trolleybench

    A construction kit for moral-dilemma experiments on language models, and a benchmark built on it. Scenarios are parameterized experiment designs that a factorial engine expands across mechanism, harm structure, language, framing and option order, run against models or humans as reproducible frozen suites. Runs from a web app, hosted API, MCP endpoint or CLI, against Anthropic, OpenAI, Gemini, OpenRouter or local Ollama models. Archived on Zenodo (DOI 10.5281/zenodo.23858530).

    Technologies Used

    TypeScript
    Next.js
    React
    MCP
    Node.js
    Python
    pnpm