Entity Resolution Strategy Selector

Entity Resolution Strategy Selector

Overview

Entity resolution determines when two data records refer to the same real-world entity — the cornerstone that lets an agent maintain a coherent worldview across fragmented systems. If the agent cannot decide when two references are the same entity, its whole reasoning framework collapses: the graph either conflates distinct entities or fragments a single one.

The chapter draws one decisive distinction, and this skill turns it into a selection:

• Evidence-based resolution examines specific features, applies

domain-specific matching rules, and builds a case from concrete evidence. It is deterministic (same input, same output), explainable (every match cites which features drove it and their scores), culturally robust (explicit rules handle Arabic / Chinese / Russian naming), and calibrated (confidence reflects actual match accuracy).

• Generalization-based AI (an LLM) infers from statistical similarity

learned in training. It is nondeterministic, its explanations are post-hoc rationalizations, it breaks on non-Western names, and its confidence is not tied to accuracy.

Evidence-based wins for identity, compliance, high-stakes, and adversarial work. The sharp case is channel separation: a money launderer appears as Bob Jones, then Bob R. Smith II at the same address with different phone formatting, then Robert Smith Jr. elsewhere with overlapping contact details — each variation engineered to pass fuzzy filters while looking distinct. Simple string matching fails catastrophically; what wins is consolidating fragmented identities on evidence from multiple overlapping features.

The selector scores a requirement profile across the six factors the chapter names (highstakes, adversarial channel-separation, explainability, determinism, culturalvariation, trainingexamples) and returns evidence-based, generalization-AI, or a hybrid (LLM for cheap candidate generation, evidence-based for the auditable final decision). The matcher makes the trade-off concrete: it scores name/address/phone similarity, aggregates to an explainable confidence with per-feature evidence metadata (the chapter's "89% because NAME 87%, ADDRESS 100%, PHONE 95%"), classifies the graph edge (RESOLVED / POSSIBLYRELATED / DISCLOSED), and flags the three edge cases that require domain and cultural knowledge.

When to Use

• Standing up entity resolution for an agent knowledge graph and deciding the

resolution strategy

• Justifying evidence-based vs LLM matching in a design doc for identity,

compliance, or fraud work

• Auditing a proposed merge: what confidence, on what evidence, and is it an

edge case?

Phrases: "entity resolution", "record linkage", "deduplication", "are these the same entity", "evidence-based vs LLM matching", "channel separation", "RESOLVED edge", "match confidence explainability".

When NOT to Use

• The extraction stage. Producing the records (LLM triple extraction,

iText2KG / RAKG / ATOM construction) is upstream; this skill decides identity once records exist.

• Picking a specific ER product. This chooses the strategy and gives a

reference matcher, not Senzing-vs-build. Vendor choice is downstream.

• Arity-2 relationship modeling. Whether a fact needs a hyperedge is the

graph-model-selector skill; this one is about identity, not structure.

• A domain with no identity stakes. Genuinely low-stakes fuzzy dedup with

abundant labeled examples and no compliance need can use generalization-AI — the selector will say so.

Process

| Step | Input | Action | Output | Verification | |------|-------|--------|--------|--------------| | 1 | Requirement profile (6 weights 0..3) | lib.score_strategies(profile) | [(strategy, score), ...] sorted desc | weights * factor-affinities, descending order | | 2 | Same | lib.recommend_strategy(profile) | {recommended, scores, rationale, hybrid_recommended} | evidence-based wins high-stakes/adversarial; generalization-AI wins example-rich low-stakes; hybrid when mixed | | 3 | Two records {name, address, phone} | lib.resolve_match(a, b, weights) | {confidence, edge_type, features_used, evidence[]} | confidence == sum of per-feature contributions; evidence names which features drove it | | 4 | A match confidence (or declared=True) | lib.classify_edge(conf, declared) | {edge_type, confidence, reason} | >=0.85 RESOLVED, >=0.5 POSSIBLYRELATED, else NOMATCH; declared -> DISCLOSED | | 5 | Two records | lib.flag_edge_cases(a, b) | list of {case, warning, action} | catches honorific-same-entity, near-identical-different-entity, address-same-location |

Rationalizations

| Agent rationalization | Documented rebuttal | |------------------------|--------------------| | "Just ask an LLM if these two records are the same person." | For identity, compliance, and fraud the LLM's answer is nondeterministic, its explanation is a post-hoc rationalization, and its confidence does not reflect accuracy. When a wrong merge denies a mortgage or voids a ballot, the org ends up in court. Evidence-based scoring is deterministic and cites the evidence. | | "Fuzzy string matching on each field is good enough." | It fails catastrophically under channel separation — variations are engineered to defeat per-attribute fuzzy filters. The win is channel consolidation from evidence across multiple overlapping features, not any single field. | | "These names are almost identical, so merge them." | flag_edge_cases catches near-identical-different-entity: John R Smith vs John E Smith differ by one letter that may mean father-and-son. High string similarity is a reason for caution, not a merge; the action is do-not-merge-without-evidence. | | "These names look totally different, so they're different people." | Wrong the other way: al-Hajj Abdullah Qardash and Abu Abdullah Qardash bin Amir look different as strings but are the same person once honorifics and Arabic naming conventions are handled. Cultural robustness is exactly where generalization-AI breaks. | | "The selector said hybrid — pick one and move on." | Hybrid is a real pattern: generalization-AI cheaply proposes candidate pairs (blocking), evidence-based makes the final auditable decision. Record it as a conscious trade-off, not a non-decision. | | "Confidence is confidence; I don't need the per-feature breakdown." | Evidence metadata (which features matched and their scores) is what makes the match explainable and auditable. "89% because NAME 87%, ADDRESS 100%, PHONE 95%" is defensible; a bare 89% is not. |

Red Flags

• All six profile weights set to 3. You have not prioritized. If every

concern is critical the selector degenerates to raw affinity sums — re-interview the use case.

• Generalization-AI recommended while adversarial or high-stakes is weighted.

Mismatch: re-check the weights. Generalization-AI scores 0 on those axes by construction.

• RESOLVED edge created on a single feature. Channel-separation resistance

comes from overlapping features. A match driven by name alone is fragile — demand corroborating evidence.

• Merging on a near-identical pair without checking edge cases. Run

flag_edge_cases first; a one-character difference can be a distinct entity.

• Trusting an LLM confidence score as calibrated. It reflects statistical

regularity, not match accuracy — do not threshold merges on it.

Non-Negotiable Verification

1. Run the benchmark battery. python cli.py benchmark must report 10/10:

• adversarial high-stakes picks evidence-based; example-rich low-stakes picks

generalization-AI; mixed flags a hybrid

• identical records resolve at 1.0 and classify RESOLVED
• resolve_match emits per-feature evidence and the confidence equals the

summed contributions

• edge thresholds hold and declared=True overrides to DISCLOSED
• all three edge cases (honorifics / near-identical / address) fire on the

chapter's examples

1. Run the scenario. python cli.py scenario fraud-channel-separation

recommends evidence-based and consolidates the engineered aliases with per-feature evidence.

1. Verify CLI help. python cli.py --help exits 0 and prints this

SKILL.md description (so any harness can discover the skill from --help).

Security Posture

• Prompt injection. The records under comparison are untrusted by

construction - channel separation IS adversarial input, fields engineered to force or dodge a merge. The matcher only computes similarity scores over field values; it never executes or interpolates them, so an attack can at most bias one match - which is why RESOLVED demands overlapping evidence, never a single feature.

• Data exfiltration. Records carry PII (names, addresses, phones). No

network calls, no file writes; PII surfaces only in the evidence output on stdout, and the caller owns where that report flows.

• Privilege escalation. No shell invocation, no eval, no dynamic import. A

RESOLVED edge is a data assertion, not an identity credential - downstream systems must not grant access on a merge without independent verification.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly, by Anthony Alcaraz and Sam Julien) Ch3 — Knowledge Representation, section "Entity Resolution: The Foundation of Agent Knowledge" and its subsections: "Why traditional approaches fail" (channel-separation fraud), "Evidence-based resolution vs generalization-based AI" (the deterministic / explainable / culturally-robust / calibrated distinction and the "89% because NAME 87%, ADDRESS 100%, PHONE 95%" example), "Entity resolution as graph building blocks" (RESOLVED / POSSIBLY_RELATED / DISCLOSED edge types and evidence metadata), and "Edge cases" (honorific same-entity, near-identical different-entity, address same-location). The feature-scoring, edge classification, and edge-case detection here are a self-contained stdlib reference implementation of those ideas, not production ER tooling.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AnthonyAlcaraz
• Source: AnthonyAlcaraz/agentic-graph-rag-skills
• License: MIT
• Homepage: https://www.oreilly.com/library/view/agentic-graphrag/9798341623163/

Install and usage instructions live in the source repository linked above.