Investigating Model Motives — Blog Companion

This repository contains supplementary details and artifacts for the blog post on investigating model motives. It provides experiment-specific information (including prompt templates, workspace files, output variants, and plots) for each of the four environments in agent-interp-envs.

Structure

Each folder corresponds to one environment:

• funding_email_details/ — Funding email environment
• sandbagging_details/ — Sandbagging environment
• eval_tampering_details/ — Eval tampering environment
• secret_number_details/ — Secret number environment

Within each folder you'll find a README describing the experiment setup, prompt templates used, workspace files provided to the agent, output variants across runs, and plots summarizing results.