Canonical knowledge set with generated renderings

[medicnick]

This is a breaking change to the file layout. No clinical content changed. Every threshold, criterion, and standard is the same as 2.1.1. What changed is how the files are organized and how the platform renderings are produced.

Why

The 2.1.1 audit found that safety-critical text had been paraphrased into each platform's instruction file rather than reproduced, and the paraphrases had lost content. The ChatGPT rendering was missing six of eleven forensic triggers, including domestic violence and sexual assault, and its standing disclaimer was truncated. Fourteen of twenty findings traced to one cause: three hand-maintained paraphrases of a single standard drift apart.

Those files are no longer written by hand.

What is different

One always-loaded block, identical on every platform. It holds only the rules that must never depend on retrieval: the editorial-tool boundary, ten NEVER rules, the attribution boundary, PHI, patient and incident separation, during-a-call guardrails, the pre-draft check, style, the standing disclaimer, and a router.

Seventeen topic files in paramedic-narrative/references/, all named WHEN-*.md. Each is named after the situation that requires it, not the subject, and opens by restating its own trigger. Filenames are the retrieval lever, so WHEN-FORENSIC.md is easier for a model to route to than a section buried in a monolith.

system-prompt.md, chatgpt-instructions.md, claude-project-instructions.md, and SKILL.md are now generated from those sources by build.py. Sources live in src/.

A verifier gates the build. It fails if the always-block is not byte-identical in every rendering, if a knowledge file is not routed, if the router points at a missing file, if the ChatGPT instructions exceed 6,500 characters, or if house style is violated. The 2.1.0 defects would not have shipped past it.

documentation-standards-primer.md and narrative-formats.md no longer exist. They duplicated each other and SKILL.md on seven topics, including the entire ACS Field Triage criteria list appearing near-verbatim in both. That content was deduplicated into the seventeen files.

Upgrading from 2.x

Remove the old files first. If SKILL.md, documentation-standards-primer.md, or narrative-formats.md remain alongside the new set, the assistant may consult a retired copy.

Platform | What to do -- | -- Claude, skill package | Download paramedic-narrative.skill below, remove the old skill, upload the new one. Single upload. Claude Project | Paste claude-project-instructions.md into Instructions. Delete the old project files, upload all seventeen references/WHEN-*.md. ChatGPT | Paste chatgpt-instructions.md into Instructions. Delete the old SKILL.md from Knowledge, upload all seventeen WHEN-*.md. Gemini / direct API | Paste system-prompt.md. Unchanged as a workflow -- that file now inlines all seventeen sections, since these platforms have no retrieval. Microsoft Copilot | chatgpt-instructions.md in Custom Instructions. Use system-prompt.md for the per-session paste instead of SKILL.md.

Provider profiles and agency configuration files are unaffected and need no changes.

On seventeen uploads

For ChatGPT and Claude Projects this is more work than the previous two files. It is one-time. The reason is that Knowledge retrieval is chunked and not guaranteed per turn, so a monolithic file means a safety rule may not reach the model on a given call. Claude users can avoid it entirely with the .skill package.

Contributing

Pull requests edit src/, never the generated files at the repo root -- those are overwritten by build.py. Run python3 build.py && python3 build.py --verify before opening a PR.

---

This discussion was created from the release Canonical knowledge set with generated renderings.