LLMFuzz
Deterministic testing tool for LLM-targeted workflows, with explicit inputs, captured outputs, and inspectable run artifacts.
Overview
LLMFuzz is a bounded fuzz / eval / triage tool for model-facing systems and LLM-agent-style targets. It executes one case or a small campaign, records deterministic artifacts under a stable work root, and makes failures reproducible enough to inspect directly.
Architecture
The system is organized around explicit test cases, command-driven target execution, captured outputs, and reproducible on-disk artifacts. Each run produces a narrow, inspectable evidence set so evaluation and triage can be repeated without relying on inferred context or hidden state.
Scope
v0.1 focuses on a minimal OSS runtime: command invocation via argv (no shell), deterministic artifact layout, single-run execution, small campaigns, deterministic evaluation, and campaign triage. The design intentionally favors bounded behavior and stable contracts over broad runtime abstraction.
Status
ACTIVE_DEVELOPMENT (v0.1.1)