LinkedIn drafts

One draft per Guidelines section, plus a launch post and a "what we learned" post. Each is at most 1,300 characters, carries one concrete finding or artefact, and ends with the repository link.

The drafts were written by Claude Fable 5.1 from the result files. The .md.in files are the sources; the .md files are generated by experiments/render_paper.py (make paper) with every number injected from results/summary.json. Edit the template, never the output: a hand-edited draft would drift from the paper the first time an experiment is re-run. A test enforces the character limit, the trailing link and the absence of unresolved placeholders.

#	Draft	Clauses	Finding it carries
01	01-launch.md	—	What the toolkit is; the automatability headline
02	02-applicability.md	§2.2-§2.5	The applicability test changed from "integrated use" to material adverse impact
03	03-governance.md	§3.3-§3.6	Section 3 is 0% automatable, and why that is correct
04	04-inventory.md	§4.2-§4.9, fn21, fn22, fn25	An inventory must account for what is not in it
05	05-materiality.md	§4.10-§4.13	Autonomy moves the control set, not the tier; no GenAI use case can be Low
06	06-lifecycle.md	§5.1-§5.26	A semi control never reports PASS; attestation is not failure
07	07-guardrails.md	§5.15, fn46	The patterns that suppress attacks refuse half of benign in-scope traffic
08	08-third-party.md	§5.10-§5.11	Third-party AI tightened; an ISO 42001 certificate does not test model behaviour
09	09-crosswalk.md	all	The ISO+NIST+EU residue is governance, not the agent clauses
10	10-what-we-learned.md	—	The refutations, and the measurement bug that produced a perfect score