) because its scope is deposit-level mineral inventory, not facilities. FF should drop occurrence rows since "no facility ever existed at that point" is a category miss for the canonical-store goal. Per Cu deposit-atlas counts, this drops ~12,161 of 44,929 Cu rows (27 %) - a significant filter; document and cite the count in the fetcher's emit-stats log.

4.2 FacilityIdentifier rows

For every admitted Facility:

The two legacy schemes need IdentifierScheme registry seed rows added alongside the fetcher; flag for the schema-side seeds commit.

4.3 Classification - industry

MRDS doesn't carry NAICS / ISIC codes directly, but commodity + dev_stat permits a high-confidence inference for most rows:

Implementation note (commit 9e523b5, 2026-05-05): the producer / past-producer commodity → NAICS / ISIC inference still applies. Plant-row industry inference is now reachable via plant_tp from the RDBMS dump (Deposits.txt) - see the §4.4 process_role table below for the live mapping of plant_tp → process_role.

The mapping table needs ~30 commodity → NAICS pairs to cover the MRDS commodity vocabulary; seed in _values.py as a commodity_to_naics_isic dict. Confidence is uniformly MEDIUM because MRDS data vintage often pre-dates current NAICS revisions and the inference is by-rule rather than published.

4.4 Classification - process_role

The FF-internal 10-stage decomposition (see process_stage.py). MRDS dev_stat + oper_type + plant_type together drive this:

4.5 Classification - process_technology

Three sources of process-technology evidence in MRDS, in priority order:

1. plant_id (when dev_stat=Plant) - the cleanest. ~50 values like "Solvent Extraction-Electrowin", "Smelter-Pyrometallurgy", "Bayer", "Heap Leach-Carbon Column-Merrill-Crowe". Map to ff_process_technology codes via a 1:1 enum; some compound values emit two Classifications (e.g., "Smelter-Pyrometallurgy" → pyrometallurgy_smelter).

2. model (any dev_stat) - USGS deposit model codes, brace- delimited and numbered. Most common values from the Cu deposit-atlas:

   USGS model code	Description	ff_process_technology code
   17	Skarn Cu	skarn_cu
   18	Skarn Pb-Zn	skarn_pb_zn
   21a	Porphyry Cu	porphyry_cu
   23	Pb-Zn (Mississippi Valley)	mvt_pb_zn
   28a	Volcanogenic massive sulfide (VMS)	vms
   28b	Kuroko massive sulfide	kuroko_vms
   30b	Sediment-hosted Cu	sediment_hosted_cu
   53	Porphyry Cu, skarn-related	porphyry_cu_skarn_related
   208	Sedimentary exhalative Zn-Pb	sedex_zn_pb

   The 30+ most common USGS deposit-model codes seed under ff_process_technology with parent codes pointing to the USGS-published deposit-model bulletin (https://pubs.usgs.gov/bul/b1693/) for traceability. Long tail of model codes: emit raw label, confidence MEDIUM, mark is_primary=False.

3. dep_type (free-form, sparse) - fall-back when model is blank but dep_type is populated. Stored as a Classification with confidence LOW because the field is not domain-controlled.

4.6 Classification - commodity_or_product

Per §3.6, commod1 / commod2 / commod3 fan out to one Classification row per (token, role). Taxonomy: usgs_commodity at version mrds_2016. is_primary is True only for commod1 tokens; commod2 and commod3 rows carry is_primary=False. The parser must:

1. Strip hyphen-qualifiers ("Copper-primary" → "Copper").
2. Split on comma + trim.
3. Look up against the seeded MRDS commodity vocabulary; emit one Classification per matched code.
4. Optionally cross-walk to UN CPC v2.1 (when the FF Cu-cathode use case demands it; v1 can defer this - UN CPC join lives in the data).

4.7 Classification - feedstock_origin

MRDS doesn't carry feedstock origin directly. Default rules:

• dev_stat ∈ {producer, past_producer} (extraction-stage) → feedstock_origin=primary_ore (HIGH confidence - by definition).
• dev_stat=plant + plant_type ∈ {Smelter, Refiner, Smelter/Refiner, Beneficiation (Mill), Acid Plant, Sinter Plant, Pellet Plant, Direct Reduced Iron Plant} → feedstock_origin=unknown (no MRDS column distinguishes primary vs secondary smelters). Flag: a meaningful fraction of MRDS Plants are secondary smelters (e.g., the Carrollton Copper Smelter at dep_id=10288750 is explicitly noted as "secondary smelter" in its description). Future Phase 1.5 work could enrich this from BREF / ICSG via cross-reference.
• dev_stat=plant + plant_type=Manufacturing Plant / Metal Plant → feedstock_origin=unknown (manufacturing inputs vary).

4.8 Classification - capability_scale

From prod_size: "Large" → mega, "Medium" → commercial_large, "Small" → commercial_small. Sparse field - most rows have it null and FF emits no Classification. When Y / N flags appear (legacy encoding), drop with a debug-level log.

4.9 OperatorReference

MRDS publishes operator and owner narratively rather than as foreign keys. For each admitted Facility:

• If operator (or owner) field non-null → emit one OperatorReference with entity_id="usgs_mrds_operator:<dep_id>" (placeholder reference; the lineage-side resolution against BQ Legal Entities is Phase 1.5+ work). role=ROLE_OPERATOR, confidence MEDIUM (operator strings often legacy - Bingham Open Pit's MRDS row says "Kennecott Copper Corp." which is a legacy spelling; current operator is "Kennecott Utah Copper LLC").
• If both operator and owner present and distinct → second OperatorReference with role=ROLE_OWNER, same placeholder ID scheme, confidence MEDIUM.

The placeholder-entity-id approach mirrors the v0 MSHA fetcher's pattern (entity_id=f"msha_operator:{operator_id}"); the synthesis layer joins on the placeholder string when reconciling against BQ.

4.10 LifecycleEvent

Three event types inferable from MRDS:

Approximate dates (yfp_ba="<" / ">") drop the year; do not emit a LifecycleEvent. The full prod_yrs descriptor ("intermittent 1850-1880, 1920-1940") is NOT parsed into multiple events at v1 - defer to v1.1; preserve raw in JM.

4.11 FacilityAttribute

When MRDS publishes resource estimates (rare - only on well-documented deposits), emit FacilityAttributes:

Resource-estimate fields are sparse (<5 % of rows have any tonnage) and the denormalised CSV doesn't expose them as separate columns the same way the RDBMS dump does. Flag as a known v1.1 work item: when the FF schema needs grade-tonnage curves as time series, re-target the RDBMS dump.

4.12 jurisdictional_metadata

The escape hatch for source-native fields without a structured FF home. Per §3, the following keys are written under Facility.jurisdictional_metadata (free-form dict[str, str]):

mrds_alt_names, mrds_region, mrds_ore_minerals, mrds_gangue_minerals, mrds_other_materials, mrds_orebody_form, mrds_work_type, mrds_model_raw, mrds_alteration, mrds_ore_control, mrds_host_rock_unit, mrds_host_rock_type, mrds_assoc_rock_unit, mrds_assoc_rock_type, mrds_structure, mrds_tectonic, mrds_yfp_qualifier, mrds_ylp_qualifier, mrds_dy_qualifier, mrds_discovery_year, mrds_last_production_year, mrds_prod_yrs_descriptor, mrds_discoverer, mrds_dep_type_raw, mrds_commodity_type, mrds_oper_type, mrds_plant_id.

Total ~25 keys per Facility maximum; most rows populate <10 of them. Keys are MRDS-prefixed to avoid namespace collision with EPA / MSHA extras when a single Facility has identifiers across sources.

---

5. Sample worked records

Three real MRDS rows + the v1 records each produces.

5.1 Producer - Bingham Open Pit Mine

Raw MRDS row (dep_id=10069296, fetched via show-mrds.php 2026-05-05):

dep_id=10069296
mrds_id=W002888
mas_id=(blank)
site_name=Bingham Open Pit Mine
names=Bingham Canyon Mine, Utah Copper
country=United States
state=Utah
county=Salt Lake
latitude=40.52077
longitude=-112.1458
dev_stat=Producer
oper_type=Surface
work_type=Open Pit
dep_type=(blank)
prod_size=Large
model={53: Porphyry Cu, 17: Skarn Cu}
commod1=Copper, Molybdenum, Silver, Gold
commod2=Lead, Bismuth, Platinum, Zinc
commod3=Palladium, Nickel, Tellurium, Selenium
ore=Bismuthinite, Chalcopyrite, Galena, Gold, Molybdenite, Pyrite, Bornite, Chalcocite, Covellite
gangue=Barite, Garnet, Orthoclase, Quartz, Rhodochrosite
hrock_unit=Bingham Stock
hrock_type=Quartz monzonite
disc_yr=1863
yr_fst_prd=1868
yr_lst_prd=2008
prod_yrs=1904-2008
operator=Kennecott Copper Corp. (Rio Tinto subsidiary)
owner=Rio Tinto PLC

Emitted v1 records (in stream order):

Facility(
  facility_id=fac_<minted_ulid>,
  canonical_name="Bingham Open Pit Mine",
  latitude=40.52077, longitude=-112.1458,
  coordinate_confidence=HIGH, coordinate_source=<prov>,
  country_iso2="US", region="Utah", locality="Salt Lake",
  status="operating",
  commissioning_year=1868, decommissioning_year=None,
  jurisdictional_metadata={
    "mrds_alt_names": "Bingham Canyon Mine, Utah Copper",
    "mrds_ore_minerals": "Bismuthinite, Chalcopyrite, ...",
    "mrds_gangue_minerals": "Barite, Garnet, Orthoclase, ...",
    "mrds_host_rock_unit": "Bingham Stock",
    "mrds_host_rock_type": "Quartz monzonite",
    "mrds_discovery_year": "1863",
    "mrds_last_production_year": "2008",
    "mrds_prod_yrs_descriptor": "1904-2008",
    "mrds_oper_type": "Surface",
    "mrds_work_type": "Open Pit",
    "mrds_model_raw": "{53: Porphyry Cu, 17: Skarn Cu}",
  },
  canonical_provenance=<prov>, minted_at=<fetch_ts>,
  minted_from_source="usgs_mrds_2016",
)

FacilityIdentifier(
  facility_id=<above>, scheme="usgs_mrds_dep_id",
  source_value="10069296", confidence=HIGH, provenance=<prov>,
)
FacilityIdentifier(
  facility_id=<above>, scheme="usgs_mrds_legacy_id",
  source_value="W002888", confidence=MEDIUM, provenance=<prov>,
)

# Industry classifications (commodity-inferred)
Classification(dimension=industry, taxonomy=(naics, 2022),
  code="212230", is_primary=True, confidence=MEDIUM, ...)
Classification(dimension=industry, taxonomy=(isic, rev_5),
  code="0729", is_primary=True, confidence=MEDIUM, ...)

# Process role - extraction (mine)
Classification(dimension=process_role, taxonomy=(ff_process_role, v1),
  code="extraction", is_primary=True, confidence=HIGH, ...)

# Process technology - porphyry Cu (skarn-related)
Classification(dimension=process_technology, taxonomy=(ff_process_technology, v1),
  code="porphyry_cu_skarn_related", is_primary=True, confidence=HIGH, ...)
Classification(dimension=process_technology, taxonomy=(ff_process_technology, v1),
  code="skarn_cu", is_primary=False, confidence=HIGH, ...)

# Commodity - primary (commod1 → 4 rows)
Classification(dimension=commodity_or_product, taxonomy=(usgs_commodity, mrds_2016),
  code="copper", is_primary=True, ...)
Classification(dimension=commodity_or_product, taxonomy=(usgs_commodity, mrds_2016),
  code="molybdenum", is_primary=True, ...)
Classification(dimension=commodity_or_product, taxonomy=(usgs_commodity, mrds_2016),
  code="silver", is_primary=True, ...)
Classification(dimension=commodity_or_product, taxonomy=(usgs_commodity, mrds_2016),
  code="gold", is_primary=True, ...)
# (commod2 + commod3: 8 more rows with is_primary=False)

# Feedstock origin
Classification(dimension=feedstock_origin, taxonomy=(ff_feedstock_origin, v1),
  code="primary_ore", is_primary=True, confidence=HIGH, ...)

# Capability scale - prod_size=Large
Classification(dimension=capability_scale, taxonomy=(ff_capability_scale, v1),
  code="mega", is_primary=True, confidence=MEDIUM, ...)

OperatorReference(
  facility_id=<above>,
  entity_id="usgs_mrds_operator:10069296",
  role=ROLE_OPERATOR, confidence=MEDIUM, ...,
  # operator string preserved in provenance.notes:
  # "Kennecott Copper Corp. (Rio Tinto subsidiary)"
)
OperatorReference(
  facility_id=<above>,
  entity_id="usgs_mrds_owner:10069296",
  role=ROLE_OWNER, confidence=MEDIUM, ...,
  # owner string in provenance.notes: "Rio Tinto PLC"
)

LifecycleEvent(
  event_id=evt_<minted_ulid>, facility_id=<above>,
  event_date=date(1868, 1, 1),
  payload=CommissionedPayload(context="MRDS yr_fst_prd"),
  confidence=MEDIUM, provenance=<prov>,
)

# (no decommissioned event - dev_stat=Producer, not Past Producer)

Total: 1 Facility + 2 FacilityIdentifiers + 17 Classifications + 2 OperatorReferences + 1 LifecycleEvent = 23 records for one MRDS row. This is the upper-end fan-out for a fully-curated row; most MRDS rows produce 6-8 records.

5.2 Past Producer - Copper Queen Mine, Nevada (dep_id=10045466)

A representative past-producer record. Selected because it has operator history without being maximally curated.

Raw MRDS row (selected fields):

dep_id=10045466
site_name=Copper Queen Mine
country=United States, state=Nevada, county=Mineral
latitude=38.3, longitude=-118.6
dev_stat=Past Producer
oper_type=Underground
dep_type=Porphyry Cu-Mo
prod_size=Small
commod1=Copper
commod2=Molybdenum, Silver
yr_fst_prd=1885
yr_lst_prd=1942
prod_yrs=1885-1942 intermittent
operator=(blank)
owner=Anaconda Copper Mining Co. (historical)

Emitted v1 records (compressed):

• 1 Facility - status=decommissioned, decommissioning_year=1942.
• 1 FacilityIdentifier - usgs_mrds_dep_id=10045466, HIGH conf.
• Industry classification: NAICS 212230 + ISIC 0729 (Cu-mining inference).
• Process role: extraction (primary).
• Process technology: porphyry_cu_mo (from dep_type; conf MEDIUM - free-form field, not USGS deposit-model code; model was blank).
• 3 commodity classifications: copper (primary, commod1), molybdenum + silver (secondary, commod2).
• Feedstock origin: primary_ore.
• 1 OperatorReference - entity_id="usgs_mrds_owner:10045466", role=ROLE_OWNER (no operator field; only owner present).
• 2 LifecycleEvents: commissioned (1885-01-01), decommissioned (1942-12-31).

Total: ~12 records. Past-producer rows commonly emit fewer records than active producers because operator strings tend to be legacy and ownership-only (not active-operator).

5.3 Plant - San Manuel Copper Smelter (dep_id=10234965)

A representative smelter-Plant record from the richer MRDS field surface. The current bulk-CSV fetcher admits this as a Plant facility but does not see plant_type, plant_id, or operator.

Raw MRDS row:

dep_id=10234965
mas_id=0040210030
site_name=(Facility) San Manuel Copper Smelter
names=San Manuel Smelter
country=United States, state=Arizona, county=Pinal
latitude=32.61621, longitude=-110.62238
dev_stat=Plant
oper_type=Processing Plant
plant_type=Smelter
plant_id=Smelter-Pyrometallurgy
commod1=Copper
com_type=M
operator=Magma Copper Co.
ownership=Private
note=1973 capacity 200000 TPY

Emitted v1 records:

Facility(
  facility_id=fac_<ulid>,
  canonical_name="San Manuel Copper Smelter",   # "(Facility)" prefix stripped
  latitude=32.61621, longitude=-110.62238,
  coordinate_confidence=HIGH, ...,
  country_iso2="US", region="Arizona", locality="Pinal",
  status="unknown",   # MRDS Plant rows don't disclose current operating status
  jurisdictional_metadata={
    "mrds_alt_names": "San Manuel Smelter",
    "mrds_oper_type": "Processing Plant",
    "mrds_plant_id": "Smelter-Pyrometallurgy",
    "mrds_commodity_type": "M",
  },
  ...,
)

FacilityIdentifier(scheme="usgs_mrds_dep_id", source_value="10234965")
FacilityIdentifier(scheme="usgs_mas_id", source_value="0040210030")

# Industry - copper smelting
Classification(industry, naics_2022, "331410", primary=True, MEDIUM)
Classification(industry, isic_rev_5, "2420", primary=True, MEDIUM)

# Process role - primary smelting
Classification(process_role, ff_process_role_v1, "primary_smelting",
  primary=True, HIGH)

# Process technology - pyrometallurgy
Classification(process_technology, ff_process_technology_v1,
  "pyrometallurgy_smelter", primary=True, HIGH)

# Commodity - copper
Classification(commodity_or_product, usgs_commodity_mrds_2016,
  "copper", primary=True, HIGH)

# Feedstock origin - UNKNOWN (MRDS doesn't tell us if primary or secondary)
Classification(feedstock_origin, ff_feedstock_origin_v1, "unknown",
  primary=True, HIGH)

OperatorReference(
  entity_id="usgs_mrds_operator:10234965",
  role=ROLE_OPERATOR, confidence=MEDIUM,
  # provenance.notes = "Magma Copper Co. (1991)"
)

# FacilityAttribute - the rare disclosed capacity
FacilityAttribute(
  attribute_name="capacity_cu_tpy",
  value=200000.0, unit="metric_tons_per_year",
  measurement_year=1973, confidence=MEDIUM, ...,
)

# (no LifecycleEvents - yr_fst_prd / yr_lst_prd not populated for plants typically)

Total: ~10 records. Plant rows commonly have richer process-tech detail (clean plant_id enum) but sparser temporal data.

---

6. Schema gaps surfaced

MRDS data exposes several axes that v1 doesn't have a clean home for. Categorise: defer-to-JM (acceptable v1 compromise) vs schema-extension (ADR-grade work).

6.1 Discovery-event LifecycleEvent type - defer or extend?

disc_yr is widely populated (~30 % of rows) and conceptually a distinct lifecycle event from commissioned. Per lifecycle.py, the v1 LifecyclePayload union covers commissioning, expansion, suspension, resumption, ownership_changed, divestiture, regulatory action, decommissioning, and correction - but no discovered event type.

Decision recommendation: defer to JM (mrds_discovery_year) for v1; flag as candidate for ADR 010 if other sources (Australia state mining, Brazil ANM) regularly publish discovery dates. Discovery is mining-specific and may not generalise to chemicals / semi-conductors / pharma.

6.2 Resource estimates / reserve figures - FacilityAttribute is the right home, but...

Bingham Canyon (dep_id=10129805) carries tonnage=1,088,000,000, length=3500m, width=2500m, thickness=800m, Cu_grade=0.59 wt%, Mo_grade=0.45 wt%, etc. The FacilityAttribute record has measurement_year and unit - the year + unit fields are sufficient. The schema does not have a structured way to express "this is a 2010 reserve estimate as opposed to a 2008 production figure"; both flow through measurement_year with a free-form attribute_name.

Decision recommendation: emit FacilityAttribute with discriminating attribute_name (reserve_estimate_tonnage_2010 vs production_tonnage_2008) and document the convention in the fetcher's docstring. ADR-grade extension (a new MeasurementType enum) is over-engineered for v1; revisit when a second source (Brazil ANM, Australia state mining) starts emitting the same distinction.

6.3 Geochemistry data (host rock + ore minerals + gangue) - no structured home

MRDS publishes:

• ore - comma-separated ore minerals ("Chalcopyrite, Bornite, Chalcocite, Covellite")
• gangue - comma-separated gangue minerals
• hrock_unit / hrock_type - host rock formation + lithology
• arock_unit / arock_type - associated rock
• alteration - alteration assemblage description

These don't fit FacilityAttribute (not quantitative) and don't fit Classification (not taxonomy-bound). For v1 they live in jurisdictional_metadata as raw strings.

Decision recommendation: defer to JM for v1. Future v1.1 may introduce a GeochemistryRecord type if cross-source analysis (MRDS + Australia state mining + Brazil ANM, all of which publish geochemistry) creates a query-time need. Flag for ADR 011.

6.4 Grade-tonnage time series - definitely v1.1+

MRDS publishes per-year tonnage and grade only for the most curated deposits (~5 % of rows), and the denormalised CSV doesn't expose them as separate columns. The richer rdbms-tab-all.zip does (multi-row production-history tables), but FF v1 doesn't ingest that.

Decision recommendation: out of scope for v1. Re-target the RDBMS dump in v1.1 when the canonical store needs grade-tonnage curves. Document the future-work hook in the fetcher's docstring.

6.5 Ownership history and operator strings - no bulk-CSV structure

The shipped MRDS bulk-CSV path does not carry operator or owner. Those fields live in the normalised RDBMS dump and, even there, are single-snapshot legacy strings rather than current legal-entity truth. The full ownership-transition history that BQ Legal Entities + SEC EDGAR can expose via LifecycleEvent.ownership_changed events is not in MRDS.

Decision recommendation: no schema change. The current MRDS fetcher emits no OperatorReference rows. Operator graph fill-in comes from complementary sources (MSHA for US mine operator strings, EPA FRS for non-mine facilities, and BrightQuery Legal Entities for the entity graph) or from a future MRDS RDBMS-dump enrichment fetcher.

6.6 RDBMS design note: plant rows with no plant_id

A meaningful fraction (~40 %) of dev_stat=Plant rows have plant_type populated but plant_id blank. The §4.4 mapping emits process_role from plant_type (e.g., Smelter → primary_smelting) but cannot emit a process_technology row when plant_id is blank.

Decision recommendation: emit no process_technology Classification when plant_id is blank. Log a counter via the fetcher's drops_no_process_technology field on the result container. Acceptable degradation; flag if cross-source enrichment becomes necessary.

6.7 RDBMS design note: dev_stat=Plant smelter/refinery ambiguity

The MRDS plant_type=Smelter/Refiner value is genuinely ambiguous - USGS uses it for vertically integrated complexes (smelter + electrolytic refinery on the same site, like the original El Paso copper complex). The §4.4 mapping emits two process_role Classifications (primary_smelting + refining) for this case, which is the right answer per ADR 007 §6 (one Facility, multiple process_role rows). But the MRDS row alone doesn't tell us which is the primary process; both Classifications carry is_primary=True, which violates the §7 constraint that at most one row per (facility_id, dimension, taxonomy_id) may be primary.

Decision recommendation: emit both primary_smelting (with is_primary=True) and refining (with is_primary=False) by convention; document the convention in the fetcher. Cross-source data (BREF, ICSG) can flip the primary marker later via correction events.

---

7. Parser implementation plan

7.1 Polars schema dict

import polars as pl

_MRDS_SCHEMA: dict[str, type[pl.DataType]] = {
    "dep_id":       pl.String,
    "mrds_id":      pl.String,
    "mas_id":       pl.String,
    "site_name":    pl.String,
    "names":        pl.String,
    "country":      pl.String,
    "state":        pl.String,
    "county":       pl.String,
    "region":       pl.String,
    "latitude":     pl.Float64,
    "longitude":    pl.Float64,
    "dev_stat":     pl.String,
    "oper_type":    pl.String,
    "dep_type":     pl.String,
    "prod_size":    pl.String,
    "model":        pl.String,
    "ore":          pl.String,
    "gangue":       pl.String,
    "other_matl":   pl.String,
    "orebody_fm":   pl.String,
    "work_type":    pl.String,
    "alteration":   pl.String,
    "conc_proc":    pl.String,
    "ore_ctrl":     pl.String,
    "hrock_unit":   pl.String,
    "hrock_type":   pl.String,
    "arock_unit":   pl.String,
    "arock_type":   pl.String,
    "structure":    pl.String,
    "tectonic":     pl.String,
    "yr_fst_prd":   pl.String,  # parse_year handles it; defensive against trailing decimals
    "yfp_ba":       pl.String,
    "yr_lst_prd":   pl.String,
    "ylp_ba":       pl.String,
    "disc_yr":      pl.String,
    "dy_ba":        pl.String,
    "prod_yrs":     pl.String,
    "discr":        pl.String,
    "com_type":     pl.String,
    "commod1":      pl.String,
    "commod2":      pl.String,
    "commod3":      pl.String,
    "plant_type":   pl.String,
    "plant_id":     pl.String,
    "score":        pl.Int32,
}

_MRDS_NULLS: list[str] = ["", "NULL", "null", "N/A", "n/a"]

7.2 Defensive coercion helpers in _values.py

def clean_text(value: str | None) -> str | None: ...
def parse_year(value: str | None) -> int | None: ...   # [1700, 2100], drop ≥0 sentinel
def parse_lat_lon(lat, lon) -> tuple[float | None, float | None]: ...
def normalise_dev_status(value: str | None) -> str: ...   # → "operating" / "decommissioned" / "project" / "unknown"
def country_to_iso2(value: str | None) -> str | None: ... # seeded ~200-entry dict
def split_commodity_field(value: str | None) -> list[str]: ...
def strip_commodity_qualifier(token: str) -> str: ...     # "Copper-primary" → "Copper"
def parse_model_field(value: str | None) -> list[tuple[int, str]]: ...  # "{53: Porphyry Cu, 17: Skarn Cu}" → [(53, "Porphyry Cu"), (17, "Skarn Cu")]
def confidence_from_score(score: int | None) -> ConfidenceTier: ...
def normalise_plant_type(value: str | None) -> str | None: ...   # 13-value enum lookup
def normalise_oper_type(value: str | None) -> str | None: ...    # 11-value enum lookup
def map_commodity_to_naics(commod1_first: str) -> str | None: ...
def map_commodity_to_isic(commod1_first: str) -> str | None: ...
def strip_plant_name_prefix(value: str | None) -> str | None: ... # "(Facility) X" → "X"

7.3 Sketched 30-line parser skeleton

def parse_mrds_csv(
    zip_path: Path, *, canonical_provenance: Provenance, minted_at: datetime,
) -> Iterator[BaseModel]:
    """Stream Facility/FacilityIdentifier/Classification/Operator/Lifecycle/Attribute records."""
    lazy = pl.scan_csv(
        zip_path, schema_overrides=_MRDS_SCHEMA,
        null_values=_MRDS_NULLS, encoding="utf8-lossy",
        truncate_ragged_lines=True,
    ).select(_MRDS_SCHEMA.keys())
    df = lazy.collect(engine="streaming")
    for row in df.iter_rows(named=True):
        dev_stat = normalise_dev_status(row["dev_stat"])
        if dev_stat == "occurrence":   # drop occurrences entirely
            continue
        lat, lon = parse_lat_lon(row["latitude"], row["longitude"])
        if lat is None or lon is None:   # geospatial-anchorless rows skip
            continue
        country_iso2 = country_to_iso2(row["country"])
        if country_iso2 is None:
            continue
        facility_id = mint_facility_id()
        yield _make_facility(facility_id, row, dev_stat, lat, lon, country_iso2,
                             canonical_provenance, minted_at)
        yield from _emit_identifiers(facility_id, row, canonical_provenance)
        yield from _emit_industry_classifications(facility_id, row, canonical_provenance)
        yield from _emit_process_role_classifications(facility_id, row, dev_stat,
                                                     canonical_provenance)
        yield from _emit_process_technology_classifications(facility_id, row,
                                                           canonical_provenance)
        yield from _emit_commodity_classifications(facility_id, row, canonical_provenance)
        yield from _emit_feedstock_origin_classification(facility_id, row, dev_stat,
                                                        canonical_provenance)
        yield from _emit_capability_scale_classification(facility_id, row, canonical_provenance)
        yield from _emit_operator_references(facility_id, row, canonical_provenance)
        yield from _emit_lifecycle_events(facility_id, row, dev_stat, canonical_provenance)
        yield from _emit_facility_attributes(facility_id, row, canonical_provenance)

Each _emit_* helper is 10-30 lines, returning 0 - N records depending on which MRDS columns are populated. Total parser surface ~400-500 lines across parse.py (split from _values.py per the v0 MSHA package shape).

---

8. Taxonomy seed requirements

The MRDS fetcher needs the following Taxonomy registry rows seeded before records can land in the canonical store. Each row is a single INSERT against the Taxonomy table with the metadata below. List the schema-side seeds commit needs to land alongside the fetcher rebuild.

Total: 9 Taxonomy registry rows. Some may already exist (e.g. ff_process_role v1 from the canonical-stage definitions) - verify during the schema-side seeds commit. The two USGS commodity vocabularies are distinct and both needed.

Identifier-scheme registry rows also required (separate from Taxonomy registry but parallel concept):

• usgs_mrds_dep_id - already present at SCHEME_USGS_MRDS_DEP_ID.
• usgs_mrds_legacy_id - needs new IdentifierScheme registry row + constant SCHEME_USGS_MRDS_LEGACY_ID.
• usgs_mas_id - needs new IdentifierScheme registry row + constant SCHEME_USGS_MAS_ID.

Total: 2 new IdentifierScheme registry rows alongside the 9 Taxonomy rows.

---

Summary of MRDS-port deliverables for the schema-side commit

1. 9 Taxonomy registry rows (above), including 2 USGS-commodity vocabularies at distinct version codes.
2. 2 IdentifierScheme registry rows (usgs_mrds_legacy_id, usgs_mas_id).
3. ~30 ProcessStage / process_technology seed rows under ff_process_technology v1 covering the most common USGS deposit-model codes + the MRDS plant_id enum.
4. ~150 commodity codes seeded under usgs_commodity mrds_2016 (the full MRDS Commodity table) plus the ~88 codes under usgs_commodity mcs2026 (the current MCS vocabulary).

Once seeded, the fetcher rebuild lands as a single PR layered on the existing v1 BaseFetcher ABC + cache-strategy infrastructure (per the v0 MSHA pattern), with ~400-500 lines of fetcher code spread across release.py, _values.py, parse.py, fetch.py, and __init__.py. Estimated record-emit volume: **~280k-310k Facilities

• ~2-3M total records** (Facilities + Identifiers + Classifications + OperatorReferences + LifecycleEvents + FacilityAttributes), of which ~80 % are US.