ECG-Bench — LLM Response-Mode Evaluation

ECG = Epistemic · Coercion · GOAT — a 143-item benchmark that measures which response modes LLMs activate under fixed answer coercion (— one answer only).

What it tests

Refuse, folk numeric, consensus pick, forced binary, hedge leak, obscure craft math — same suffix, different stems across domains.

Progress

• Full eval suite across Anthropic, OpenAI, DeepSeek, Gemini
• A/B intervention studies and paper draft with figures
• Dataset on Hugging Face

Stack

Python · multi-provider LLM APIs · Pydantic graders · matplotlib