CRASM-Gate: Deterministic-First Constraint- and Role-Aware Semantic Mapping with Selective Model Assistance Across Heterogeneous Industrial Standards

Kabeh Mohsenzadegan; Vahid Tavakkoli; Kyandoghere Kyamakya

cs.AISep 25, 2026

CRASM-Gate: Deterministic-First Constraint- and Role-Aware Semantic Mapping with Selective Model Assistance Across Heterogeneous Industrial Standards

Authors: Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya

Organizations: Institue of Smart System Technologies, University of Klagenfurt, 9020 Klagenfurt, Austria · Faculté Polytechnique, Université de Kinshasa, Kinshasa, Democratic Republic of the Congo

Abstract

Industrial standards often encode the same engineering concept through incompatible hierarchies, identifiers, roles, and structural constraints, so the nearest lexical or embedding match can still be technically inadmissible. This article presents CRASM, a deterministic constraint- and role-aware semantic mapping method, and CRASM-Gate, its selectively model-assisted extension. The framework separates standard-specific canonicalization, bounded retrieval, deterministic rules, destination-versus-origin role interpretation, semantic and structural ranking, ambiguity refusal, and target validation. CRASM-Gate adds a confidence/disagreement gate that may invoke a candidate-constrained large language model, while final authority remains with deterministic validation. A controlled artifact covers six directed industrial-standard pairs, three difficulty levels, and ten configurations, yielding 14,400 sample-level decisions. With a fixed local model endpoint, CRASM-Gate reaches mean F1 0.9938 and an implemented structural-validity rate of 1.0000; deterministic CRASM reaches 0.9931 without generative-model calls; and the model-only baseline reaches 0.5347. Relative to the model-only baseline, CRASM reduces top-1 errors from 670 to 10 while exhibiting 0.0688 s/sample rather than 24.5263 s/sample observed latency. CRASM-Gate improves CRASM by one additional correct decision out of 1,440, with observed latency increasing to 15.7666 s/sample. The results support a deterministic-first interoperability architecture in which model assistance is optional, measurable, candidate-bounded, and unable to bypass structural validation.

Figures & tables

Dimension	Preliminary ISynKGR	This article
Primary contribution	Reproducible benchmark framework	CRASM mapping method + CRASM-Gate framework
Model role	Component of hybrid benchmark pipeline	Optional bounded uncertainty resolver
Destination/origin roles	Not explicit	Explicit deterministic interpretation
Final authority	Hybrid pipeline evaluation	Deterministic validation contract
Structured generation	Lightweight/bounded use	Strict typed JSON Schema + downstream validation
Experimental emphasis	Eight scenario families; benchmark comparison	Method decomposition, context-conflict difficulty, cost/quality, artifact audit

Table 1: Scientific separation from the preliminary ISynKGR benchmark paper.

Paradigm	Typical strength	Typical limitation for this task	CRASM-Gate relation
Model-driven transformation	Explicit metamodel and transformation semantics	High engineering cost when heterogeneous standards expose different role/context conventions	Retains deterministic adapters and validation while allowing evidence fusion
Ontology matching	Mature lexical, structural, and logical alignment	Usually assumes ontology entities rather than protocol-qualified industrial instance paths	Adopts multi-evidence ranking and refusal; external matchers remain future baselines
Knowledge-graph integration	Relational context, provenance, explainability	Graph similarity alone does not define target-admissibility authority	Uses graph/retrieval evidence as input, not as final decision authority
LLM-supported matching	Flexible language and metadata interpretation	Output can be structurally invalid, unbounded, or difficult to reproduce	Restricts model calls to a gate, a bounded candidate vocabulary, typed schema, and downstream validator
CRASM	Deterministic, auditable role/context reasoning	Cannot recover information absent from candidates or metadata	Default decision path
CRASM-Gate	Deterministic core plus selective language-model evidence	Requires profiling of gate utilization and external validation	Complete framework proposed in this article

Table 2: Conceptual positioning of CRASM/CRASM-Gate relative to representative interoperability paradigms.

Representation	Typical mapping unit	Ambiguity when context is ignored	Evidence used by CRASM
OPC UA	Object/variable path	Same label under different namespace, object, or instance context	Qualified path, parent/object context, datatype, unit, instance identifiers
AAS	Asset/submodel element	Same property name can occur under different assets or submodels	Asset/submodel context, element role, semantic label, datatype, unit
IEEE 1451	TEDS/channel element	Similar measurements can occupy different transducer/channel contexts	Channel identity, signal family, datatype, unit, parent context
IEC 61499	Function-block I/O	Measurement names can be confused with control/interface artifacts	Device/resource/FB path, I/O role, datatype, unit, signal family
ISO 15926-style	Class/relation projection	Label similarity can ignore relation direction or process-plant context	Class/relation tokens, target path, signal family, contextual identity

Table 3: Representative semantic ambiguities exercised by the benchmark. The final column describes implemented checks, not full standards conformance.

Figure 1: Vector-rendered CRASM-Gate control pipeline with orthogonal decision and feedback paths. CRASM provides the deterministic route from canonicalization through role-aware ranking and validation. CRASM-Gate can request bounded model evidence only through the gate, and every model-supported proposal returns to the same deterministic validator before acceptance.

Audit question	Recorded evidence
What was mapped?	source/target protocol-qualified paths
What alternatives existed?	bounded candidate list and ranking trace
Why did a candidate win?	score breakdown and support sources
Was a model used?	gate reason, invocation trace, model evidence
Why was output accepted?	validation outcome, rejection reasons, no-match/repair trace

Table 4: Traceability questions and corresponding CRASM-Gate evidence.

Publication name	Frozen ID
CRASM-Gate	full_framework
CRASM	agentic_interpreter
LLM-only	llm_only
RAG-only	rag_only
Embedding Similarity	embedding_similarity
Semantic Graph Calibrated	semantic_graph_calibrated

Table 5: Publication names and frozen artifact identifiers.

Configuration	F1	Pair SD	Validity	Observed latency (s/sample)	Correct / 1440
CRASM-Gate	0.9938	0.0153	1.0000	15.7666	1431
CRASM	0.9931	0.0170	1.0000	0.0688	1430
LLM-only	0.5347	0.1141	1.0000	24.5263	770
Legacy w/o Rules $\dagger$	0.5056	0.1035	1.0000	11.0734	728
Legacy w/o Retrieval $\dagger$	0.4750	0.1722	0.9993	2.0750	684
Semantic Graph Calibrated	0.3333	0.0000	1.0000	0.0753	480

Table 6: Aggregate frozen-artifact results across six directed mapping pairs. SD is descriptive variation across pair-level values, not repeated-seed uncertainty. Rows marked

\dagger

are legacy compound ablations and are not interpreted as one-component causal estimates.

Figure 3: Aggregate F1 by frozen-artifact configuration. CRASM is almost indistinguishable from CRASM-Gate in mapping quality. Legacy compound ablations are shown for provenance but are excluded from isolated component claims.

Quantity	CRASM-Gate $-$ CRASM
F1	+0.00069
Correct decisions / 1440	+1
Validity	+0.0000
Latency / sample	+15.6978 s
Latency ratio	229.3 $\times$

Table 7: Observed marginal effect of CRASM-Gate over deterministic CRASM.

Method	Correct	Errors	Latency (s/sample)
CRASM-Gate	1431	9	15.7666
CRASM	1430	10	0.0688
LLM-only	770	670	24.5263

Table 8: Error accounting for the three principal operating points.

Mapping pair	Configuration	F1	Latency/sample (s)
AAS $\rightarrow$ OPC UA	CRASM-Gate	0.9625	31.0413
	CRASM	0.9583	0.0965
	LLM-only	0.3333	31.2947
IEC 61499 $\rightarrow$ IEEE 1451	CRASM-Gate	1.0000	37.7348
	CRASM	1.0000	0.1128
	LLM-only	0.6750	34.3918

Table 9: Pair-level F1 and observed latency for CRASM-Gate, CRASM, and LLM-only.

Figure 4: Accuracy by controlled difficulty. Medium and hard samples require contextual-role disambiguation rather than nearest-similarity selection.

Figure 5: Complementary diagnostics. (a) Reported confidence calibration error. (b) Retrieval recall for configurations sharing the retrieval layer. Retrieval recall is substantially below final CRASM quality, showing that downstream role interpretation and validation materially change the decision.

Configuration	Cal. error	$\mathrm{R}@1$	$\mathrm{R}@5$
CRASM-Gate	0.0070	0.2618	0.6597
CRASM	0.0102	0.2618	0.6597
LLM-only	0.3251	0	0
RAG-only	0.1394	0.2618	0.6597
Rules-only	0.2901	0	0

Table 10: Selected diagnostic metrics averaged across six pairs.

Figure 6: Quality–latency operating points. The horizontal axis is logarithmic. Timings are descriptive artifact measurements, not hardware-normalized inference benchmarks.

Profile	Intended role	Required observability	Principal caveat
CRASM	Default deterministic mapping	candidate list, feature scores, rules, role interpretation, validator trace	quality depends on available candidates, metadata, and implemented validators
CRASM-Gate	Selective escalation for unresolved cases	all CRASM evidence plus gate reason, model identifier, prompt/schema version, invocation and token/latency record	selectivity and cost must be demonstrated with a dedicated gated run
LLM-only	Experimental baseline	prompt, model output, candidate vocabulary	lower quality in this artifact and insufficient authority controls for deployment
No-match / manual review	Safe fallback when evidence is insufficient	ambiguity reason and rejected candidates	requires downstream review workflow

Table 11: Recommended operational interpretation of the evaluated configurations.

Semantic Mapping

Semantic mapping links observations and data to meaningful concepts, categories, entities, and relations for machine reasoning and action.

See all 93 papers

Constraint-Guided Enterprise Data Mapping with Large Language Models
Aug 25, 2026Sebastian Monka, Pramod Anantharam, Thien Vo Minh +1Entity Resolution Semantic Mapping
Bosch Center for Artificial Intelligence, Renningen, Germany · Bosch Center for Artificial Intelligence, Pittsburgh, USA · Robert Bosch GmbH, Ho Chi Minh City, Vietnam
Harness as an Asset: Enforcing Determinism via the Convergent AI Agent Framework (CAAF)
Apr 18, 2026Tianbao ZhangSafety-Critical Control LLM Guardrails
Independent Researcher
Talk Freely, Execute Strictly: Schema-Gated Agentic AI for Flexible and Reproducible Scientific Workflows
Mar 6, 2026Joel Strickland, Arjun Vijeta, Chris Moores +7LLM Agent Orchestration Scientific Workflow Automation
Intellegens, The Studio, Chesterton Mill, Cambridge, CB4 3NP, UK · Independent Researcher · Cavendish Laboratory, J.J. Thomson Avenue, Cambridge, UK

Aug 25, 2026J/K move · Enter open · S save

Constraint-Guided Enterprise Data Mapping with Large Language Models

Sebastian Monka, Pramod Anantharam, Thien Vo Minh, Lavdim Halilaj

Bosch Center for Artificial Intelligence, Renningen, Germany · Bosch Center for Artificial Intelligence, Pittsburgh, USA · Robert Bosch GmbH, Ho Chi Minh City, Vietnam

Entity Resolution Semantic Mapping

Open paper AskPDF arXiv

Enterprise entity alignment must handle semi-structured records, implicit attributes, and unit or granularity mismatches. Manual matching is still common in practice, but does not scale as schemas and providers evolve. LLM-only matching improves semantic recall, yet can violate structural and physical invariants, producing fluent yet operationally invalid correspondences. We propose constraint-guided mapping (CGM), a neuro-symbolic method with three stages: (i) schema-grounded admissibility constraints with metadata mc = <tau_c, delta_c>, where tau_c denotes the constraint type and delta_c provides executable relation and normalization logic; (ii) constraint-restricted candidate generation with cascade relaxation to guarantee a nonempty feasible set under noise; and (iii) neural ranking with bounded LLM disambiguation restricted to that feasible set. Methodologically, constraints operate as hypothesis-space operators rather than post-hoc validators, enabling controlled degradation under relaxation and auditable, human-guidable decisions. On a controlled structural-decoy benchmark, hard admissibility shrinks the candidate space by ~480x without dropping the GT, and a layer-by-layer ablation shows this gate, not the LLM, is the decisive lift (F1 0.08 to 0.66). The benefit is model-independent and adds no extra inference cost: a small model with constraints matches a frontier LLM used without them at ~28x lower cost. The method, not a single tuned configuration, transfers across seven enterprise makes (macro F1 0.70), each under its own automatically discovered, expert-refinable constraints, and lowers expert effort by ~7x versus spreadsheet workflows. Public Valentine results add an external ranking sanity check and mark the boundary: constraints should be hard only where structural invariants are match-determining.

Abstract

Semantic Mapping

Explore similar work

Constraint-Guided Enterprise Data Mapping with Large Language Models

Harness as an Asset: Enforcing Determinism via the Convergent AI Agent Framework (CAAF)

Talk Freely, Execute Strictly: Schema-Gated Agentic AI for Flexible and Reproducible Scientific Workflows