The simulation numbers were not the weak point. In 136 accepted comparisons (132 external references and 4 other Studio meshes), the median difference was 0.13%.
The weak point was the writing about them.
NovaSolver ran a small pipeline in November 2026. Claude Code wrote Python MCP client scripts that drove NovaSolver Studio (an open-source FEM/CFD app) through its MCP server, computed reference values, and then drafted engineering articles from the logged results. The old seven sets of drafts went through these review stages:
A separate Claude agent fact-checked it against the logs and the source files (sets 4, 5 and 7 only)
Codex (OpenAI) reviewed it read-only, after the fact-check
Reflection status was recorded: applied, partly applied, unapplied or unknown
Across seven sets of drafts, the reviewers found 135 review findings , including errors and optional improvements.…