Embedding Retrieval Surfaces Similarly Named, Unrelated Subsidiary in Corporate-Structure Chart

Issue: A Due-Diligence Agent Building a Target Company’s Corporate-Structure Chart From Filings and Registry Data, Using Semantic Similarity Search to Match Entity Names Across Documents, Merges or Links a Subsidiary Into the Target’s Structure Based on Name Similarity Alone, When the Matched Entity Is in Fact an Unrelated Company With a Coincidentally Similar Name

Frequency: Occasional

Symptoms

• A buyer’s risk assessment flags litigation or liability exposure that actually belongs to an unrelated company, while the target’s real subsidiary – carrying a different, undisclosed liability – goes unreviewed because the chart never linked it in
• The entity the agent linked in has its own registration number and tax ID on file, distinct from the target’s genuine subsidiary of a similar name, but the linking step never queried either registry field
• Because the agent is pulling from filings, a foreign companies registry, and a UCC lien database simultaneously, a name that reads as unique within any single source can collide with an unrelated entity the moment it’s cross-referenced against the other two
• Reviewers treat every link on the chart as equally load-bearing during diligence, since nothing distinguishes a link backed by a filed ownership disclosure from one the agent inferred purely from name overlap
• Discovery happens only when someone manually traces one specific entity against the target’s actual filed ownership disclosures for an unrelated reason (usually because a number downstream looks wrong), not because the chart-building process itself flagged the link as uncertain

Root Cause The agent’s entity-linking step treats “same name across sources” and “same legal entity” as interchangeable, because its matching function only ever sees the name string – it was never given a registration-ID or tax-ID field to check against when a candidate match is proposed across the three source types it ingests. Cross-jurisdictional due diligence compounds this: the same descriptive words (“Logistics,” “Holdings,” a shared geographic prefix) get reused across genuinely unrelated corporate families far more often than they would within one country’s registry, so a matching threshold that would rarely misfire on a single, well-normalized company database misfires routinely once filings, a foreign registry, and a lien database are being reconciled against each other.

Example

Due-diligence agent compiles a corporate-structure chart for "Meridian Logistics Holdings" from a mix of SEC filings, a foreign companies registry, and a UCC lien database
Target's actual filings reference a subsidiary "Meridian Logistics Holdings (Asia) Pte Ltd"
Foreign registry separately lists an unrelated company "Meridian Logistics Pte Ltd," coincidentally similar in name but with a different registration number and no filed relationship to the target
Agent's name-similarity matching links the unrelated "Meridian Logistics Pte Ltd" into the target's structure chart as a subsidiary
Buyer's risk assessment, relying on the chart, flags exposure to a litigation matter actually involving the unrelated entity, while missing that the target's real Asia subsidiary carries a different, undisclosed liability

Key Statistics

Contributing Factors

• Entity matching across filings, registry data, and lien databases is performed via name similarity rather than registration number, tax ID, or other unique identifier
• No validation step confirms a matched entity shares a registration identifier or a documented ownership filing with the target before it is added to the structure chart
• Jurisdictions or industries with high naming-convention overlap are not flagged for mandatory identifier-based verification before similarity matching is trusted

Mitigation Strategies

Prevention

1. Enforce identifier-primary entity matching with fallback detection: Implement entity resolution pipeline: (1) All entity matches must first attempt lookup via registration number/tax ID/LEI against authoritative registry (SEC EDGAR, national business registries, Refinitiv/FactSet), (2) Only if no identifier match found, apply semantic similarity scoring as secondary signal, (3) Require explicit user confirmation before including any similarity-matched entity, (4) Flag all similarity-matched entities with confidence metadata in structure chart, (5) Maintain audit trail: {matched_entity, match_method, confidence_score, override_reason}. Root cause mitigation: Prevents false positives by enforcing identifier-first methodology rather than similarity-first.

2. High-collision jurisdiction/industry flagging with mandatory verification gates: Build registry of high-collision-risk contexts: (a) jurisdictions with common naming conventions (e.g., Hong Kong holding companies), (b) industries with generic terms (e.g., “Holdings”, “Capital”, “Partners”), (c) multi-national subsidiary families where different countries use similar names. For any entity match in flagged contexts, require: secondary verification step (look up in alternative registry), document approval from compliance/legal reviewer before chart inclusion, cross-reference against filed ownership disclosures for contradictions. Root cause: Prevents matching based solely on similarity in high-risk scenarios.

3. Ownership-filing validation before structure-chart inclusion: Before adding any entity link to chart, require corroborating ownership filing from at least one independent source: (a) Specific SEC 13D/13G/Schedule 13A filing naming the relationship, (b) Foreign registry deed/certificate of incorporation naming parent, (c) UCC liens or financing statements naming control relationship, (d) Bankruptcy filings or creditor lists showing ownership. Match required filing document ID and effective date. If filing unavailable, mark entity as “[UNVERIFIED - no filed ownership document found]” and escalate to analyst. Root cause: Ensures relationships are documented, not inferred.

Detection & Response

1. Entity match audit logging with verification signals: For each entity link in structure chart, capture: entity name, target company, match method (identifier vs. similarity), match confidence score (0-1.0), corroborating ownership filing (yes/no, document ID), analyst override (yes/no, reason), effective date. Alert if: >10% of chart entities linked via similarity only (no identifier confirmation), >1 entity per chart failed verification audit, entity linked without corroborating filing. Monitor weekly: audit random 10% of finalized charts; validate each link against alternative data sources (alternative registry lookups, document re-verification).

2. Post-chart discovery reconciliation on new filings: When new corporate filings received post-chart publication, parse for entity relationship data. If new filing contradicts previous chart (e.g., “this entity is not a subsidiary of target”), trigger automated chart audit: flag affected entities, re-run identifier-based matching, generate reconciliation report. Escalate to due-diligence team if contradiction affects material findings (e.g., liability exposure calculation).

Architecture Patterns

1. Entity Resolution Pipeline with Identifier-Priority Matching: Ingestion → Parse Entity Names → Identifier Lookup (against SEC EDGAR, Refinitiv, national registries) → If match found, confirm via filing document → If no match, apply semantic similarity scoring with threshold (>0.85) → Require explicit verification/override → Add to chart with match metadata {entity_id, match_method, confidence, filing_doc_id}. Prevents similarity-only matches in primary path.

2. High-Collision Jurisdiction Registry & Verification Gate: Maintains curated list of high-collision contexts (e.g., “Hong Kong + Holding Company”, “Singapore + Capital”). When entity matches in flagged context: (a) Cross-check against alternative registries (e.g., if EDGAR lookup succeeds but foreign subsidiary suspected, verify with foreign registration authority), (b) Flag for manual compliance review, (c) Require override confirmation before inclusion. Reduces false-positive family relationships.

3. Ownership Filing Validator: Searchable index of filed ownership documents (SEC 13D/13G, foreign deeds, UCC filings, corporate bylaws naming control). On entity link: query index for corroborating filing. If found, auto-confirm link + document ID. If not found, flag for analyst verification. Maintains audit trail of all lookup attempts.

Key Metrics

Alerts & Escalation

References

• Classifying and Addressing the Diversity of Errors in Retrieval-Augmented Generation Systems
• Towards Reliable Retrieval in RAG Systems for Large Legal Datasets
• A Survey on Knowledge-Oriented Retrieval-Augmented Generation