Methodology

Chain-of-title walk

Where the walk stops

• reached_ubo: a parent carries both an SEC CIK and a ticker, or is a government; or no further parent is recorded and the last organization reached does not match the patterns under opacity. An organization with no recorded parent at all ends here too.
• opacity: no further parent is recorded and the last organization reached has a Cayman Islands, British Virgin Islands, Bahamas, Jersey or Guernsey country (stored as KY, VG, BS, JE, GG) and no SEC CIK, or a name such as HoldCo, SPV, special purpose, nominees or blind trust. A CIK marks an SEC registrant, whose ownership is filed, so a domicile alone does not make it opaque.
• no_controlling_holder: the organization has recorded holders and the control review withholds every one, so no source establishes that any of them controls it. The holders are listed with their shares; where a filing states that the organization has no controlling shareholder, the review row’s quote, document and page are printed.
• cycle_detected: the chosen parent was already visited.
• depth_capped: the eighth hop was reached.
• contested: the two strongest candidates are different parents that tie on confidence and authority tier, with no later as-of date to separate them, and they are not named together in one document. Every claim is shown with where it leads, and no owner is named.

Algorithm

// walkOwnership() in src/lib/data/ownership-investigator.ts
function walkOwnership(entityId) {
  start = resolveCanonicalId(entityId)
  visited = { start }
  frontier = expandWithLicensedDbBridge(getAllIdsForOrg(start))
  chain = [hop0(start)]
  for hop in 0..MAX_HOPS-1:
    raw = SELECT * FROM ownership_upward_edges WHERE child_org_id IN frontier
    rows = filterPhantomSelfEdges(raw, frontier)
    if rows.empty: terminal = looksOpaque(tail)? opacity : reached_ubo; break
    edges = enrichEdges(rows)  // join source_url, authority_tier, share_pct
    edges = edges.map(e => withControlVerdict(e, ownership_edge_control_review))
    groups = groupBy(edges, canonicalParent)
    holders = groups.filter(g => every edge withheld)
    tail.holders = nestHolders(holders, unwithheldEdgesBetween(holders))  // "Of which", each citing its edge
    candidates = groups.filter(g => some edge not withheld).map(g => pickStrongestEdge(not withheld))
    if candidates.empty: terminal = no_controlling_holder (+ filing statement); break
    candidates.sort(compareEdgeStrength, then share desc, then canonical id before vendor id, then id)
    if tiesWithoutArbitration(candidates[0], candidates[1])
       && !sameDocument(candidates[0], candidates[1]): terminal = contested; break
    primary = candidates[0]; alternates = candidates[1..]
    if primary.canonicalParent in visited: terminal = cycle_detected; break
    visited.add(primary.canonicalParent)
    chain.push({ primary, alternates })
    if parent.cik && parent.ticker: terminal = reached_ubo; break
    if parent.org_type == 'government': terminal = reached_ubo; break
    if hop == MAX_HOPS-1: terminal = depth_capped
    frontier = expandWithLicensedDbBridge(getAllIdsForOrg(primary.canonicalParent))
  return { chain, terminal, terminal_detail }
}

At each hop compareEdgeStrength() ranks the candidate parents by confidence (HIGH, then MEDIUM, then LOW; an edge without a label counts as LOW), then by authority tier, then by the later as-of date, and the walk follows the first. Candidates still equal after that are ordered by the larger recorded share, then a canonical identifier before a vendor-keyed one, then identifier, so the database’s row order never chooses a parent. pickStrongestEdge() keeps one edge per parent; the others at that hop are listed as alternates with their own sources and dates. A recorded share is a fact about a holding and never decides control on its own: a 54.82% nominee holding confers none, and a 20.09% holder the issuer names as its controlling shareholder is a parent. Tier and confidence come from each source’s row: most SEC EDGAR edges are tier 1, most licensed edges tier 2, and every hop shows its own. Both functions live in src/lib/data/ownership-investigator.ts.

Edge cases handled

• Phantom self-edges (filterPhantomSelfEdges()). When one firm is held under two identifiers with a parent edge between them, merging the identifiers turns that edge into a self-loop; the walk drops it instead of stopping on a false cycle.
• Placeholder names. A hop can land on a licensed record that carries only a placeholder name. The walk continues through it, and the page shows the record’s identifier without the placeholder, so the link still resolves.
• Co-parents. Parents named together in one document (a founder’s holding vehicles, a stake split across two holding companies) are co-parents rather than a contest: the walk follows the strongest and shows where each of the others leads.
• A holding inside another (nestHolders()). A holder’s figure can count shares held through companies it controls, and one of those companies can be listed as a holder too. SMIC’s 2025 annual report gives CICT 14.97%, which includes the shares of Datang HK, a wholly-owned subsidiary of CICT that the same report lists at 14.06%. The walk draws Datang HK under CICT as “Of which” on a recorded edge from Datang HK to CICT that cites CICT’s own document, and the entry cites that edge; an edge citing anything else counts only where a review row for it establishes control, and an edge the review withholds never counts. CICT’s figure must count holdings through controlled companies, and its document must give Datang HK a share at the same date, no larger than CICT’s and not a share of the voting power. Holders are placed from the top down, so a figure is judged as it is drawn, and a holder whose recorded container has a figure that does not count it goes under the nearest holder above that container whose figure does. Edges reaching two branches of one holding put a holder under the nearest holder above both whose figure counts it; edges reaching different holdings, or running in a loop, leave it at the top, because the records contradict one another. A nested holder’s strongest record, when it is another reading, gets a line of its own; its weaker records are not shown, and no remainder is computed.

Outside this walk

• Limited-partner stakes. A limited partner’s stake in a fund is not control of that fund’s portfolio companies, so sovereign-wealth and pension-fund commitments are excluded from this walk. They appear separately, in the sovereign-LP attributions.
• Ownership through debt. The ownership tables carry equity edges only; convertible notes, mezzanine instruments and contingent-equity structures are not modelled.

Worked example

BIS designation reach

Algorithm

// getDesignationReachForEntity() in src/lib/data/ownership-investigator.ts
function getDesignationReachForEntity(orgId) {
  allIds = getAllIdsForOrg(orgId)
  direct = []
  for match in precomputed_crossref:
    if match.confidence_score >= 0.9 && match.matched_org_id in allIds: direct.push(match)
  // Belt-and-suspenders ILIKE on entity_name (catches entities the crossref pipeline missed)
  direct ∪= SELECT * FROM export_control_designations WHERE entity_name ILIKE orgName
  direct ∪= SELECT * FROM export_control_designations WHERE id IN allIds
  closures = SELECT * FROM designation_closures WHERE canonical_closure_entry_id IN allIds
  return direct.map(d => { ...d, hop_count: 0, is_direct: true, authority_tier: 1 }) ∪
    closures.map(c => { ...c, is_direct: false })
}

A walk starts from an anchor: the licensed ownership record taken to be the designated entity. A record anchors a designation only when it carries the designation’s exact name, legal form aside, in the designation’s own country; when that match survives a reading that keeps “holding” and “holdings”, which tell a holding company from the company it holds; and when no differently named designation claims the same record (admissible_anchors() in scripts/lib/bis_closure_anchor.py). The walks from the anchors fill designation_closures, and the matches are drawn from them offline (scripts/lib/bis-cascade.ts) by four routes:

// scripts/lib/bis-cascade.ts (offline; produces bis_crossref_matches AFFILIATE_CLOSURE rows)
  (1) direct      — the closure entry IS a compute-universe org
  (2) canonical   — the closure entry resolves to one via canonical_org_map
  (3) edge_extend — the closure entry's own SUBSIDIARY is one, one recorded edge down
    // confidence drops one tier (HIGH→MEDIUM, MEDIUM→LOW, LOW→LOW)
    // never the entry's PARENT, a JV or an AFFILIATE: a designation does not reach them
  (4) designation_chain — SUBSIDIARY edges down from the org an active name match
      links the designation to, at most 6 hops
  // A reviewed refusal (bis_matcher_overrides, match_decision 'reject') holds on
  // every route; route 4 walks nothing below a refused org.
  // Routes 3 and 4 follow no edge the ownership control review records as
  // conferring no control, so a designation does not reach a company through
  // a stake in it (a funding-round investment, for instance).

Edge cases handled

• A designation’s own record. Each designation creates an organization record of its own; a walk that returns to that record is not counted as reach (isCascadable() in scripts/lib/bis-cascade.ts).
• Removed designations. When a Federal Register notice removes a designation, its matches stop counting as active and stay on record, so what it once reached can still be traced.
• Direct and affiliate matches. The page marks a direct designation as such, at tier 1, and shows each affiliate match with the relationships it crossed.

Coverage ceiling

Worked example

Direct designations are tier 1, cited to the Federal Register. Affiliate matches carry the tier of the ownership records they walk; a match that needs a one-edge bridge beyond the end of a walk is downgraded one confidence step (HIGH to MEDIUM, MEDIUM to LOW).

Section 1260H, NS-CMIC and FASCSA listings

How a listed name is matched to a record

• Names compared. The name as listed, any former name the source prints, and the names a Federal Register notice says the entity is also known as, against every organization’s name and recorded aliases, ignoring case, runs of whitespace and full stops. A match on a merged duplicate counts as the record it was merged into, and only records the site publishes count.
• Names not compared. The short names a notice prints in parentheses, and OFAC’s a.k.a. names. They abbreviate or brand: OFAC prints CHINA TELECOM as an a.k.a. of two entries, CHINA TELECOM CORPORATION LIMITED and CHINA TELECOMMUNICATIONS CORPORATION; compared, it would put both listings on one record.
• One record, several, none. A name that matches exactly one record is attributed to it, and the record shows the listing with the printed name that matched. A name that matches several records, or none, is attributed to no record and held for review; the MCP company record returns the rows that name it among several, with the other records.

What a missing listing means

Each listing is tier 1, cited to its Federal Register notice or OFAC list file. Rows: us_restrictive_listings.

Sovereign LP attributions

Algorithm

// getSovereignAttributionsForEntity() in src/lib/data/ownership-investigator.ts
function getSovereignAttributionsForEntity(orgId) {
  allIds = getAllIdsForOrg(orgId)
  return SELECT id, sovereign_lp_name,
    intermediate_fund_name, intermediate_fund_id,
    commitment_date, commitment_usd_m, is_dollar_disclosed,
    commitment_status, data_source, source_url, data_vintage,
    authority_tier
    FROM sovereign_investment_chains
    WHERE target_canonical_org_id IN allIds
    ORDER BY commitment_date DESC
}

Reading the ledger

• Missing statuses. The source-status column shows a dash when none of an investor’s commitments has a recorded status. Where statuses are recorded, the ledger keeps the source’s wording and counts commitments under each label.
• Commitment amounts. A commitment covers the fund’s portfolio, so its amount cannot be assigned to one company. The public ledger counts how many commitments have a dollar amount in the source; fund names and amounts are withheld under licence.

National AI budgets and project spending are compiled separately in the sovereign AI programmes.

Ownership-change events

Algorithm

// getOwnershipChangeEventsForEntity() in src/lib/data/ownership-investigator.ts
function getOwnershipChangeEventsForEntity(orgId) {
  allIds = getAllIdsForOrg(orgId)
  partyDeals = SELECT pe_investment_id FROM pe_deal_investors
    WHERE canonical_party_org_id IN allIds
  // licensed rows: year + size band; date, amount, source URL withheld
  deals = SELECT * FROM v_pe_investments_grey
    WHERE canonical_target_org_id IN allIds OR id IN partyDeals
    ORDER BY deal_year DESC, announced_year DESC, id
  if deals.empty: return []
  investors = SELECT * FROM pe_deal_investors WHERE pe_investment_id IN deals.ids
  partiesByDeal = groupBy(investors, pe_investment_id).map(invs =>
    invs.map(inv => {
      name = inv.canonical_party_org_id ? getCanonicalName(inv.canonical_party_org_id) : null
      party_name = suppressPlaceholderStubName(name)  // placeholder label → null
      return { party_role, canonical_party_org_id, party_name, fund_name, fund_id }
    })
  basis = loadConsiderationBasis(deals.ids)   // raw fields of filing-checked deals only
  return deals.map(d => {
    e = { ...d, parties: partiesByDeal[d.id] ?? [] }
    terms = basis[d.id] ? qualifyDealConsideration({ ...basis[d.id], parties: e.parties }) : null
    return terms ? { ...e, primary_consideration: terms, amount_band: null } : e
  })
}

A party with no resolved name reads “Name withheld under licence” on a licensed record; otherwise the panel falls back to the fund name, then to “— name not recorded”. A licensed deal reads its year and size band (“2005 · Completed · $100M to $1B”); a deal with no recorded value reads “value unrecorded”.

Outside this chain

The licensed databases

Subscription sources

Published and withheld records

The three provenance markers

Supply relationships named from a licensed database

Growth and spending bands from licensed annual statements