FairStack AI: AI-Enabled Orchestration of a Scientific Data Software Stack

Scientific data centers must coordinate many specialized tools to complete submission records, build data-asset inventories, and assess data quality. Standardized software stacks provide these tools, but the workflow still requires substantial manual interpretation, cross-system coordination, and compliance review. We present FairStack AI V1.0, an orchestration framework that encodes the domain specification as versioned rule artifacts and enforces three runtime constraints—Gate, Contract, and Abstain—across a three-stage submission lifecycle. The order of operations is fixed by the specification, so the agent determines content rather than sequence; operations not permitted at a stage are refused, not attempted, and ambiguous or compliance-critical values go to a human operator. The framework is implemented on FairStack and evaluated on all 200 de-identified National Key Research and Development Program projects. Data-asset inventories attain a graded match score of 0.803 (precision 0.864, recall 0.751) against the reference plans and recover 3927 of the 5029 reference entries; omissions outnumber unsupported assertions by about four to one, so the residual error is one of coverage rather than of correctness. Contract compliance holds at runtime: 96.3% of entries carry all eleven prescribed keys under their prescribed spellings, free-text names reach 0.754 average cosine similarity with the human reference, and the top-level container shape is identical in all 200 runs. On the machine side a project completes in a median of 26.7 min over 1.09 agent rounds. The results show that rule constraints reduce repetitive work while a human operator retains every ambiguous or compliance-critical decision, and that an externalized specification makes both the shape and the completeness of the output mechanically checkable against the source.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-16
DOI
https://doi.org/10.3390/app162211350
Primary Topic
Research Data Management Practices
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

FairStack AI: AI-Enabled Orchestration of a Scientific Data Software Stack

Cifeng Wang, Junlei Duan, Lei Jiang, Changfa Lu et al.
2 links
Applied Sciences
Research Data Management Practices
article

FairStack AI: AI-Enabled Orchestration of a Scientific Data Software Stack

Cifeng Wang, Junlei Duan, Lei Jiang, Changfa Lu, Zhaochen Wu
article en

Abstract

Scientific data centers must coordinate many specialized tools to complete submission records, build data-asset inventories, and assess data quality. Standardized software stacks provide these tools, but the workflow still requires substantial manual interpretation, cross-system coordination, and compliance review. We present FairStack AI V1.0, an orchestration framework that encodes the domain specification as versioned rule artifacts and enforces three runtime constraints—Gate, Contract, and Abstain—across a three-stage submission lifecycle. The order of operations is fixed by the specification, so the agent determines content rather than sequence; operations not permitted at a stage are refused, not attempted, and ambiguous or compliance-critical values go to a human operator. The framework is implemented on FairStack and evaluated on all 200 de-identified National Key Research and Development Program projects. Data-asset inventories attain a graded match score of 0.803 (precision 0.864, recall 0.751) against the reference plans and recover 3927 of the 5029 reference entries; omissions outnumber unsupported assertions by about four to one, so the residual error is one of coverage rather than of correctness. Contract compliance holds at runtime: 96.3% of entries carry all eleven prescribed keys under their prescribed spellings, free-text names reach 0.754 average cosine similarity with the human reference, and the top-level container shape is identical in all 200 runs. On the machine side a project completes in a median of 26.7 min over 1.09 agent rounds. The results show that rule constraints reduce repetitive work while a human operator retains every ambiguous or compliance-critical decision, and that an externalized specification makes both the shape and the completeness of the output mechanically checkable against the source.

Applied SciencesVol. 16(22)
Chinese Academy of Sciences (CN), Computer Network Information Center (CN)
Chinese Academy of Sciences
Research Data Management Practices
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