[Harness] Detecting Numerical instability in QMC codes

[hz-xiaxz]

Primary Harness Type

Test / evaluation harness - checks, examples, and review loops

Project Goal

In Quantum Monte Carlo codes, numerical instabilities often kicks in when dealing with ill-conditioned green function matrices etc. However, debugging the origin of them can be time-consuming, and due to the randomness nature of Monte Carlo Sampling, the overflows events can be rear.

I haven't got clear picture in mind how to build this, but basically I want Agents to place breakpoints in potentially overflow variables, detecting overflow, inferred from function calling that how this variable changes, and dates back to the origin overflow code line, like SuperMario dealing with plumbing leaking.

Planned Components

• Knowledge base, examples, or domain notes
• Skill file or agent instructions
• Command line scripts, Makefile targets, or a small CLI
• MCP server or external service integration
• Tests, checks, evaluation cases, or review prompts
• Setup guide, tutorial, or onboarding flow
• Demo repo, sample data, or reproducible example

Before-Event Plan

OpenCode with suspicious GPT5.4 API (from a transfer site). Quantum Monte Carlo Code (DQMC, VMC).

Target AI Coding Tool

OpenCode

Participation Readiness

• I have access to a Bash or Zsh terminal. Windows users can use WSL 2.
• I have installed Codex CLI, Claude Code, or another AI coding tool.
• I have tried logging in to my AI coding tool before the event.
• I have a concrete workflow, sample input, or starter repo to work on.

Team / Contact

No response