Tapestry scale: seconds per pixel

How dense a slice can get before each lane stops reading, measured against how often real transcripts change speaker, change topic and produce a quote. The sessions below are synthetic, sampled from the measured distributions.

Measured on real test recordings

Quantity	p10	p25	median	p75	p90	n
Speaker turn length	2 s	5 s	10 s	26 s	46 s	1,854
Topic segment length (s08)	48 s	86 s	181 s	266 s	362 s	194
Gap between quotes (s09)	30 s	45 s	64 s	92 s	128 s	397
Theme span in one session, share of session	0%	0%	10%	40%	60%	70

36 sessions of at least 10 minutes with at least two speakers, drawn from 9 trial-run projects; 16 of them fully analysed. Rates over the analysed sessions: 3.1 topics and 6.6 quotes per 10 minutes, and about 4–5 themes per session. Turn lengths are dominated by the conversational usability sessions. The oral histories (fossda) show 4–24 turns an hour, which is long-monologue behaviour and probably also under-split diarisation. Half the theme–session pairs are one or two quotes, so a theme usually lands as a point or a short span, not a band across the session.

What each lane needs

Lane	Legibility floor	Event it must survive	Needs at least
Speaker clips, rounded with a gap	4 px per clip	median turn, 10 s	≤ 2.5 s/px
Speaker clips, as slivers (no rounding)	2 px	p25 turn, 5 s	≤ 2.5 s/px
Topic flag with a few letters of label	50 px	median segment, 181 s	≤ 3.6 s/px
Topic flag, pennant only	14 px	p10 segment, 48 s	≤ 3.4 s/px
Sentiment bar, 4 px plus a gap	8 px	p10 quote gap, 30 s	≤ 3.75 s/px

The speaker lane is the binding constraint. Every other lane survives at about 4 s/px, but the turns stop being individual clips beyond about 2.5 s/px. Long monologues (p90, 46 s) still show as clips at 4 s/px (11 px).

Decision: one shared scale per project, fit to the longest session, clamped to 1–4 s/px (15–60 px per minute).

Upper clamp, 1 s/px: beyond this a short project would be blown up to fill the width, with nothing gained. Every topic label already fits at the median, and every turn is a clip.
Lower clamp, 4 s/px: a 60-minute session is then 900 px, which fits the Sessions grid on a laptop. Topics and sentiment still read. The speaker lane switches to slivers (no rounding, no gap, no code) beyond 2.5 s/px. It becomes a talk-pattern texture in which monologues stand out, the way clips do in Final Cut zoomed out.
Longer than about 60 minutes (2 of the 36 measured sessions are over 90 minutes): the slice scrolls horizontally at 4 s/px rather than compressing further. Every lane stays legible, at the cost of one scrolling row.

15-minute session

60-minute session

Both in one project: shared scale, fit to the 60-minute session

Zoom: one control for the project's shared scale

Sessions

Commentary

Post-script, 6 Oct 2026 — what shipped differs from this page in five ways (the measurements and the verdict above stand): (1) no zoom shortcut — ⌘+ and ⌘− stay the browser's text zoom, by decision; (2) the zoom is shared but each slice scrolls on its own; (3) clip colour is the per-turn scene colour for every session with video, not an option — the bullet below about neutral clips is the earlier position; (4) a click on the speaker lane opens the transcript at exactly the clicked moment, not the start of the turn; (5) the top lane is sections, not topics — "topic" below names the s08 data the rates were measured from. Sentiment bars shipped at 6 px, not 4: at 4 s/px two bars at the p10 gap (30 s, 7.5 px) now sit 1.5 px apart rather than 3.5, still separate. Each has an invisible click target 2 px wider than the mark that stops halfway to its neighbour, so close bars stay individually clickable. Themes carry their name in full on a chip at the span's start. Shipped design: experiments/session-tapestry/README.md.
Clip colour is neutral here: moderator muted, participant solid. This follows the design review, so that the sentiment hues are the only colour in the slice. The frame-colour idea from session-tapestry.html can come back as an option for speaker-view recordings.
Themes are drawn as spans from first to last quote, not as a tag at the first quote. Per the measurements, most are points or short spans. A one-quote theme gets a short cap so it can still be hovered.
Turns are synthetic: lognormal, median 10 s, p90 46 s (σ ≈ 1.19). Participant turns run ×1.5 and moderator turns ×0.6. Topic segments are lognormal with a median of 181 s, and quote gaps lognormal with a median of 64 s.
Hover and click: hovering the speaker lane raises the flag for the topic under the pointer, with its full label drawn over its neighbours. Labels at rest are truncated so nothing overlaps. Clicking a clip jumps to the transcript at the start of that turn, and clicking a flag jumps to the start of that topic. (In the mockup the destination is shown in the frame's title bar.) The hit area is the whole speaker lane, so slivers at 4 s/px are as easy to hover as clips.
Zoom sets the project's shared scale, so every expanded slice zooms together and the cross-session comparison survives. The control is borrowed from Photos and the Final Cut timeline: a 64 px slider between zoom-out and zoom-in magnifiers, about 104 px in all, placed in the Sessions header and not repeated per row. The left end is Fit, the default (fit to the longest session, clamped at 4 s/px). The right end is 0.25 s/px, for reading individual turns in a dense stretch. The steps are logarithmic, and ⌘− / ⌘+ / ⌘0 (fit) follow the View-menu convention. Double-clicking the slider also fits. Beyond Fit the slices scroll horizontally.
Clip colour (since this page): the per-turn scene colour, settled in the colour lab (experiments/session-tapestry/colour-lab.html) and computed by bristlenose/utils/scene_colour.py. The real-data build (experiments/session-tapestry/build.py) draws it inside the actual Sessions grid.
Not settled: whether oral-history projects, which have few long turns, want a coarser floor. Their speaker lane is legible at any scale. Their topics are the dense lane.