Person actions, everywhere

This is the whole of step 1 for a participant. No submenus except Role, no scope choices, no header. The anchor is the scope: you clicked p4 in session 4, so the name you type is p4's — and because participant codes are numbered across the whole study, that name is p4's everywhere. Nothing needs to say so.

Deliberately absent: everything from steps 2–6. No Confirm (there is no guess to confirm), no Same Person As, no Turn items — a badge in a sessions row is not standing over any particular turn. A menu that shows six disabled items to advertise a roadmap is a menu nobody reads.

Where the name comes from

NSFullUserName() — the Mac's own account name. The app offers it and never assumes it, which is the standing rule for profile connection: the researcher may say "I'm m1", the system may not decide it for them.

Scope

This study. m1 in every session where the app currently believes m1 is one speaker. The sweep line confirms the count afterwards — You are m1 in 8 sessions — because a one-click action with an eight-session reach has to say so somewhere.

Fallback

Where there is no full name to offer (a browser session, or an account called admin), the item degrades to That’s Me… with an ellipsis, asks once, and never asks again.

Two reasons this earns first place over any matching algorithm. It is the only identity claim in the product with a perfect source — the person asserting it is the person. And it is the most reused: a freelancer's own name is in every study they will ever run, so one click here is the last time the question needs asking, in this project or any other.

The item sits above Set Name, against the usual instinct to lead with the generic verb, because it is both the more common answer for an m code and the cheaper one. On a p code it does not appear at all.

Why it is its own case

Not unknown, not misidentified — the right person, mis-transcribed. It is the safest edit in the whole vocabulary: no words move, no identity changes, nothing merges. And it is the most frequent, because a speech model mangles proper nouns on every raw-audio study while a platform transcript, which reads the name off an account, gets it right.

Why the researcher wins every time

She recruited him. The screener has the spelling. This is the one class where the human is not probably right but definitely right, so the interaction should cost nothing and ask nothing.

What only a rename can offer

The old string is known-wrong, so it can be found exactly. Setting a name on an empty field has nothing to search for; changing one does. That is the rule: rename offers propagation, naming does not.

The label and the words are two different places, and today only the label is fixable. The transcript body still says "Michel Hurlly" wherever anyone said it aloud — so the report shows the corrected name at the top of a quote whose text spells it wrong. Nobody would design that; it is just what happens when a name field and a transcript are edited by different features. The sheet is the cheapest possible bridge, and it is a scoped exact-string replace, not free-form transcript editing.

Neither button is destructive and neither is Cancel — "Just the Label" is a real answer, not a retreat, and it is what a researcher picks when the participant genuinely said a different name aloud. Escape takes the same path.

A name belongs to a code, and a code has one name. Correcting the spelling here corrects it in session 7, in every quote, in the participant table and in the export — there is no such thing as p7 being spelled one way here and another way there, and the sheet does not offer a scope because there is no scope to offer. That invariant is load-bearing well beyond spelling: it is what makes the moderator renumber in bench 9 forced rather than chosen.

The bigger version of this is a recruitment list. The researcher knew the spelling before the session existed — the names were in a screener days earlier. An imported roster would let the app match what it heard against names it was given, rather than guessing spellings from audio, and would collapse this whole class. Out of scope here, worth noting as the thing steps 1–2 are a manual stand-in for.

"in Session 4" is in the item text, not implied. Role is a fact about a person in a session, not about a person — the database has always modelled it that way, and the edge case is real: the same colleague can observe one session and moderate the next. Naming the scope in the item costs four words and prevents the reasonable assumption that this is a global switch. At a person-level anchor the same item reads Role in All 12 Sessions, and where the roles genuinely differ it says so rather than flattening them.

What it does not do: retract the six quotes already pulled from Jane's asides. Role governs who is eligible for quote extraction, and extraction has already run. Changing a label cannot un-ask a question that was already answered — so the surface offers Analyse again and lets the researcher decide, rather than silently spending a cloud call and discarding edits.

And it changes who is named in an export — today, by accident. Anonymising blanks names whose code starts with p and keeps m and o. Jane as a participant is blanked; Jane as an observer is named. One three-item submenu quietly moves a colleague across the anonymisation line, in the direction of more disclosure, with nothing on screen saying so. That is Part 5's problem, and it is live in shipped code.

Same name, different people

p3 and p5 are two Marys who have never met. Correct as they stand, and the app proposes nothing — name equality is not evidence at any altitude, including inside one study. The only work is telling them apart on screen, which the shipped short-name collision handling already does: Mary A. and Mary O.

Same person, two codes

p3 and p6 are one Mary, back for a second session on another part of the journey. Normal, not a defect. Participant codes are numbered across the study so they never collide — which means one human returning simply holds two of them.

The only action on this screen

Naming p6 by picking the Mary who is already in the study. The codes join as a consequence. One pick.

This is the common participant case, and it should be almost invisible. The temptation is to build matching — two "Mary"s in one study, surely worth a suggestion? No: the app cannot tell these three apart and would be wrong a third of the time. What it can do is put the ambiguity where the researcher will see it, name both candidates, and let the person who was in the room answer in one click.

Note what the contested row does not say. Not a confidence score, not "possible duplicate", not a merge proposal. Just the two names that are already in this study and the word Which? — because the researcher knows, and the app does not.

The second frame is the same shape one namespace over. Moderator and observer codes fragment as well as collide, for a dull reason: the number comes from within-session ordering, which has nothing to do with identity. Jane observes session 1 alone and is o1; she observes session 3 alongside Tom, who happens to speak first, and she is o2. Same person, two codes, and nothing about her changed.

So all three namespaces need the join, and only m/o need the split — and every one of those repairs is the same gesture: name the person in front of you. Here the app offers, because both names are platform records rather than guesses; a Mary heard once gets the silent Which? treatment instead. Offering is not merging, and the strength of the evidence decides which you get.

The grain that matters

Not a turn. All 41 turns at once, because when the diariser is wrong about a speaker it is usually wrong about all of them. Cheap too: one bulk update inside a session — no word-timing division, no split, no merge, no quote cascade.

Why it is an identity item, not a Turn item

It moves words, but what the researcher asserted is who p3 is — that the diariser invented a speaker who is really the moderator. The families are told apart by what is asserted, not by what the database ends up doing. Same shape as a rename causing a renumber.

Grouping is the role

Picking inside Observers sets an o code; picking New Moderator… mints an m. Role and identity are chosen in one gesture rather than two, and the groups run in p/m/o order — the order codes are read in everywhere else — rather than by guessed frequency.

The gap this closes: a person the system never noticed. Diarisation merges two people into one code and the second has no code at all, so there is nobody to reattribute to. This is the exact inverse of Not a Speaker, and without a create path the picker simply cannot express the correction. New Participant… mints the code on the spot.

Create and name are one action. A nameless p7 is a second chore, and the researcher knows the name at precisely the moment they are creating them. For a new moderator or observer the field offers the bank, because a new colleague is usually a known colleague; for a new participant it is blank, because they usually are new. Note this is the one place the two lists meet — the cast (who could this speech belong to, in this session) and the bank (who is this person, across studies) are otherwise different questions and different lists.

The whole cast fits without scrolling because the cast is small: a dozen participants, three or four colleagues. At that scale a grouped list is faster than any search field, and a search field would be a confession that the list is too long.

Why it outranks fixing the turn

A quote is published. It goes in the report, the spreadsheet, the Miro board, the deck. A transcript is a working document nobody outside the team reads. A slightly-wrong transcript is tolerable; a quote credited to the wrong participant is a factual error in a deliverable.

Why it is cheap

One field on one quote. No segment surgery, no word-timing division, no per-speaker stats recompute, and crucially no cascade — which is the unsolved part of doing it at segment level. The speaker-editing design already reaches this conclusion and calls it "recommendation B, simplest first".

Why crosstalk may never be fixable upstream

Two voices genuinely occupy the same three seconds. No boundary is correct. So the honest design does not promise a perfect transcript — it gives the researcher a way to make the output right regardless.

Quotes exist only for participants, so the current value is always a p code and there is no such thing as a moderator’s quote. That makes the two halves of this list mean different things. Picking another participant re-credits — genuine crosstalk, the words are a finding, credited wrongly. Picking a moderator or observer says this is not a participant quote at all — the speech is re-attributed, and the card leaves the lens because the lens only shows participant quotes. The group header says which is which at the point of action.

That second half is the everyday repair from bench 11 arriving on a different surface: an observer mis-filed as a participant had their asides mined as findings, and this is where a researcher meets the consequence — one odd quote at a time, while reading.

The card disappearing is a consequence, not the action. A quote is a view onto speech, not a copy of it — so re-attributing it moves the words, and going to session 3 afterwards shows them as the moderator’s. Had only the card changed, the two surfaces would disagree: the quote gone from here, the transcript still showing p4 saying it. Hence Show in transcript sitting next to Undo — the quote was not destroyed, it moved to where it belongs, and the researcher should be able to go and see that.

And nothing is withdrawn. There is no flag, no reason field, no pile it moves to. The Quotes lens shows participant quotes; re-attribute the speech to m1 and the card is simply no longer in that set — reversible for free, because putting the speech back puts the card back.

Two orthogonal axes, and keeping them apart is the whole point. Membership — is this a participant quote? — is a fact about who spoke, derived from speaker identity, with no state of its own. Emphasis — star promotes, hide demotes — is a judgement about evidence quality that the researcher makes deliberately and owns. Star and hide are one spectrum: picking and choosing the best evidence. Speaker identity is not on that spectrum at all. Which is why the hidden count stays meaningful — it counts judgements, not facts — and why a correction must never be expressed through it.

Note the sentence in the frame above avoids the word hidden. Borrowing the curation verb for a membership fact is the exact confusion this whole distinction exists to prevent.

And then the pattern this design uses everywhere: do the narrow thing, then offer the wide one in a sentence. One quote is re-attributed, because that is what was asked. The app, now holding strong evidence, says so and notes the five quotes still credited to p4 — inviting review, claiming nothing, and never performing the sweep uninvited. It is how a reading surface can default to the thing under the cursor while the underlying error is usually study-wide.

Trim and Hide sit in the same menu deliberately. When speech overlaps, the extracted text is often a genuine mash of two voices, and re-crediting all of it to one person swaps one wrong attribution for another. The real repair is to trim to the part one person actually said and then re-credit — or, when it is unsalvageable, to hide it. Both already ship; the menu just puts the three answers in one place, in the order a researcher tries them.

Note the noun. This Quote Is — a unit of speech, so the fix moves words. Role, Change Name, Same Person As — units of identity, so they only relabel. The rule generalises from the earlier one: the item names the object it acts on, and speech nouns are the ones that move things.

Two items say Turn and they are the two that move words to a different speaker. One says All of p4’s Turns Are — the whole-speaker sweep, named for the speech it moves and classified attribution, where its effect belongs. (An earlier draft spelled it p4 Is Actually — identity wording on a 41-turn word-move, distinguishable from the pure relabel m1 Is only by an adverb; the review killed it, so the rule now holds with no exceptions.) The rest relabel and move nothing. A researcher does not have to be taught this; they can read it off the menu.

Reach, top to bottom

This Turn Is — the thing under the cursor. All of p4’s Turns Are — all 41 turns in this transcript. The wider one is deliberately one group down: this is a reading surface, so the narrow act is the default and the sweep is available with its scope written into its name.

With a selection

The first item reads These 4 Turns Are. Selection carries the scope, which is the standard text-editor idiom and introduces no new concept — shift-click a run of misattributed turns, then fix them together.

Cost, which is not evenly spread

This Turn Is and All of p4’s Turns Are are bulk updates — cheap. Only Split and Merge need segment surgery: word-timing division, index management, a quote cascade. They are the rare grain, and they are the reason step 5 of the sequence is the largest item in it.

One exception worth carrying forward: when a split is needed because a quote sits inside a longer turn, the quote’s own timecode range supplies the split point. That case needs no cursor placement and no new interaction, which makes the flavour of turn surgery that arrives from a reading surface markedly cheaper than the general one drawn here.

Drawn before

A sheet titled "Separate m1 into two people?" with three buttons, a chain diagram and a nested question. It priced a tail case as though it were a feature.

What the researcher does

Opens session 9, says the moderator here is Mike, carries on. They never think "separate m1" — they think about the person in front of them.

What the app does

Notices m1 is Martin elsewhere, concludes this is a different person, renumbers Mike to m2, and says so in one sentence with an Undo. No confirmation, because the house rule is that nothing confirms and everything ⌘Z’s.

Renumbering is a consequence of naming, not a verb. That is the whole of decision 1, and it is why the same menu from bench 2 absorbs the tail case without growing an item. Three things that looked structural — joining two participant codes, separating a collided moderator, linking across studies — are all bookkeeping behind an act the researcher has already performed, provided naming is built as an identity assertion rather than a string edit.

The heavy sheet survives for one path only: noticing at study level, in the People lens, that a code covers two names — where there is no single session to open and no rename in flight. Secondary door, not the primary one.

The model is already right

full_name and short_name have been on PersonEditable from the start, and suggest_short_names() derives the second from the first with real care — honorific stripping, family-name-first detection, 337 surnames, and collision handling that produces "Sarah J." and "Sarah K." when two participants share a first name.

The UI is not

The pencil in the Sessions grid edits short_name and displays short_name || name. full_name appears only as a hover tooltip, and only when it differs. Nothing in the SPA can edit a full name at all.

So the user's case is half-served

You can already type "Mike" — but you cannot see that you are editing the display name rather than the record, and if Teams supplied a wrong or over-formal full name it is uncorrectable. The field is not inert: it is what the participant table shows, and it is in the export.

The sheet's job is almost entirely teaching by consequence. Two fields, each with one line saying where it shows up. "Called" first, because it is the one being changed nine times out of ten and the one with visible effect. Derivation still runs — leave "Called" empty and it fills from the full name — but it is now visibly a default rather than an invisible one.

This is also the quiet answer to the anonymisation question. A study that never records full names and works entirely in short ones has less to strip on the way out. Making the distinction visible lets a researcher choose that, which is a governance decision the product currently makes for them by hiding one of the fields.

Observers cannot be classified by the machinery as built, and they fall through to participant by construction. Role identification samples roughly the first five minutes (seg.start_time > 300), so anyone who first speaks at 31:40 is not in the sample at all. The heuristic that covers the rest scores question ratio and moderator phrases — an observer fires neither. And the code assignment maps PARTICIPANT and UNKNOWN to a p code, so "could not tell" and "is a participant" produce identical output. The observer is not unlucky; they are the default.

The signals that would catch them are computable and already stored: very low word count, a late first utterance, and never being the addressee of a moderator question. None needs a model. One absent-when-zero line is enough — it costs no column width, it is a single translated string, and it targets a bias rather than nagging about every speaker.

Now the perverse part. Correcting Jane from participant to observer moves her across the anonymisation line, towards disclosure: the export blanks names whose code starts with p and preserves m and o. As a mis-filed participant her name is stripped from the client's leave-behind. Correctly filed as an observer, it is published. The most common role error currently produces the safer export, and fixing it produces the riskier one — which means the sweep above cannot ship before Part 5's second decision, or it will do net harm while looking like a correctness feature.

What the bank is

Not a new store. Person rows are already instance-scoped — no project_id, deliberately, with a docstring saying merging is a future human-driven action. The bank is a query over rows that already exist, ordered by how often each has been used.

Why moderators and not everyone

The recurrence is wildly lopsided. Three or four researchers appear in every study; a participant appears once and never again. So the bank is short, its hit rate is high, and it earns its place for exactly the population the user described.

The part that changes the plan

Picking a name from the bank is a link, not a copy. Choosing "Steve Nakamura" asserts that this m1 is the Steve who already exists — which is precisely the cross-project identity link that was step 6. Obtained here for free, at step 1, as a side effect of not retyping.

Prevention beats reconciliation. The linking feature as designed is a reconciliation engine: import everything, generate duplicates, then match names within a folder, propose, confirm, and handle transitive chains. The bank stops the duplicates being created. What is left for a links table is a back-fill for people named before it existed — a much smaller and much less dangerous feature, and one with no matching algorithm in it at all.

"Someone New…" has to be the visible escape, and the safe default. The failure mode is a researcher picking "Rachel Okonjo" for a different Rachel, which silently welds two people together. So the bank never pre-selects, never auto-completes on a keystroke, and the study counts beside each name are there to make "which Rachel" answerable in the menu itself. Cross-client picks still deserve the confirmation from the linking work — the bank changes how a link is created, not whether crossing a client boundary is worth a sentence.

"Me" sits above the separator and outside the team group, because it is not a person you might be confusing with another person. It is also the row that seeds an empty bank: the first time anyone uses this, NSFullUserName() is the only entry there is.