If you discover a security issue in AVA — a model serving exploit, an unsafe tool path (e.g. python execution sandbox escape), credential leakage, or a supply-chain risk in the build/CI pipeline — please do not open a public GitHub issue.

Instead, email lancearmour24200@gmail.com with:

1. A clear description of the issue
2. Steps to reproduce
3. The affected component (training script, eval harness, MCP tool, GGUF build, etc.)
4. Optional: a proposed fix or mitigation

You should expect an acknowledgement within 7 days and a substantive response within 30 days.

In scope:

• Code in this repository (src/, scripts/, experiments/, site/, CI workflows)
• Build pipelines (GitHub Actions, GGUF conversion path)
• The released LoRA adapter and GGUF artifacts on HuggingFace
• The MCP tool catalog when it ships in v3 (the python sandbox is the highest-priority target)

Out of scope:

• Vulnerabilities in upstream dependencies (Qwen base model, PyTorch, BitsAndBytes, llama.cpp, Triton, FLA, etc.) — please report those upstream
• Generic prompt-injection issues with public LLMs
• Issues only reproducible on private forks

Once a fix is in place, the issue will be disclosed in the changelog with credit to the reporter (unless they prefer to remain anonymous). If a coordinated disclosure window is needed, we'll arrange one.

AVA pins exact versions of training-critical dependencies in docs/REPRODUCE.md and the CI workflow. The released adapter weights are reproducible from the listed corpus and config — the build is not signed but is bit-stable on the reference hardware. If you need stronger supply-chain guarantees, open a discussion.