Case Studies

Real-world examples of SciAgent reproducing scientific research from peer-reviewed publications.


Overview

These pages are written as narrative summaries of benchmarked reproductions. Each one answers four questions:

  • what scientific task was attempted
  • why that task matters for understanding SciAgent
  • what happened in the audited run
  • what the audit layer revealed beyond a simple pass/fail result

Throughout these pages, cc-bare means the Claude Code baseline without access to SciAgent’s registry or compute subagents.

The full benchmark bundle is not shipped in the main repo. These case studies are meant to stand on their own as reader-facing summaries rather than file-by-file archive guides.

Available Studies

Study Domain Benchmark criterion What it shows
AR Waveguide Metasurface Photonics / RCWA MFE >= 0.25 Long-horizon scientific workflow and why verifier recursion matters
BRCA1 Fitness-Structure Analysis Bioinformatics Mapping rate >= 0.95 A correct result with a caught scope downgrade in the audit trail
Datacenter Temperature with OpenFOAM CFD / OpenFOAM Mean temperature in 294-298 K Clean end-to-end cloud execution with a strong verifier story
Digital IC Synthesis EDA (coming soon) - OpenROAD-based benchmark in progress

Table of contents