When a quote and its transcript disagree

Decision 6 (docs/design-transcript-editing.md, revised 24 Oct 2026): a correction flags an edited quote only when it lands inside the span the quote kept and the words there were reworded. This page asks how loud that flag is, where it shows, and what it does to the analysis already made from the old words. Quote cards, the transcript paragraph and the signal-card quote are drawn with the shipped CSS; anything inside a dashed purple outline, and every string marked (new string), is proposed.

artefact: what would ship commentary: not the product proposed element or string

1. The case

transcript · session 3 · after the correctionshipped CSS

[08:12] p2

My pharmacist said it’s not fine to take it with food, so I’d been doing that wrong for months.

Taking medication with meals

Confidence

Commentary

Whisper dropped a “not”. The researcher restored it in the transcript (a correction, decision 6).
Last week the same researcher had tidied the quote from this passage to Fine to take with food, apparently. It is starred, tagged Confidence, and counted in a positive signal card.
The correction lands inside the quote's kept span, but the tidied words don't contain “it's fine”, so it can't flow in. The quote now says the opposite of the record. That is the only case that flags.
How often: rare. It needs a reworded quote and a correction inside it. A flag can afford to be louder than a common one.

2. How loud: three versions of the quote card

A · glyph

Quotes lensproposed

[08:12]
“Fine to take with food, apparently.” p2
Confidencemedication routine

B · the transcript's words, inline

Quotes lensproposed

[08:12]
“Fine to take with food, apparently.” p2
Confidencemedication routine
The transcript now says My pharmacist said it’s not fine to take it with food

C · B, and held out of the counts

Quotes lensproposed

[08:12]
“Fine to take with food, apparently.” p2
Confidencemedication routine
The transcript now says My pharmacist said it’s not fine to take it with food
Not counted until you choose

Commentary

	A · glyph	B · inline words	C · held
What the researcher sees	An amber triangle after the quote; the reason only on hover.	The transcript's words under the quote, the corrected word marked, and two choices.	B, with the quote dimmed and a label saying it doesn't count.
Missed?	Easily: hover-only, in a lens of hundreds of cards. Fails the medical case.	Hard to miss while scrolling; the words do the explaining.	Impossible to miss.
Sibling	`warning` kind (`ui_kinds.py`), today only whole-run states.	The proposed AutoCode tag's accept / deny: the one "decide this" state a card has.	A hidden quote counting nowhere (`analysis_quotes`, 22 Oct).
Cost	None.	Height on one card. It stays until a choice is made.	Silently changes counts, signal cards and the Dashboard until a choice, which could surprise more than the flag.
Proposed	B. The case is rare and the meaning is inverted, so the words belong on the card, not behind a hover. C's hold is a separate question (§4).

3. Where it shows

Quotes lens · view switchernew item

Quotes lens

The card carries the variant chosen in §2.
A third view, Quotes to review (new string), beside All and Starred (viewMode in utils/filter.ts), present only while a flag exists. The same place a later "to review" state (AutoCode proposals) could join.
The section's group header could count it as it counts hidden quotes (.bn-hidden-badge). Proposed: no, the view is enough.

transcript · guttermark in the warning colour

[08:12] p2

My pharmacist said it’s not fine to take it with food, so I’d been doing that wrong for months.

Taking medication with meals

Confidence

Transcript page

The quote's gutter mark takes the warning glyph and colour, with the same hover text (GutterEntry, beside its data-hidden state). Clicking it opens the quote in the Quotes lens, where the choice is made.
The corrected word is marked in the text (it is the transcript's own correction mark, decided at step 2a, not this flag's).
No note in the transcript: the transcript is right. The quote is what disagrees.

Signals · a card's quotesglyph only

[08:12]Fine to take with food, apparently. p2

[14:03]I just take them with breakfast, it's easiest. p5

Signal cards

The quote shows the glyph and hover text, never the inline note: a card is for reading evidence, not deciding.
It still counts toward the card unless §4 holds it out. The card's written headline was made from the old words; decision 21 rewrites a card only for a meaning change once 2b can tell, and this is one.

Surfaces with no drawing

Surface	Proposed	Why
Dashboard	Nothing.	No to-review tile exists, and one flag doesn't earn a new widget. The Quotes view is where review happens.
Mac sidebar	Nothing.	The sidebar's warning glyph is for a run's state; a per-quote review isn't one.
Exports (CSV, XLSX, Miro, FigJam, HTML)	Ship the researcher's wording, no flag; no question before exporting.	An export is the researcher's deliverable, and the flag is their to-do, not a fact for the reader. A question at export time is needy. Open: §5, question 3.
Chat lens and MCP	Tell the model.	An agent cites the quote with its sentiment. Grounding should say the transcript disagrees, so an answer doesn't repeat the inverted fact as evidence. See below.

What the model would be given beside the quote (model-facing, English, through grounding.py):

[q-p2-492] p2 · "Fine to take with food, apparently." · sentiment: confidence note: the transcript under this quote was corrected after the researcher edited it; the transcript now reads "My pharmacist said it's not fine to take it with food". Do not cite this quote as evidence for what it says.

And one line added to INVARIANTS, so the chat lens and MCP follow it alike.

4. How it meets the analysis already made

Moment	The quote's words	Sentiment and tags	Counts (signals, Dashboard, sentiment bars)	The flag
Flagged (the correction lands)	The researcher's edit, unchanged (decision 6).	As scored: confidence, made on the old words.	Open (question 2): counted as now (A, B), or held out like a hidden quote (C).	On.
Use the transcript's words	The edit is replaced by the transcript's words for the kept span, opened in the edit field so the researcher can tidy them again.	This is a meaning change: step 2b re-scores sentiment or marks the session owed. Until 2b exists, the sentiment badge shows it was scored on other words.	Follow the new sentiment once re-scored; the signal card is rewritten (decision 21).	Off.
Keep my wording	Unchanged.	Unchanged.	Unchanged.	Off, and remembered: who chose and when, so the same correction never flags again.
Edit by hand	Whatever the researcher writes.	As "Use the transcript's words" if the new words change the meaning (2b).	As above.	Off once the edit is saved.
Undo the correction (⌘Z, or revert in the transcript)	Unchanged.	Unchanged.	Back as before.	Off: nothing disagrees now.
Analyse re-reads the session (owed for another reason)	The researcher's edit still wins (decision 6).	Re-scored by the model on the corrected transcript, but the card shows the edit, so they can still disagree.	From the re-read.	Stays on until the researcher chooses.
Re-analyse	Gone with every edit (decision 12).	Fresh, from the transcript without its corrections, which go too.	Fresh.	Gone. The word list re-offers the correction (decision 22), which can flag nothing, since no quote is edited yet.

Decided 2 Nov 2026 (owner)

A, the glyph, with the detail in a popover; and the edit is applied, not flagged. The analysis re-runs for the quote about 30 seconds after the last edit, with a shimmer while it runs; a trivial result settles with no glyph; a strong one shows the glyph, whose popover reads “⚠ Transcript edit changed analysis” with a green and red diff and Undo Edit / Keep changes. The quote takes the corrected transcript's words. B, C, “Keep my wording” and §4's flagged rows are superseded; questions 2–5 below fell away with them. Record: decision 26 of docs/design-transcript-editing.md.

5. Questions for the owner (as asked)

Loudness. A glyph (A), the transcript's words inline with two choices (B, proposed), or B and held out of the counts (C)?
Counts while flagged. Does the quote keep counting toward signals, sentiment bars and the Dashboard until the researcher chooses (proposed), or count nowhere, like a hidden quote?
Exports. Ship the researcher's wording with no flag (proposed), carry a note in the export, or ask before exporting?
Agents. Tell the chat lens and MCP that the transcript disagrees, with an INVARIANTS line (proposed)?
Keep my wording. Is a single click enough to clear a flag on an inverted medical fact, or should it ask for a reason, as a rejected AutoCode tag can?

New strings, all owner-written drafts: The transcript was corrected here, and this quote says something else (hover), The transcript now says, Use the transcript’s words, Keep my wording, Not counted until you choose (C only), Quotes to review. Each is one key in all 21 locales once settled.