STRATA · How it's built
From public registers to a supply-chain graph
A critical element passes through many hands between the ground and a finished product, and no single register follows it. Strata infers that chain at the level of individual organisations: it reads 255 sources (250 public, plus a licensed commercial company-identity corpus), resolves the same mines and companies across them on evidence, and builds a graph that answers who does what at each stage, in what quantity, and under whose control. This page walks through how, one claim at a time, with copper as the worked example and a status on every claim.
store figures measured 2026-11-17 on edition synth_71d8dc8770c5 (synthesised 2026-11-16) · copper figures from its 2026-11-16 profile · source counts read live
01 · The claims
Seven claims, each with a status
The method makes seven claims, from the sources it reads to how it stays current. Each is worked through in its own section below, and each carries a status: running today, running with stated limits, or on the roadmap. Select a claim to jump to it.
Facility-level and commodity-level sources
Every source has a catalogue card. Fetchers are routed by the record kinds they emit; curated tables carry a declared domain.
Evidence-based entity resolution
Only HIGH-tier evidence merges. MEDIUM and LOW matches are kept as links, and curated vetoes block known false merges. Registry numbers match only inside the register that issued them.
An automatically constructed, federated knowledge graph
A typed property graph in partitioned parquet, queried with DuckDB. Each source stays its own partition, joined only through the resolution layer.
Element-specific actors across a nine-stage lifecycle, including recycling
Nine stages are populated for copper. Two are demand estimates rather than actor censuses, and the recycling census is led by battery recyclers.
Estimated quantities and concentration per stage
Country concentration (HHI) is computed stage by stage. Company-level capacity is ranked, but outside uranium enrichment no owner-level HHI is computed yet.
Each organisation's controlling authority
A dated, tenure-aware control graph with an ultimate-parent walk. Coverage is strongest where registers are open; joint-venture splits come from a curated, cited layer.
Updated from real-time news and social media
The update path exists (vintages, lineage, dated edges). A news and social feed is designed but not ingested today.
02 · Sources
255 sources, in two kinds
Facility-level sources feed the graph: sites, companies, identifiers and ownership. Commodity-level sources feed the commodity profiles: production, trade, prices, reserves and demand. 46% sit in the official-survey tier (geological surveys, statistical and customs agencies, intergovernmental bodies such as the IEA, and national labs), and every one has a catalogue card with its publisher, licence, cadence and authority tier.
Facility level
67
Sources that feed the graph: sites, companies, identifiers and ownership.
e.g. MSHA, USGS MRDS facilities, EPA FRS, Brazil's ANM SIGMINE cadastre, GLEIF, SEC 13D/G, EDINET
Commodity level
188
Sources that feed the commodity profiles: production, trade, prices, reserves and demand, plus deposit records and curated plant rosters not loaded into the graph.
e.g. USGS Mineral Commodity Summaries, BGS World Mineral Statistics, UN Comtrade, Australia's Resources and Energy Quarterly
For the 148 fetchers the split is computed: a source is facility-level only if every record kind it declares is one the graph stores, and the fetch command, the store build and the catalogue apply that same test. The 107 curated and derived cards carry a declared domain. Of the site-record sources, 11 feed the current graph; BLM mining claims and the EEA register are catalogued but not loaded. 255 sources are catalogued today; each has a card with its publisher, licence, cadence, authority tier and the identifier schemes it contributes.
Catalogue by authority tier
210 sources
Every one of the 210 carries a metadata card - publisher, homepage, coverage, cadence, access method, format, licence, access URL, and an explicit list of known traps in that source.
03 · Pipeline
Fetch, parse, store, resolve, build
Construction is a pipeline of five steps, each with one guarantee it enforces. Sources are never merged at rest: each keeps its own partition and its own lineage, and they meet only in the resolution step. That is what federated means here.
- 01Fetch
One fetcher per source. Bulk files, ArcGIS and WFS services, EDGAR and national registers, each with a declared cache policy.
Raw artifacts are cached per source and vintage; a new release adds a vintage instead of overwriting the last, and the primary artifact is hashed at fetch.
data/raw/<source>/<vintage>/
- 02Parse
Each artifact streams into typed, validated records (facility, entity, ownership edge, supply, trade flow, price).
A record without provenance is a hard error, not a warning.
pydantic record models, extra fields forbidden
- 03Store
Graph records land in Hive-partitioned parquet, one partition per source and vintage, with a manifest of counts and warnings. Commodity series are re-parsed from the raw tier at each build.
Sources are never merged at rest. Each file names its source, vintage and fetch date in its footer.
data/resolved/<record_type>/source_id=/vintage=/
- 04Resolve
Identity resolution across sources: identifier passes, guarded name blocks, facility scoring, curated vetoes.
HIGH merges automatically; MEDIUM and LOW are kept as links. Every run writes an audit manifest of its match decisions.
data/ready/ (the reconciled tier)
- 05Build
Per-commodity profiles assemble production, trade, stage censuses, concentration and control on top of the graph.
Invariant rules sweep every shipped profile in the test suite; a violation fails the tests.
one profile per commodity
source_idsource_urlfetch_datesource_hashpage_or_sectionlicensecitation_urlprovenance_chainderivation_note04 · Resolution
Merge only on evidence
The same company appears as a CIK in one register, an LEI in another and a bare name string in a third. Resolution decides which records are one company. It merges on a shared identifier, an identifier-anchored name block or a high facility score, and keeps weaker matches as visible links rather than merging them. Three stories below, one unit each: the records, the pairs that decide them, and what reaches the map.
Story 1: from records to companies
Counted in records. A record is one row a source holds for an organisation.
7,107,677 organisation records from 36 sources (GLEIF 3,167,800 · SEC EDGAR 1,408,134 · BrightQuery 1,122,723 · Companies House PSC 1,023,809); 1,699,692 carry a CIK, 3,276,485 an LEI, 1,133,105 a BrightQuery id and 4,005,652 a register number.
- 1Organisation recordsone per source record, from 36 sources7,107,677
- 2Records that mergematched to at least one other record with HIGH evidence1,619,106
- 3Companies they becomeone company per cluster of merged records453,604
- 4Records standing aloneno HIGH match to any other record: each is its own company5,488,571
22.8%
of records merge with at least one other record
144
records in the largest company
29,254
facilities merged from records in two or more sources
One company, 54 records
FREEPORT-MCMORAN INC. is 54 records from 7 sources; 41 joined by a key or lineage, 13 by an anchored name. Pairs among its records, with the evidence each stores:
- HIGHBrightQuery record and SEC EDGAR record, shared CIKcik-key match, cik=0000831259
- HIGHBrightQuery record and GLEIF record, shared LEIlei-key match, lei=549300IRDTHJQ1PVET45
- HIGHBrightQuery record and BrightQuery record, BrightQuery lineageBrightQuery lineage: two of its own records of one company, cited to the primary filing
- HIGHUSGS MRDS record and GLEIF record, curated paircurated_identity v1.0
- MEDIUMEPA FRS record and USGS MRDS record, name and country onlyblock_key='freeport mcmoran copper and gold', name + country (no hard anchor)
- LOWBrightQuery record and GLEIF record, name in a crowded blockblock_key='freeport mcmoran', multi-anchor-block (n=42)
A record merges only on HIGH evidence; the pairs that carry that evidence are story 2.
Story 2: merge only on evidence
Counted in pairs. A pair is two records that may be one company.
The matcher judged 3,303,398 pairs. A company of n records carries up to n(n-1)/2 pairs, so pairs outnumber the records they fold (219,409 pairs are asserted both ways). Only HIGH pairs merge; the rest stay visible links.
How a pair is decided, in the order the matcher tries
- 1Shared identifier · both records carry the same CIK, BrightQuery id, LEI or register numberHIGH1,570,043
- SEC CIKboth records carry the same US filer identifier744,821
- BrightQuery org idboth records carry the same hub identifier656,290
- LEIboth records carry the same legal-entity identifier104,776
- Register numberthe same number in the register that issued it; 9,300 of them join a BrightQuery record64,156
- 2Anchored name block · one distinctive name in one jurisdiction, a CIK or LEI on one sideHIGH97,376
- 3Lineage and registers · BrightQuery lineage, US state filings, regrouping and the GLEIF mapping bridgeHIGH715,362
- BrightQuery lineagethe hub's own record-to-record lineage, cited to the primary filing692,662
- Other register evidenceUS state filings, regrouping and the GLEIF mapping bridge17,317
- Other passesHIGH edges from passes not named above5,383
- 4Curated pair · a person admitted it, with a citationHIGH165
- 5Name and country alone · names and country agree, nothing anchors themMEDIUM179,761
- 6Name across borders, or in a crowded block · names agree in different jurisdictions, or the block holds many anchored companiesLOW740,691
- 7Never-merge list · a person pinned the pair apartnever5,699
A pair falls through to the next gate when one does not hold. Identifiers beat names; registers and dates beat lineage; a veto beats everything.
2,382,946 HIGH pairs tie 1,619,106 records into 453,604 companies; only the HIGH share changes what counts as one company.
Story 3: into the map
Counted in records, then in companies. The map is the critical-minerals package the product ships, cut from the same edition and resolved the same way.
Of the 7,107,677 organisation records in the store, 173,457 are carried into the map. A company is on the map by one of four layers: direct, a stated operator, owner or parent role on a site in a commodity profile; parent, above such a company by control, at any depth; shareholder, the other end of a shareholding edge touching one of those; or subsidiary, up to three levels below a direct company whose stated role is on an upstream or midstream site.
- 1Records carried into the maporganisation records the public package ships, cut from the same edition173,457
- 2Records that mergematched to at least one other record with HIGH evidence71,673
- 3Companies they becomeone company per cluster of merged records20,712
- 4Records standing aloneno HIGH match to any other record: each is its own company101,784
On the map, copper: the companies the commodity bridge names as operators of copper sites, and how many have an owner recorded above them.
- 1Companies tied to copper sitesnamed as an operator of a copper site on the commodity bridge15,897
- 2With any live ownership edgean owner, controller or parent recorded above the company, no end date799
- 3With a controlling parent recordedconsolidates, controls, ultimate parent, or a stake of 50% or more517
122,496
companies on the map: merged clusters and standing records together
942,330
facility sites on the map
564,697
site-to-commodity rows tying sites to the minerals they handle
3.3%
of copper-site companies have a controlling parent recorded in force
Stories 1 and 2 span the whole store; this is the cut a reader meets on the map, and copper is one commodity on it. The walk from a plant to its owner is section 07.
Is it the same site? A weighted score
coordinates 0.40 · name 0.40 · state / region 0.10 · commodity 0.10
Distance decays smoothly with a 10 km half-distance rather than a hard radius, because a register's shaft entrance and a survey's deposit centroid for the same mine can sit kilometres apart.
Thresholds
At or above 0.72 the records merge when each source contributes one record; otherwise, and between 0.65 and 0.72, they are linked as the same complex. When a high-scoring cluster holds several records from one source, it is split by site name with stage words kept, so Chuquicamata's pit records merge while its smelter and refinery stay separate. Three gates then check every merge: a stage word in a name keeps a mine apart from its smelter, one source's records must sit within 1 km of each other, and a merge resting on coordinates alone is cut unless the names corroborate. Whatever a gate cuts stays linked as the same complex.
05 · The graph
Three node types, typed edges, a source on every one
The result is a typed property graph: facilities, organisations and commodities as nodes, and operator, ownership, control and identity relationships as edges. Every edge names its source. Control and shareholding edges carry a start date, and an end date where a register closes the spell; identity, co-location and most operator edges carry no dates. It lives in partitioned parquet and is queried with DuckDB, not a triple store. Same-entity edges are the resolution itself and are counted apart from ownership.
- Commodity505 codes
↓ commodity bridge
564,697
- Facility942,330 sites
↻ same_complex_as 1,356,941
↓ operator · owner · parent
772,470
- Organisation7,107,677 entity records
↻ same_entity_as 3,303,398
↓ controls · owns_share · consolidates
2,045,150
ultimate_parent_via 739,251
- Ultimate parentorganisation at the top
06 · Nine stages
Who does what, stage by stage, for copper
Process roles that a source states, or that a facility-type register records, are kept apart from roles inferred from industry codes, and each company inherits the stages of the plants it runs. Seven of copper's nine stages are named-plant rosters drawn from company filings: partial censuses, in which one plant can sit in more than one stage. The last two are demand estimates, because no register lists every manufacturer that uses copper. Dashed cards mark them.
- 1Mineactor census
274mines · 58 companies
Zijin 1.41 · BHP 1.19 · Anglo American 1.09 Mt Cu, whole-site basis
The register-based roster behind the product's copper page lists 2,156 operating copper sites.
- 2Concentrateactor census
221concentrators · 54 companies
Zijin 1.41 · BHP 1.19 · Anglo American 1.09 Mt Cu, whole-site basis
- 3Smelting & refiningactor census
208plants · 127 smelt, 152 refine, 71 do both
Jiangxi Copper 1.78 Mt/yr leads smelting · Freeport-McMoRan 1.81 Mt/yr leads refining
- 4Chemical processingactor census
92plants · 46 companies
rod, tube and foil mills, battery-material plants and recyclers, DRC copper-cobalt hydromet plants, smelter acid circuits
- 5Fabricationactor census
83semis mills, the major fabricators
against 66.1 Mt/yr global semis capacity (ICSG, 2024)
Wire rod is 59% of that capacity.
- 6Recyclingactor census
29plants · 20 companies
17.1% of refined copper was secondary in 2024 (ICSG)
27 of the 29 are battery recyclers that recover copper; the only one with a stated capacity is Boliden's Harjavalta smelter.
- 7Componentsactor census
141plants
cable, rod, tube, alloy and foil
- 8Manufacturingdemand estimate
26%of copper use is building construction (2024)
consumer and electronic 23 · power and telecom 17 · transport 13 · industrial 12% (ICSG)
- 9Final technologydemand estimate
8.1Mt for grids in 2040 (IEA APS)
EVs 4.0 · solar 2.6 · wind 0.9 Mt
07 · Quantities
How much, and how concentrated, at each stage
Country-level quantities come from the agency or study group that publishes each stage (USGS, IEA, ICSG); plant-level capacity comes from company filings. Concentration is the Herfindahl-Hirschman index over country shares. Estimates are labelled as estimates: a share of named countries is scaled by coverage before it becomes a world share, and contained metal is a declared quantity times a stated grade or coefficient.
Copper, country HHI by stage
Mine output
share of named-country output, 2025 (estimated)
Refined output
refined supply, 2024
Refining capacity
ICSG top-20 refineries (21 with ties)
Smelting capacity
ICSG top-20 smelters (22 with ties)
US import partners
refined imports, 2021-24
The same metal reads diversified at the mine and concentrated at the smelter. Each rung names its own source and year (USGS MCS 2026, IEA Critical Minerals Dataset 2025, ICSG World Copper Factbook 2025), because the panels are not one dataset and are never blended into one.
World share from a named-country share
26.48% × 86.96% coverage
Chile ≈ 23.0% of world mine output
Shares are computed over named countries, then scaled by coverage before a world claim is made.
Contained metal in a deposit
20,731 Mt ore × 0.62% Cu
El Teniente ≈ 128.5 Mt contained copper
Identified resources, from the deposit's grade-tonnage record. Grade is a property of the deposit, not of a lifecycle stage.
Contained metal in a trade flow
15,912 t HS 2603 concentrate × 0.27
≈ 4,296 t contained copper
One Colombian exporter's 2026 declarations to one Peruvian buyer. The weight is declared; 0.27 is strata's representative grade for HS 2603, so the content is an estimate.
Contained capacity by ultimate parent
10.92 Mt/yr across 53 operating plants with a stated figure · 712 plants in the census
Freeport-McMoRan
11 plants · US, Chile, Peru, Spain
1.81 Mt/yr
Jiangxi Copper
4 plants · China
1.78 Mt/yr
Southern Copper (Grupo México)
7 plants · Mexico, Peru
0.85 Mt/yr
Plants roll up to the ultimate parent named in their own filings, not to the graph walk in §08. The ranking is published; no owner-level HHI is computed over it yet.
08 · Control
Who ultimately controls each plant
Control is a walk from the operator up through control edges to an ultimate parent. The store holds 1,026,638 control edges, most parsed from SEC Exhibit 21 subsidiary lists, 48,893 shareholdings with an exact percentage, and a US mine-level controller history from MSHA of 111,636 spells · 47,107 mines · 33,207 controllers. Three rules govern the walk.
A closed edge ends the walk
An edge a source has closed (a ceased UK PSC holding, a reported SEC transfer, an MSHA controller change) ends the walk, and the former parent is shown as history. An edge with no recorded end is treated as current and carries the last year a filing attested it.
A minority stake is not a parent
A sized shareholding is walked only above 50%, so an index fund's or a partner's minority holding is never promoted to controller.
Identity before ownership
An operator is resolved to its registered company through the same_entity_as layer before the walk starts. A bare name a source wrote (an MRDS owner field, an ICMM site label) is labelled as such and never starts a walk.
Bagdad copper mine, Arizona: who controlled it, and when
MSHA controller history
Each spell has a start and an end, so the register answers who controlled the mine in any year. MSHA labels a controller by its latest name, so early spells can carry a name the company took later. The ownership walk itself answers for today and shows ended edges as history.
Joint ventures: control is a split, not a name
BHP57.5%Rio Tinto30%JECO12.5%
Ivanhoe39.6%Zijin39.6%DRC state20%Crystal River0.8%
MMG (China Minmetals)62.5%Guoxin22.5%CITIC Metal15%
CMOC80%Gécamines20%
Rio Tinto66%Mongolia state34%
These splits come from a curated, hand-maintained layer with one cited source per row: company reports for most, Wikipedia for Tenke Fungurume and Oyu Tolgoi. It is shown beside the automated walk and labelled as curated, not merged into it.
09 · Staying current
Built to change, and the feed that comes next
The graph is built so new facts can land beside old ones: sources are stored under vintage labels, every record names where it came from, and control edges carry dates where a register records them. Rebuilds run on demand today. A news and social signal lane is designed on top of that, and is shown here as roadmap, not as a running feature.
Vintaged sources
Each source is stored under a vintage label. A release with a new label adds a partition beside the old one; re-fetching the same label replaces it.
Lineage on every record
Each record names the source, URL and fetch date it came from. Shipped commodity profiles are versioned in git and can be reverted; the reconciled store is rebuilt, not versioned.
Dated control edges
Ownership edges carry valid_from and valid_to. Where a register records an end, the edge closes and the next one opens; most edges carry a start date only.
Rebuild
Refresh, fetch, resolve and reseed are commands, run on demand today rather than on a schedule.
Change-of-control filings
A discovery worklist of SEC 8-K Item 5.01 change-of-control filings is built; extracting them into dated control edges is not yet.
News and social signals
Designed as a low-authority lane: an event proposes a candidate edge, and a register or filing must corroborate it before it changes the graph.
The event lane, as designed
roadmap- 1Event observed
- 2Candidate edge, LOW
- 3Register or filing corroborates
- 4Promoted, dated, cited
A news item or post could only propose. It would enter below every existing authority tier (none exists for press today) and change the graph only once a register or filing agrees, so a fast signal never overrides a primary source.
10 · Known limits
What the numbers do and do not say
Stated here first, so nobody has to find them on their own.
Facility-level means graph-facing
The facility domain is the sources meant for the facility, organisation and ownership graph. Most are company and ownership registers, not site lists, and two site sources are catalogued but not loaded today.
Corroborated means the commodity
A corroborated facility is one where two or more sources agree on its commodity, not two independent confirmations that the site exists.
Coverage follows publication
Open bulk registers dominate the facility count: USGS deposit records and the EPA facility registry in the US, and the national mining cadastres of Brazil and Peru. Counts describe publication regimes, not geology.
Owner concentration is ranked, not indexed
Capacity is attributed to the ultimate parents named in company filings and ranked, but an owner-level HHI is not computed yet (uranium enrichment, by operator group, is the one exception). Plant-level HHI is withheld where fewer than half the plants disclose capacity.
Joint-venture splits are curated
Headline stake splits for major assets come from a cited, hand-maintained layer kept beside the automated walk, not from the walk itself.
Modelled content is labelled
Contained metal per shipment or plant applies an engineering coefficient to a declared weight or a nameplate capacity. It is an estimate and is shown as one.
The facility gates trade a little recall for precision
The name gate is right on about 91% of the merges it cuts in a held-out sample. The true matches it cuts, like an operator's name against a differently named site, stay linked as the same complex rather than lost.