Evidence-aware retrieval and bounded orchestration for heterogeneous materials R&D

Abstract Materials research and development (R&D) questions often require evidence distributed across measurements, process histories, and revision-controlled investigations. Relevance ranking alone cannot determine whether that evidence is sufficient, current, or numerically interpretable. We present an evidence-aware retrieval methodology with three elements: provenance-preserving representations, a question-specific evidence ledger for bounded orchestration, and layered evaluation of retrieval, claim admission, and response behavior. An IN718 additive-manufacturing testbed combines public sources from the National Aeronautics and Space Administration (NASA) and the National Institute of Standards and Technology (NIST) with an authored benchmark of 60 questions in 30 paired intents. Hybrid retrieval increased the top-five hit rate from 0.7885 to 0.9231, whereas reranking the Hybrid candidates reduced it to 0.8654: component effects are non-monotonic. Evidence-conditioned control exposed unresolved requirements but did not guarantee closure: every integrated case with multiple retrievals retained a residual. The fixed integrated system, executed once on 24 evaluation questions with prior component-level exposure, answered one abstention-labeled question and withheld three answerable ones, while all emitted citation identifiers were valid and all rule-rejected claims were excluded. Revision-controlled investigation examples and numerical probes show how historical authority, recorded event status, and calculation provenance constrain what an answer can establish. The methodology is inspectable: strong retrieval scores do not conceal response or semantic failures. The labels are authored, not expert-adjudicated; independent adjudication and an untouched evaluation cohort are the next validation axes.

Authors

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-11746659/v1
Primary Topic
Information Retrieval and Search Behavior
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Evidence-aware retrieval and bounded orchestration for heterogeneous materials R&D

Seifallah Fetni, Theodore Tyrikos‐Ergas
4 links
Research Square
Information Retrieval and Search Behavior
preprint

Evidence-aware retrieval and bounded orchestration for heterogeneous materials R&D

Seifallah Fetni, Theodore Tyrikos‐Ergas
preprint en

Abstract

Abstract Materials research and development (R&D) questions often require evidence distributed across measurements, process histories, and revision-controlled investigations. Relevance ranking alone cannot determine whether that evidence is sufficient, current, or numerically interpretable. We present an evidence-aware retrieval methodology with three elements: provenance-preserving representations, a question-specific evidence ledger for bounded orchestration, and layered evaluation of retrieval, claim admission, and response behavior. An IN718 additive-manufacturing testbed combines public sources from the National Aeronautics and Space Administration (NASA) and the National Institute of Standards and Technology (NIST) with an authored benchmark of 60 questions in 30 paired intents. Hybrid retrieval increased the top-five hit rate from 0.7885 to 0.9231, whereas reranking the Hybrid candidates reduced it to 0.8654: component effects are non-monotonic. Evidence-conditioned control exposed unresolved requirements but did not guarantee closure: every integrated case with multiple retrievals retained a residual. The fixed integrated system, executed once on 24 evaluation questions with prior component-level exposure, answered one abstention-labeled question and withheld three answerable ones, while all emitted citation identifiers were valid and all rule-rejected claims were excluded. Revision-controlled investigation examples and numerical probes show how historical authority, recorded event status, and calculation provenance constrain what an answer can establish. The methodology is inspectable: strong retrieval scores do not conceal response or semantic failures. The labels are authored, not expert-adjudicated; independent adjudication and an untouched evaluation cohort are the next validation axes.

Research Square
Information Retrieval and Search Behavior
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.

Laika access coming soon