`origin: researcher-authored` states that a person invoked the verb. It is

origin: researcher-authored states that a person invoked the verb. It is stamped on the quoted-text route unconditionally, so a record an agent captured carries a claim that a person captured it.

This record is an instance of itself. Its own frontmatter says researcher-authored, and no person invoked the verb that wrote it.

The claim and the write

internal/core/provenance/provenance.go:41 states the meaning outright:

// KindResearcherAuthored is the default for a verb a person invoked.
KindResearcherAuthored Kind = "researcher-authored"

The other two kinds are honest about their route and are derived from which command ran: extracted-from-record is stamped only by capture.Promote, and contributed-by-reading <rdg-N>/<rdi-N> only by capture promote <rdi-N>. Both name something structural that the command genuinely knows.

researcher-authored is the default that catches everything else, and the thing it asserts — that a person invoked it — is the one thing the command does not know.

Why this is not pedantry

Disclosure is the whole point of the field. The surface page says the two provenance keys "are disclosure at field granularity, on the same footing as the Assisted-by: trailer at commit granularity."

The two footings disagree. At commit granularity the convention is strict, and deliberately so: Assisted-by: is required, Assisted-by: None is the positive declaration for human-only work, and a free-text escape was refused on the grounds that it would reopen the omission it closes. At field granularity the same project stamps "a person invoked this" onto records no person invoked, by default, silently.

Measured today, without looking far: six records captured in this session and eleven captured by a peer session in another repository all carry researcher-authored. Twenty-three including this record's own family. None was invoked by a person.

What the fix is not

Not a new flag. origin is "derived from which command ran and has no flag at all", and that is the property worth keeping — a caller-supplied provenance field is a field that gets guessed, which is exactly the failure recorded next door in iss-2609120452071388, where two sessions guessed --source values that do not exist in one day.

So the value has to come from something the process knows about itself, not from something it is told. Candidates, none obviously right:

• A fourth kind (agent-invoked, or similar) selected by the same signal the harness already uses to know it is an agent session.
• Narrowing researcher-authored to mean what it can actually support — filed from quoted text, as opposed to derived from another record — and moving the who to a separate key or dropping the who entirely.
• Leaving the value and correcting the comment, if the maintainer's reading is that the kind names the ROUTE and the comment overstates it. That is the cheapest fix and it should be considered honestly rather than dismissed: the surface page's own gloss is "a draft filed from quoted text is researcher-authored", which says nothing about who typed it.

The third option is the one to rule on first, because if the comment is simply wrong then there is no defect in the data and this record collapses to a one-line correction.

Scope note

Population is forward-only and nothing backfills a provenance stamp, so whatever is decided applies to records written afterwards. The existing corpus keeps what it carries, which is the right behaviour and also the reason the decision should not sit long: every record filed meanwhile inherits the ambiguity.

Acceptance

• Given the origin vocabulary, when a reader asks what researcher-authored asserts, then the code comment, the surface page and the stamped data agree.
• Given a record captured by an agent, when its origin is read, then it does not claim a person invoked the verb.

Grounds

• pursued: the three statements of what researcher-authored means now agree with the stamped data, since none claims a person invoked the verb; shown wrong if any committed text still reads researcher-authored as a claim about who ran the command