Person picker — decided states

The owner's answers of 6 Oct 2026 (design-people.md §J8, points 8–10) drawn on the shipped picker — and, in section 5, dashed and marked speculative, three undecided ways to tell a paragraph's speaker from the whole speaker, and in section 7 a fourth, the owner's sticky Speaker | Paragraph switch: the Export menu's box and rows, the role toggle, the split person badge, the house check from components/Icon.tsx, and the grid and transcript classes, with CSS inlined from the tree. Each chrome bar names what is new in that frame. Codes follow Phase 1 (per person and role), which is not on main yet.

Artefact: real pixels, nothing addedCommentary: not the product

1 · An unknown speaker

Commentary

1a is the grid as it stands for an unknown moderator under Phase 1: m? and the grey italic role word. 1b is the picker it opens. There is nothing to confirm, so the cursor starts in the new-person row and the researcher just types. The hint is New moderator, the shipped string, here in grey italic. Whether the hint is italic is the open question from the tick page, so 1b and 1c show both. The row shows m3, the next free code after Martin's m1 and Kerri's m2 (§J8.8), not this slot's own.

1d is the guard against duplicates. Typing "Mar" with Martin already known moves the highlight onto Martin, so Return picks the existing person instead of making a second Martin (answer 2: never merge on a name). To make someone new with a matching name, the researcher arrows back to the field. This is the type-ahead the owner anticipated, and it is the one genuinely new interaction on this page.

2 · A known speaker

3 · Looking without changing

4 · Two unknown moderators in one session

5 · Paragraph or speaker? — speculative

Commentary — speculative, not decided

The owner's reading of 6 Oct 2026: opened from a transcript paragraph, "this quote was actually from m1" is about that paragraph, but the picker as built changes the whole speaker for the session. Renaming is not in question — a name is always the speaker's. Reassigning is. Three ways to draw it, none decided; research on how Otter, Descript, Dovetail and Marvin handle the same question is under way, and this section should be read beside it.

A · paragraph first, then the offer. The paragraph badge asks only "who said this?" — no role toggle, since a paragraph has a speaker, not a role. After the change, one line offers the wider fix: p1 has 42 more paragraphs. The scope is decided from evidence, not in the abstract, at the cost of a second question.
B · scope inside the picker. One picker, with a second toggle above the roles: this paragraph, or all of p1's. Explicit, but every visit asks the question, including the common speaker-wide case.
C · two places. The speaker picker moves to a speaker strip at the top of the transcript (and stays on the Sessions grid, where only the speaker question makes sense); a paragraph's badge asks only about that paragraph. No question asked, but two controls that look alike do different things.

Every paragraph-scoped option needs storage the pipeline respects — the next run replaces paragraphs — and quotes are credited to the session's participant, not to each paragraph's speaker, so moving one paragraph changes the evidence only once membership becomes paragraph-aware (jobs 10 and 11). The strings in these frames are owner-unseen placeholders, named in each chrome bar; none is a locale key.

6 · Split a paragraph, then reassign the second half — speculative

Stage 1 · split

Stage 2 · reassign the second half

Commentary — speculative, not decided

Stage 1 is the common case (owner, 6 Oct 2026): a long paragraph that does not read as one thing, split in two with the same speaker. It follows the convention of the research tools that support splitting at all — Dovetail, Trint and Condens: place the caret in the text and press Return. 6a is the caret; 6b the result: two paragraphs, one speaker, the second with its own start time. Where the transcript has word timings, the split point gives that time exactly (07:58 here is the start of the word after the caret); without them the second half keeps the first's time. 6c: ⌘Z joins them again, and so does Delete at the start of the second, the same convention in reverse.

Stage 2 is the rarer case: the second half was really another speaker. Its badge opens the paragraph-scoped picker from section 5's option A or C — "who said this?" — and choosing Martin moves that paragraph alone (6d, 6e). Nothing here changes Simon's other paragraphs.

Both stages meet the wall the jobs map names: the next run replaces paragraphs, so a split and a reassigned half have to be held somewhere the pipeline respects. Stage 2 also meets the evidence question — a quote whose time window covers the moved half is still credited to the participant until membership is paragraph-aware (job 11). Strings here are placeholders; none is a locale key.

7 · A scope switch: speaker or paragraph — speculative

From a transcript paragraph

Live: click a badge, flip the scope, pick someone

The Mac popover, at the shipped small size

Sticky: a run of wrong paragraphs

From the Sessions grid: no switch

Commentary — speculative, not decided

The idea (owner, 6 Oct 2026): on the transcript only, a two-way switch above Moderator · Participant · Observer, first drawn as Speaker | Paragraph and now labelled Session | Paragraph (below), defaulting to Session. Below it the picker is exactly today's. With Speaker it acts on every one of p1's turns in the session; with Paragraph it is an atomic "who said this one?". The switch is sticky: leave it on Paragraph and a run of mis-credited paragraphs is click, pick, click, pick. It is section 5's option B with the default settled and the stickiness added, which answers B's main cost (asking every time): you are asked nothing, and the scope is visible before you act rather than offered afterwards.

What it settles. The scope is chosen first and is on screen while you choose. The common case (the whole speaker is wrong, or you are naming them) costs nothing extra. The Sessions grid never shows the switch (7e): a grid row is a speaker. Offered on the transcript only, to start with (owner, 6 Oct 2026); any other surface waits until this one has been used. And it gives each scope room to differ later without the two looking alike, which was option C's weakness.

Questions it raises, in rough order of weight:

What Paragraph does to a name. A name is always the speaker's, so with Paragraph on, the name is not a field (7b: plain row, no rename). Typing a new person under Paragraph makes that person and credits this paragraph to them. That is a new speaker in the session, which Speaker scope never creates.
The switch is plain text, not a second segmented control (owner, 6 Oct 2026). Two stacked .dimension-toggles read as one control in two rows; drawn as words, semibold in body ink for the chosen one and regular in muted grey for the other (owner, 6 Oct; --bn-weight-semibold, 600, added to the scale for it), it reads as a heading. The eye starts at the role toggle, which is where the act is. No box, no blue, no rule. It stays a radio group for VoiceOver and the keyboard. Note it is new: the house has no text-weight toggle yet, so this is a small new primitive, not a reuse.
Where the stickiness lives. Per transcript page and reset on leaving it? Per session? Per window? A sticky Paragraph forgotten on is the risk: the next "rename Simon" would not be possible, which is at least a visible failure, not a silent one. Resetting when the page changes is the conservative answer.
The labels: Session | Paragraph (owner, 6 Oct 2026). Speaker was too high level; All of p1 overclaimed for a moderator, since m1 Martin runs through every session while a pick or a recode reaches only this session's slot; This session | This paragraph repeated "This". The pair names the reach of the act, as the prior art does (Grain This speaker block / All speaker blocks, Marvin one instance / across the transcript, Rev a Change all [label] to: checkbox, §K2). A rename is the person's everywhere whichever is chosen, as names are elsewhere. "Paragraph" is already the transcript's word (Edit ▸ Undo Split Paragraph). Two plain keys, no interpolation.
What a paragraph move costs. Speaker scope on a moderator ↔ participant change is today's recode, then Re-analyse…. A paragraph move needs storage the next run respects: the split/join table (transcript_layout_edits, keyed by position and first words) can hold a "speaker" edit the same way. The evidence follows only once quote membership is paragraph-aware (§K, job 11). Until then a paragraph moved off p1 still counts as p1's evidence. Moving a moderator's question off the participant is the useful case, and it is the one evidence-membership has to honour.
Sticky could carry the pick too. On a run of a moderator's questions, the picker could open on the role and person last picked (Moderator, Martin focused), so each paragraph is click then Return. Cheap, and it is what makes "click click click" literal.
Two pickers, two drawings. Frames 7a–7f draw the web picker, whose role toggle is the house .dimension-toggle (blue-washed active segment); the "Mac app (SF)" switch at the top changes only its typeface. The Mac popover is AppKit (7g, 7h): an NSSegmentedControl at the small size, 11pt type, 24pt rows, all from PickerMetrics. In a key window the selected segment is filled with the accent and set in white, as the owner's screenshot of the real control shows; the grey pill in moderator-identity-picker-v3.html was sampled with the window not key. On the Mac the scope words would be two borderless 11pt labels, labelColor semibold and secondaryLabelColor regular, centred over the control.
Undo. Each paragraph pick is its own ⌘Z entry, so a run of five is five undos. A whole-run undo is not proposed.