Do Frontier LLMs Adapt to Neurodivergence? A Cross-Model Audit

Code, configs, and data for an empirical audit of how GPT-5-chat and Claude Sonnet 4.6 respond to neurodivergence (ND) context supplied via the system prompt.

See spec.md for the full experimental design and research questions.

Layout

configs/        # profiles, prompts, queries — single source of truth for the experiment
ndbench/        # runner, metrics, judges, analysis code
data/responses/ # raw model outputs (JSONL), one per (model, condition, profile)
paper/          # LaTeX source

Quickstart

1. cp .env.example .env and paste real API keys into .env (never commit).
2. python -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
3. python -m ndbench.runner — dispatches all model calls, caches to data/responses/.
4. python -m ndbench.metrics.run — computes structural, surface, and harm metrics.
5. python -m ndbench.analyze — fits models, emits figures and tables to paper/figures/.

Models under audit

• gpt-5-chat-latest (OpenAI, non-reasoning chat variant)
• claude-sonnet-4-6 (Anthropic)

License

Code: MIT. Data and paper artifacts: CC-BY-4.0.