Skip to content

AI, Retrieval, and Publication Evidence

Part VII — AI, retrieval and publication evidence

27. Neuro-symbolic execution model

27.1 Reported structured-decomposition evidence

The source baseline reviewed Sadowski and Chudziak, “Structured Decomposition for LLM Reasoning: Cross-Domain Validation and Semantic Web Integration” (arXiv:2601.01609). The reported architecture separated:

  • LLM entity identification and assertion extraction;
  • deterministic OWL/SWRL reasoning;
  • result inspection.

The reported ablation showed that the reasoner-driven condition outperformed direct LLM rule application by 9.7 F1 points overall and by 19.9 points on the clinical-trial task. The source analysis used this as evidence for the execution boundary, not as evidence about Product Holons specifically.

27.2 Product Holon application

The analogous Product Holon pattern is:

  1. resolve the product identity and retrieve authoritative data;
  2. construct a bounded graph;
  3. validate with SHACL;
  4. infer and transform with approved deterministic rules;
  5. export a target graph;
  6. let the LLM explain or compare the resulting facts.

27.3 Auditability

The architecture can retain:

  • the source record;
  • the Product Holon graph;
  • the validation report;
  • the mapping records;
  • the rule and query versions;
  • the target graph;
  • the generated narrative.

This provides a reviewable path from a natural-language response to the product facts and rules that supported it.

28. Graph-grounded retrieval

28.1 Retrieval pattern

The Product Holon is created by structural graph traversal rather than by retrieving arbitrary text chunks:

  • identifier to product record;
  • product record to GPC Brick;
  • Brick to ancestor chain;
  • Brick to Attribute Types and Values;
  • product to populated GDSN records;
  • product to parent, child or component links.

28.2 Advantages over unrestricted vector retrieval

  • controlled values are matched exactly;
  • category scope prevents irrelevant attributes;
  • multi-hop parent/child and ancestor relationships remain explicit;
  • source identifiers and provenance are preserved;
  • current resolver-linked data can be retrieved rather than relying only on a stale text index;
  • the bounded graph fits the intended reasoning context.

28.3 Distinction from graph construction from text

Microsoft GraphRAG work commonly constructs graphs from unstructured text with LLM assistance. The Product Holon architecture starts from governed semantic and product-data sources. An LLM may navigate or narrate the graph but does not define the authoritative classification or code values.

29. Resolver and publication evidence

29.1 Standards-backed mechanisms

The source baseline identified these GS1-Conformant Resolver mechanisms as relevant:

  • Accept-based media-type negotiation;
  • JSON-LD linkset responses;
  • linkType selection;
  • Accept-Language and context differentiation;
  • batch/lot and serial qualifiers;
  • links inherited or selected according to resolver rules.

29.2 Proposed Product Holon use

A conformant resolver could link a GTIN to:

  • a human-readable product page;
  • an approved Web Vocabulary/schema.org JSON-LD representation;
  • a nutrition resource;
  • a product-safety or regulatory resource;
  • an authenticated partner representation;
  • a Product Holon API.

A Product Holon-specific link type or authenticated audience convention is proposed, not established by the evidence as existing GS1 guidance.

29.3 Live versus cached construction

Two implementation modes are consistent with the architecture:

  • pre-build and cache validated Product Holons and their views;
  • construct on request from current authoritative data, then validate and cache.

The source evidence did not benchmark either mode. Production choice requires latency, freshness and governance analysis.