Evidence-Gated Agent Architecture (EGAA): Evidence Sufficiency, Policy Hard-Gating, Plan Validation, and Result Grounding for Reliable Retrieval-Augmented Tool Use

Retrieval-augmented generation (RAG) reduces factual errors by giving language models access to external information, but retrieval relevance alone does not establish that an agent has sufficient evidence to justify a tool call. This gap becomes safety-critical when retrieved context is used to construct executable plans for enterprise or ERP systems. We propose the Evidence- Gated Agent Architecture (EGAA), a pre-generation and pre-execution control architecture that separates evidence sufficiency from authorization and execution validity. EGAA introduces an Evidence Sufficiency Gate that evaluates semantic relevance, atomic-fact coverage, contradiction/consistency, schema compati- bility, freshness, entity resolution, risk, and security constraints. Hard constraints such as authorization, tenant isolation, authentication state, and tool safety are enforced independently from probabilistic evidence scores. A structured Evidence Contract carries the gate’s decision to the planner, while a Plan Validator checks every proposed tool call against explicit evidence references and the policy context. After execution, a Result Validator grounds generated claims against the actual tool response. We formalize the gate as a hy- brid hard/soft decision function, provide inference and validation algorithms, define an ERP-oriented threat model, and propose an evaluation protocol based on task success, unsupported-action rate, policy-violation rate, evidence coverage, clarification quality, grounding error rate, abstention, and latency. The architecture is deliberately presented as a research proposal: no new empirical results are claimed in this manuscript. The work is motivated by and explicitly extends the pre-generation trust-decision perspective of Zhang et al.’s 2026 Memory Decision Layer (MDL) preprint, but moves the unit of decision from generic memory trust toward evidence sufficiency for a specific action.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-21
DOI
https://doi.org/10.5281/zenodo.22871292
Primary Topic
Access Control and Trust
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

Evidence-Gated Agent Architecture (EGAA): Evidence Sufficiency, Policy Hard-Gating, Plan Validation, and Result Grounding for Reliable Retrieval-Augmented Tool Use

Ezz eldeen Abdelrahman Alaa
Zenodo (CERN European Organization for Nuclear Research)
Access Control and Trust
preprint

Evidence-Gated Agent Architecture (EGAA): Evidence Sufficiency, Policy Hard-Gating, Plan Validation, and Result Grounding for Reliable Retrieval-Augmented Tool Use

Ezz eldeen Abdelrahman Alaa
preprint en

Abstract

Retrieval-augmented generation (RAG) reduces factual errors by giving language models access to external information, but retrieval relevance alone does not establish that an agent has sufficient evidence to justify a tool call. This gap becomes safety-critical when retrieved context is used to construct executable plans for enterprise or ERP systems. We propose the Evidence- Gated Agent Architecture (EGAA), a pre-generation and pre-execution control architecture that separates evidence sufficiency from authorization and execution validity. EGAA introduces an Evidence Sufficiency Gate that evaluates semantic relevance, atomic-fact coverage, contradiction/consistency, schema compati- bility, freshness, entity resolution, risk, and security constraints. Hard constraints such as authorization, tenant isolation, authentication state, and tool safety are enforced independently from probabilistic evidence scores. A structured Evidence Contract carries the gate’s decision to the planner, while a Plan Validator checks every proposed tool call against explicit evidence references and the policy context. After execution, a Result Validator grounds generated claims against the actual tool response. We formalize the gate as a hy- brid hard/soft decision function, provide inference and validation algorithms, define an ERP-oriented threat model, and propose an evaluation protocol based on task success, unsupported-action rate, policy-violation rate, evidence coverage, clarification quality, grounding error rate, abstention, and latency. The architecture is deliberately presented as a research proposal: no new empirical results are claimed in this manuscript. The work is motivated by and explicitly extends the pre-generation trust-decision perspective of Zhang et al.’s 2026 Memory Decision Layer (MDL) preprint, but moves the unit of decision from generic memory trust toward evidence sufficiency for a specific action.

Zenodo (CERN European Organization for Nuclear Research)
Peace, Justice and strong institutions
Access Control and Trust
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Evidence-Gated Agent Architecture (EGAA): Evidence Sufficiency, Policy Hard-Gating, Plan Validation, and Result Grounding for Reliable Retrieval-Augmented Tool Use — Ezz eldeen Abdelrahman Alaa · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS