v0.12.45: @newsroom.writer: System prompts as behaviour policies

A new article written here from a voice memo: can an existing system prompt be represented as an agent behaviour policy? Tested on Anthropic's published Claude Haiku 5.5 system prompt (7 October 2026): 4,255 words, 181 sentences, 82 rules, sorted into eight kinds (harm restriction, judgement, referral, conversation state, presentation, tool use, knowledge, input trust); 36 could be held by the harness, 21 in part by classifiers, 25 stay with the model. A leaked ChatGPT prompt, unverified, used for its shape: about six words in seven under its tool sections. The Model Spec's chain of command as an authority column. A proposed ontology and JSON shape (policy, mandate, reach, inputs and trust, state, rules, tests) with three of Claude's rules written in it. Formats for agent policies are arriving; published agent behaviour policies were not found. Two figures.