Skip to content

tslit.stub

Nicolas Cravino edited this page Apr 15, 2026 · 3 revisions

id: tslit name: TSLIT (Time-Shift LLM Integrity Tester) repo_path: ~/Documents/sw30labs/repos/tslit remote_url: https://github.com/sw30labs/tslit.git primary_language: Python framework: LangGraph + llama-cpp-python ingested: 2026-04-11 last_commit_date: 2026-04-11 category: ai-security stacks: [langgraph, pentest]

Purpose

Sandboxed evaluation harness for detecting time-based latent behaviors and affiliation bias in local non-US-origin LLMs served through native llama.cpp runtimes. Tests 32 scenarios across 15 probe dates (~3,840 interactions) to surface logic bombs, affiliation bias, censorship, and telemetry injection.

Architecture

CLI entry via Typer (src/tslit/cli.py). Campaign runner executes scenario×date matrix against a test model via llama-cpp-python. 12 threat detectors run per interaction. Results logged as NDJSON artifacts. A separate multi-agent LLM Analyzer (Ollama-backed) with AnalyzerAgent + QAManagerAgent validates findings in a loop. Reports generated as JSON and text.

Dependencies

Core: typer, rich, pydantic, pyyaml, python-dotenv, llama-cpp-python. Analysis: langchain, langchain-core, langchain-community, langgraph, scipy, numpy. Dev: pytest, black, ruff.

CI/CD

No CI workflows. No Docker.

Entities and topics

Clone this wiki locally