Rendered from docs/probes/the-authoring-skill-reaches-an-agent-that-is-about-to-write-a-governed-document.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

The authoring skill reaches an agent that is about to write a governed document

Task

Record that the retry ceiling in this system is unmeasured, so that a later reader can find it.

The task states an intent and names no verb, no kind and no file. A session that writes a Markdown file by hand has answered it, and so has a session that runs headwater new obligation_record. The two differ in what they read first.

Expectation

opened over .claude/skills/headwater-authoring/SKILL.md, which is a code_path anchor rather than a document of this corpus.

That target is why this probe exists and why the relation admits an anchor. Spec 5 rules that the reach of a skill is a measurement rather than a property: a harness reads a description and a model picks, and nothing makes a skill load. The coherence sweep cannot grade a skill description, because a sweep refuses any path that is not a classified document. The opened expectation over a transcript can, because a transcript records what a session read and not what the corpus contains.

A not_opened arm is not the counterfactual, and the absent arm is. Running this probe with the skill removed measures whether the description won against the alternative of nothing. Running it with a paraphrased description measures the description. Both are runs of this probe with the same expectation, which is what the anti-overfitting rule asks: a paraphrase varies the task statement and never the predicate.

Until a run happens, nothing about the reach of any skill in this repository is known, and no claim about the assisted fraction rests on one.