-
Notifications
You must be signed in to change notification settings - Fork 0
tslit dspy dgx.stub
id: tslit-dspy-dgx name: TSLIT-DSPy — DGX Spark Port repo_path: ~/Documents/sw30labs/repos/tslit-dspy-dgx remote_url: https://github.com/sw30labs/tslit-dspy-dgx.git primary_language: Python framework: DSPy (MIPROv2) + vLLM category: ai-security stacks: [pentest, agentic, dgx] ingested: 2026-07-22 last_commit_date: 2026-07-21
Port of tslit-dspy-ar to NVIDIA DGX Spark (GB10 / Linux aarch64) with a local vLLM stack, following the same Desktop DGX conventions as driftlab-dgx (and the private grey-swan-dgx / youtube-scribe-dgx siblings). A defensive research pipeline that analyzes synthetic probe responses for affiliation bias, temporal logic bombs, and combined threats.
Enforces a detector-provenance policy: the detection brain (compile, inference, autoresearch) must use non-adversary / American models — NVIDIA Nemotron via local vLLM by default — while adversary-origin open weights (Qwen, DeepSeek, MiniMax, …) are scan targets only, never part of the detector. Ships both tslit/ and tslit_dspy/ trees plus workspace/, config/, QUICKSTART and setup-venv.sh.
dspy>=2.5, python-dotenv, openai (client for the vLLM OpenAI-compatible endpoint); pytest + ruff in requirements. Python 3.12; package tslit-dspy-dgx 0.2.0; 3 test files.
No CI workflows in-repo.
- tslit-dspy-ar — the upstream DSPy analyzer this ports
- tslit — v0.1 origin
- driftlab-dgx — same DGX host conventions and shared local vLLM stack
- Tags: [python, security, llm, dspy, vllm, cuda, pentesting]