Skip to content

42 — Watch faces: what it takes to design a good one

2026-11-16 · research report, kept up to date with the lessons of each design round · web research (about 350 sources: watch platforms’ guidelines, Garmin’s docs and manuals, horology, perception and safety-critical display research, diabetes care, CGM watch apps and their issue trackers; about 220 store reviews and forum threads by people who wear one) and Sukrara’s own rules · no design: each face’s design lives in its own place

The owner wants Sukrara watch faces, the first for their Garmin Descent Mk3i, later more designed by other agents. The brief, in their words: a face that “doesn’t scream ‘I have diabetes’, yet the owner is able to read the data he needs”; “the artistic and creativity aspect is very important”, while being rigorous and useful day to day; digital time, seconds, battery time remaining, month, day of month, day of week, glucose value and trend arrow; custom data fields like Garmin’s own faces; and a face that “reacts to the blood glucose and context/situation that is going on and actions [the] user has taken”.

This report is the reusable part: what a person with diabetes needs from the wrist, which data and states a face has to carry, what makes a face legible, calm and discreet, the ways each datum can be drawn, and the facts of the platforms. On purpose it holds no design and no preferred style, so every face can be its own.

  • Read it before sketching. It is a toolbox and the evidence behind each trade-off, not a style guide: no layout, palette or look in it is a recommendation, and the examples are there to widen the space, not to be copied.
  • The hard rules are few and listed in §4; everything else is a trade-off each face decides, with its evidence.
  • Platform limits are facts, not design rules (§10). Design for the person; use a platform’s features where they help; a limit is solved later, not designed around.
  • Your face’s design (sketches, mockups, choices) goes in its own place. What a design round teaches that helps any face (a reason, a usability fix, an idea that failed and why) goes in §12.
  • The full research notes, with every source: evidence/42-watch-faces (platforms, Garmin, horology and encodings, human factors, diabetes, user voices).
  • Glances are short. About half of all smartwatch uses are time checks of about 1.9 s; half of 798,423 logged sessions lasted 5 s or less, partly because that’s when screens time out (Pizza et al. 2016, Visuri et al. 2017). Apple’s budget is “about five seconds or less” per interaction (WWDC15); people check the time about 150 times a day (Wear OS). A glucose face is read inside those 2 s, many times a day.
  • The wrist is the discreet place. “Smartwatch facilitated discrete access” to glucose (Sehgal 2024); Dexcom markets exactly the meeting, the date, the dinner (Dexcom). Pulling out a phone is visible and slow; a glance at a watch reads as checking the time.
  • For some it is the only warning. About a quarter of people with type 1 have impaired awareness of hypoglycaemia (27 % pooled, 17–21 % in CGM-era cohorts: PMC11178233).
Moment What a glance must answer What the face could know (on the watch or from Sukrara)
Any time, 2 s The time; am I OK?: value, direction, and is it current Reading age
Meeting, date, meal out, class, cinema The same, without anyone noticing; no sound, no colour flash Recent insulin / carbs; calendar
Around a meal Did I bolus? How long ago? (the most asked-for thing next to the value: docs/39); a rise within ~2 h of a bolus is expected (Brazilian statement 2021) Last treatment, kind, amount, time
Night, bed Lows matter most and are least trustworthy (compression lows, see 1.3); a check that doesn’t wake a partner (“so bright that I can’t even sleep”: xDrip #1137) Time of day, sleep schedule; evening exercise or alcohol earlier
Sport Zone readable in motion (“it is good enough to see ‘green’”: AAPS #2271); the low line moves up (1.4) Activity, heart rate (on the watch)
Before driving Above the driving line (≥ 90 mg/dL ADA, ≥ 5 mmol/L UK) and its trend; nothing to operate while driving Not reliably detectable
Phone far, dead or locked Whether the value is current, and which link broke (urchin shows each hop) Reading age; phone connection
Shower, swim, dive A gap is expected (Bluetooth doesn’t cross water); before a dive, a sequence of checks (1.4) Water activity; dive (native Garmin only)
Morning after a night out Lows peak the next morning (1.4) User-logged
Sick day Ketones and frequent checks User-declared
Travel A time-zone change shifts insulin timing (PMC7962589) Time-zone change
Alerts quiet or muted What is quiet and until when (users lose track: GDH #259) Sukrara’s alert status
Sensor change day Days left (carry a spare: GDH #295); warm-up has no value Sukrara’s sensor life
  • Ranges (international consensus, Battelino 2019): five named states, not three: very low < 54 mg/dL (3.0), low 54–69, in range 70–180 (3.9–10.0), high 181–250, very high

    250 (13.9). Targets: > 70 % in range, < 4 % below 70, < 1 % below 54. Pregnancy uses 63–140; older or high-risk adults > 50 %. → Zones are personal settings, never fixed numbers in a design. Sukrara’s own bands and limits: §3.

  • A trend arrow looks back (about the last 15 min) and implies the next 15–30 min; it can still point down after glucose has levelled, so guidance says to read it with the last ~10 min of the curve (Brazilian statement). → A short history disambiguates an arrow.
  • The same glyph means different speeds on different systems: Dexcom has 7 states (↑↑ > 3 mg/dL/min, ↑ 2–3, ↗ 1–2, → < 1), Libre 5 (no double), Medtronic up to three arrows (table). → A face that draws its own arrow chooses its vocabulary; Sukrara has one (§3). Every surface must draw the same arrow or people notice (AAPS #3414).
  • Context changes the meaning of a rise or fall: rising right after a meal bolus is expected, rising 4 h later may need action; arrow-based corrections aren’t advised within ~2 h of a bolus (Endocrine Society on Kudva 2018). → Time since the last insulin is the context that decides between “fine” and “act”.
  • The number trails the blood by 5–25 min (interstitial lag about 5–8 min, more after meals: Basu 2013, Sinha 2017). During fast moves (exercise, rescue carbs) direction and speed matter more than the exact value.
  • Least trustworthy: at low values, in the first 12 h of a sensor, and at night, where lying on the sensor makes false drops of more than 25 mg/dL that rebound in 15–30 min (median 45 min long), roughly one every 3–6 nights (docs/33, Mensh 2013).
  • Stale data is a known safety failure: a Dexcom G7 app froze on a stale value (FDA class 2 recall, 2025) and G7 apps including the watch app failed to warn of missed data (FDA). Conventions users know: Dexcom shows dashes after 10.5 min; Dexcom and Libre alarm signal loss at 20 min; DIY faces strike or grey the value at 10–15 min (evidence: diabetes §6). Sukrara’s steps: §3.
  • Sensor life: warm-up 30 min (G7) to 60 min (Libre 3), wear 10–15 days, G7 has a 12 h grace period. Warm-up, the end countdown and “no sensor” are real, recurring states.
  • Exercise moves the danger line up and into the night: start aerobic exercise at 126–180 mg/dL; during exercise the low alert at 100 mg/dL (5.6); 4.4–10 mmol/L for 90 min after; after evening or long exercise a night low alert at ≥ 80 mg/dL (Riddell 2017, Moser 2020). Aerobic effort lowers glucose, sprints and competition stress can raise it: “exercising” alone doesn’t predict the direction.
  • Driving: UK “five to drive” (≥ 5 mmol/L; 4–5 eat first; check before and every 2 h; after a hypo wait 45 min), ADA: raise it if < 90 mg/dL (Diabetes UK, ADA 2024). The DVLA’s 2019 wording said the CGM must not be actively used while driving; whether a glance at the wrist counts as active use isn’t addressed anywhere.
  • Night: 55 % of severe hypos happen during sleep; only 1 of 16 people with type 1 woke at 2.2 mmol/L against 10 of 16 controls (Schultes 2007). Heart rate from a watch detects early-night lows weakly (AUROC 0.78–0.83, PPV 0.13–0.25) and late-night lows not at all (PMC11851596): context, never a substitute.
  • Alcohol: lows peak the morning after (5 of 6 participants between 10:00 and 12:00), and a hypo is easily taken for drunkenness (Turner 2001, Browne 2014).
  • Diving (the Descent is a dive computer): ≥ 150 mg/dL (8.3) and stable or rising on three checks, 60, 30 and 0 min before; no dive below 150 or above 300; the buddy signal for a low is an “L” with thumb and index (DAN). Sensors are rated to 1–2.4 m and Bluetooth fails underwater, so a face’s role is on the surface: before, between and after dives. A protocol is a sequence with a trend, not a single number.
  • Sick days add ketones (≥ 0.6 mmol/L act, ≥ 1.5 high, ≥ 3.0 DKA range: ISPAD 2022); continuous ketone sensors are coming. Menstrual cycle: more time above range in the late luteal phase, more below around menstruation (PMC9962060).
  • Stigma is the norm, not a niche: 76 % of people with type 1 say diabetes carries stigma (n = 5,422, Liu 2017). Adults inject in toilet cubicles, avoid swimming to hide a pump, hide it from partners and at work, get blamed for eating sweets (Browne 2014). Teens hide devices at first, some later use them to educate (PHCFM 2023).
  • Alarms in public embarrass, people turn them off at work and social events, and some feel alarms “signal personal failure” (Keeping safe, Natale 2023). → How a high value looks to the wearer carries weight too: a red number can read as a verdict.
  • Abstract displays already exist in the wild: Nightscout Clock’s faces where glucose drives a unicorn’s mane, a race car that stops when data is stale, a dragon whose fire goes out (nightscout-clock); SugarPixel’s emoji and few-lights night mode (Diabetech); a wristband that shows a colour instead of a number (UW–Madison).
  • Gap: no study measures how bystanders perceive a number, a colour or a graph on someone’s watch. Stigma research predates wrist displays.
  • More alarms isn’t more safety: in 95 Dexcom users, lower high thresholds doubled high alarms (30.9 vs 14.2) and went with longer highs; 9 % had the low alarm off (Lin 2020). Half of former CGM users cite too many alarms; patients responded to 29 % of alarms (review 2025).
  • Some states are designed to break discretion: Dexcom’s urgent low (55) and failure alerts can’t be silenced and escalate to sound (Dexcom). Sukrara’s equivalents (extremely high and no data are never muted by a treatment) are in contracts/alerts/SPEC.md §2.
  • Vibration is less embarrassing than sound and noticed in noise (review 2025); awake users recognise wrist-vibration patterns 89–99 % of the time (GoodVibes); in device studies pulsed vibration woke 100 % of sleepers and continuous 92 % (vendor summary), though a watch alarm can be dismissed without truly waking. Users ask for one small tick per reading while falling fast (GDH #226) and for the watch to alert when worn, the phone when not (GDH #230).
  • The ecosystem is split on whether the wrist alerts at all (Garmin’s Dexcom app and Sugar Sense: never; Juggluco, AAPS, xDrip+, GDH: yes). A face is the display; the alert itself is Sukrara’s (phone, and the watch when it can).

From about 220 Garmin store reviews of 15 CGM faces and apps, Play reviews from watch users, Diabetes UK forum threads, Hacker News and a study of young adults given smartwatches (all quotes and links; Reddit stayed blocked). Many = 5 or more independent voices, several = 2–4.

  • The point is not taking the phone out (many): “keeps me from pulling out my phone constantly”; “I don’t have to stop running to check my blood sugar”. So the face has to answer the whole question in the glance: value, direction, freshness.
  • It must be there the moment the wrist comes up, and in always-on (several): “Every time I raise my hand, I’ll see this data”; a complaint that the dimmed face shows only the time in the dark; “please implement the ‘screen always on’… Very important for me”.
  • Stale data is the normal case, not an edge case (many; most 1–2★ reviews): “the number simply grey”, “— (no data) even after 30 minutes”, “stop updating for hours”, upstream servers and logins breaking. How it’s shown decides whether the face stays useful: “when there is no BT connection the watch face is usless… a big BT symbol accross the watch” (a stale mark must never cover the time); people value seeing when the last reading came and why (“No phone connection”).
  • Wanted next to the value: delta (several; a face that showed it per minute got “Incorrect BG delta”: use the unit people expect or say it), minutes since the reading (several; hidden or cryptic versions confuse), insulin on board “more than my steps” (several; Sukrara computes no IOB today), watch battery (many; one face shows it only when low), a 3 h average. Heart rate, steps and seconds are wanted by about as many as want them gone (“in night I don’t need them”).
  • The hierarchy is personal (several each way): “the BG is more important than the time itself” vs “clock a bit bigger”; “option to swap time and BG positions”. Same for the graph: “Finally can have Dexcom graph on my watch face” vs “option of not displaying the graph, and instead a larger display of current reading”; a graph that “always starts blank” annoys; a graph line that looks like a stale strike-out confuses.
  • Small grey secondary text fails (several): “hard to read since it’s grey… for older eyes”; “not legible in very dark grey and small font”; one app is called No Glasses Required CGM.
  • Expected details: mmol/L everywhere, including the range settings (8+ reviews across 4 faces); Dexcom’s double arrows; settings with “a sample of what you’re changing”.
  • Non-numeric works for some: “Very easy to read at a glance, without needing to necessarily look at the BG number” (a dial face). Emotion counts: “The little emoticon at the bottom warms my heart every time I see it”.
  • Discretion is the gesture as much as the graphics: “discreet monitoring of my levels during meetings”; “one shuffle of my sleeve and one tap” against a phone’s unlock; but “I still don’t want people to get the wrong idea” (a long look at the watch reads as boredom). In the dark the screen itself is the problem: “I would not use a smart watch in the cinema either, the screen is also bright”.
  • Parents are a large share of the voices (many): they read their child’s glucose on their own watch, and want the child’s watch to work without a phone at school. A child’s watch failed: “vibrations… very short, and the kid did not fell it”.
  • The wrist is the quiet first step of an alert (several): “notify my iWatch (as the vibration will wake me) but not… my iPhone” (a partner asleep); buzz at 70 before sound at 55. Too much is intrusive (“a bit intrusive with alarms on my wrist. Instead I have G7 as a complication”), and a watch that suddenly made “VERY LOUD noises” surprised.
  • A good face prevents alarms: “so easy to just look at the watch and have a toffee or bolus before blood glucose gets low / high enough for the alarms to go off”.
  • Contexts named: trail running, cycling with gloves, physical work “in the bush”, a bus driver (“I need to ensure my glucose does not go low while driving”), a dance class, the cinema; and a wish after drinking for “a little thing on my watch… ‘eat something before you go to sleep’”. Nothing found on swimming or diving.
  • Not found: first-hand stories of strangers asking about the number or colours drawing looks (that lived mostly on Reddit).

Asked by the owner in bold; the rest is what the evidence says people use or ask for. “Sukrara has it” means the phone already computes it (docs/FEATURES.md).

Datum The question it answers Precision needed Changes Source Notes
Time (digital) What time is it? 1 min 1/min Watch Always present; 12/24 h is the watch’s setting
Seconds Exact second; is it alive? 1 s, or only “running” 1/s Watch Usually only while the face is awake (§10)
Day of week Which day? Exact 1/day Watch Locale words (en/es); first day of week is a setting
Day of month Which date? Exact 1/day Watch
Month Which month? Exact 1/month Watch Rarely needed at a glance; a calendar question
Battery time left When must I charge? ~1 day Slow Watch Garmin gives days left directly
Glucose value Where am I? 1 mg/dL / 0.1 mmol/L, or the zone only 1/min (Libre 3) to 1/5 min Sukrara mg/dL 40–400 and LO/HI; mmol/L 2.2–22.2 with one decimal
Trend Where am I heading, how fast? 5–7 classes, or continuous 1/min Sukrara One trend for every surface (§3)
Reading age Is it current? “now” / minutes / old Continuous Sukrara + watch clock The most safety-relevant datum after the value (1.3)
Zone / range word Am I in range, how far out? 7 bands With the value Sukrara Never by colour alone
Delta How much since the last reading? 1 mg/dL 1/min Sukrara Asked for in every DIY face; redundant with a continuous trend
Short history What has it done lately? Shape 1/min Sukrara Disambiguates the arrow (1.3)
Last insulin / carbs Did I bolus? How long ago? How much? Minutes; units / grams On action Sukrara Kind matters (rapid, long, carbs)
Quiet / muted state Is something holding alerts, until when? Minutes left On action Sukrara Quiet after insulin 60 min, after carbs 15 min (defaults)
Alert needed Must I act now? Kind On event Sukrara Words exist: Eat carbs, Take insulin (card words)
Sensor life Days left; warming up; ended Days → hours → minutes Slow Sukrara Warm-up never as a number that passes for glucose
Time in range How is my day going? 1 % (5 % meaningful) Slow Sukrara Motivating for some, stressful for others: opt-in (docs/39)
Phone / link Is the chain working? Up / down On change Watch
Prediction Where will I be in 15–30 min? Zone 1/min Planned Predictive alerts cut time below range by > 40 % in teens (Drug Topics)
Heart rate, body battery, stress, steps, floors, calories Body and activity context As the watch gives 1/min–slow Watch Exercise context for glucose
Sunrise / sunset, weather, altitude, pressure, compass Outdoor context As given Slow Watch Weather needs the phone
Notifications, alarms set, do-not-disturb Is anything waiting? Count / on-off On change Watch
Second time zone Travel, people abroad 1 h 1/min Watch Descent’s bezel has an ALT TZ mark
Dive: surface interval, no-fly time Can I dive / fly? Hours Slow Watch (native only) Not readable by third-party Garmin faces

Sukrara decides these once, for every surface (docs/FEATURES.md, contracts/graph, contracts/alerts/SPEC.md). A face shows them; it never decides them differently (CLAUDE.md “One graph everywhere”: a face is a surface).

State True when (defaults) The face must convey Never
Fresh Reading ≤ 90 s old Value, trend, zone —
Late 90 s – 5 min The value is from a few minutes ago —
Stale 5–20 min Old: Sukrara strikes it through and greys it; the trend becomes a hollow ring Show it as current
No readings 20 min – 3 h No reading since + time; no value as if live A number that looks live
Long gap ≥ 3 h No value, no history
Very low / low / in range (tight 70–140, target 141–179) / high / very high / extremely high Bands: ≤ 55, 56–69, 70–179, 180–249, 250–299, ≥ 300 (bands.json); the user’s own thresholds Which band; a word under an out-of-range value (Low, Very high…) The band by colour alone (#26); very high’s orange and low’s red look alike to protan / deutan eyes
LO / HI Sensor says < 40 / > 400 LO / HI, no number, no trend 39 or 401
Trend Rate from the readings (Trend arrow setting) Direction and speed: drawn angle = clamp(rate / 3, −1, 1) × 90°, doubled from ±3 held 2 min; text arrow bands ±1/2/3 with 0.2 hysteresis (docs/34) A trend computed differently from the phone
No trend Too few readings, LO / HI, stale A hollow ring where the arrow goes (not a dot, not ?) A flat arrow (reads as “steady”)
Alert needed Low / high side out of range, not quiet The need: Eat carbs / Take insulin (card words), or the alert kind
Quiet after a treatment Insulin mutes highs 60 min, carbs mute lows 15 min; or glucose clearly heading back (20 min) That it’s handled, and until when (10m left) Hide that an alert is being held
Muted by the user Mute of a side, or Mute all (with or without an end) Muted and time left
Warm-up New sensor, no reading yet (60 min + 15 grace) Warming up · 42m in words, the disc filling A figure that could pass for glucose (#27)
Sensor ending Last 8 h (setting 1–24) A countdown
Sensor ended Past its end, readings stopped Sensor ended (checkered flag) The last reading’s time (looks like a gap)
Ended but still sending Past its end, readings arrive The value + Ended · Replace sensor
No sensor / not monitoring Nothing paired Not monitoring, plainly An empty slot that looks like a gap
Unit mg/dL or mmol/L (per device) Widest values fit: 400, LO, 22.2, 10.0
Watch can’t reach Sukrara No data delivered to the watch The age keeps counting from the reading’s time, so it shows as late / stale by itself Count age from when the watch received it

Everything else in this report is a trade-off. These are not, because breaking them makes the face unsafe or makes it contradict the phone:

  1. Never show an old value as current. The age is part of the value (§3, 1.3 recalls). Wear OS and Fitbit replace stale live numbers with -- rather than let them mislead (Wear OS, Fitbit).
  2. Same decisions as every other Sukrara surface: bands, words, trend, freshness, sensor states (§3).
  3. Never colour alone for anything that matters: colour-blind eyes, tinted and night modes (Apple’s tinted mode, Garmin’s Red Shift re-tints the whole screen), always-on dimming. Shape, position, a word or a glyph carries it too (Apple HIG).
  4. The wearer can always get the exact number and its age, in at most one action, whatever the face’s discretion level (the brief: “able to read the data he needs”).
  5. Warm-up, ended, LO / HI and no-data never show a figure that could be read as glucose.
  6. Every state has a visible way to see it, change it and leave it (CLAUDE.md Complete flows): a mode the face enters (discreet, quiet, night…) shows that it’s on and how it ends.
  7. The time is always readable, in every state and mode.

Numbers for sizes, contrast, colour and motion (full notes).

  • People read a watch at 28 cm (SD 5) with the watch pitched 50° and the line of sight 10° off the face’s normal: slightly oblique, closer than a phone (Blascheck 2019).
  • On the Descent 43 mm (12.8 px/mm), at 28 cm 1 arcmin ≈ 0.08 mm ≈ 1 px. The whole face is about 6° wide.
Character (cap) height Arcmin At 28 cm At 33 cm Source
Non-critical, not time-critical 10′ ≈ 0.8 mm · 10 px 12 px FAA HFDS
Minimum 16′ ≈ 1.3 mm · 17 px 20 px ISO 9241-303, MIL-STD-1472F
Preferred 20–22′ ≈ 1.7–1.8 mm · 21–23 px 25–27 px ISO 9241-303
Preferred on colour displays 30′ ≈ 2.4 mm · 31 px 37 px Vendor summary of MIL-STD

These are for steady reading at a desk; a 2-s glance, walking, an oblique wrist and sunlight all need more. Garmin’s smallest built-in fonts (19–27 px line height, roughly 13–19 px caps) sit at or under the minimum: fine for labels, not for the value (sources).

  • A time check takes about 1.9 s; time is 50 % of watch interactions but 14 % of the time spent; more than 70 % of activity-tracker use is glances with no further interaction (Pizza 2016, Gouveia 2016). The face is the product; detail screens are secondary.
  • Comparing two values on a watch: donut 159–216 ms, bar 245–285 ms, concentric radial bars 1.5–1.8 s, about 6× slower (Blascheck 2019). Reading a proportion: bar 318 ms, radial 397 ms, text “70 %” 503 ms and the most errors; people rated the ring the most beautiful and the bar the most efficient: about 80 ms of speed against beauty, a free choice. A single live value (like glucose) is untested (Blascheck 2023).
  • Walking lowers accuracy reading watch visualizations; radial progress reads best at the quarters, and people asked for tick marks (Grioui 2024). With the watch flat, OLED contrast drops so far the researchers had to recolour the stimuli (Blascheck 2023).
  • What real faces carry: a median of 5 data items, icon + text the most common form, charts rare; of 358 best-selling faces 60 % are digital, 80 % have no animation and none animates data; 62 % of people never change their face: a face is chosen once and lived with (Islam 2020, Islam review).
  • Six qualities of glanceable feedback: abstract; part of existing routines; compared with a target or norm; actionable; builds a checking habit; a door to more detail (Gouveia 2016).
  • Accuracy order (Cleveland & McGill, replicated by Heer & Bostock): position on a common scale > position on unaligned scales > length, direction, angle > area > volume, curvature > shading, saturation. A ring encodes by angle; a tape or bar by position.
  • Bertin: only position and size read as quantities; lightness is ordered; hue is for categories, not magnitudes (summary).
  • Pre-attentive in < 200–250 ms, whatever the clutter: hue, intensity, size, orientation, length, curvature, closure, flicker, motion. A target defined by a combination of features doesn’t pop out, and colour variation masks shape patterns (Healey). An urgent state carried by one unique feature pops out; one carried by a combination doesn’t.
  • A number alone hides the margin: “it may be difficult… to determine the margin relative to targets or limits… a scale, dial, or tape may be needed”; digits should match the data’s accuracy (FAA AC 25-11B §5.7). The three-pointer altimeter took ~7 s to read with > 11 % big errors; the fix was a counter plus a pointer (summary).
  • Sparklines: Tufte’s own worked example is a patient’s last 80 glucose readings: the normal range as a grey band, the newest point tied to the printed number by the same colour accent, no “vulgar stoplight” colours (Tufte).
  • Gestalt: on a 6° face, proximity, common region and similarity alone tell which number goes with which unit.
  • Humanist faces read faster at a glance: Frutiger cut total glance time by 10.6 % against a square “technical” grotesque (Eurostile) in a driving study (Reimer 2014).
  • Lowercase took 26 % more time to read at a glance, condensed 11 % more; capitals read better partly because they are bigger, but confuse more letters (Sawyer 2017). Small text cost 27 % more glance time, tight spacing 9 % (Dobres 2018).
  • Digit confusions: segment digits that differ by few segments are confused most (1/7, 3/8, 5/6, 6/8/9/0); a redesigned seven-segment font halved errors (Gunderson 1991); segment decimals invite out-by-ten errors (Thimbleby). An mmol/L value (5.6, 15.6) is exactly that error surface.
  • Fixed-width figures keep a changing number from changing its own width each minute (11:11 vs 18:18).
  • Rolling or animated digits must not suggest a motion that doesn’t match the real trend (AC 25-11B §5.7.2.1).
  • 1 in 12 men and 1 in 200 women have a colour-vision deficiency, mostly red–green (NEI).
  • AC 25-11B §5.8: colour coding always with a second code (size, shape, place); at most 6 coding colours; no distinctions by shades of one hue; pure blue is too dark to carry anything important; ageing eyes lose red focus and blue–green discrimination. Pairs to avoid: saturated red/blue, red/green, blue/green, yellow/green, magenta, blue or red on black.
  • Learned meanings the wearer brings: red warning, amber caution, green normal, white for values and scales (aviation, cars, CGM apps). The Descent itself codes its dive bars green / yellow / red (ascent rate, tissue load).
  • Contrast floors: 3:1 for graphics needed to understand content (WCAG 1.4.11); in cars 5:1 at night, 3:1 by day, 2:1 in direct sun (ISO 15008).
  • OLED in sunlight: contrast “declines sharply as ambient light gets brighter” (Chen 2017); a palette that works indoors can wash out on a beach or a boat.
  • Night: full dark adaptation takes ≥ 30 min, bright light for > 1 s spoils it; red light preserves it but “severely distorts colors” and erases red’s meaning (FAA). In dark surroundings light-on-dark text gets harder to read as it gets smaller (Dobres 2017): a tension with dark always-on faces.
  • Underwater, reflected red disappears by ~5–6 m; an emissive screen at 30 cm is probably much less affected (no study found).
  • Change blindness: a change between two glances is invisible, because the eye has no memory of the last frame (Simons & Rensink 2005). If “something happened since you last looked” matters (a new low, a treatment, an alert), the face has to encode it.
  • Motion is the strongest peripheral signal: colour targets in the periphery were missed 25 % of the time, motion under 2 %; motion “can contribute strongly to distraction and irritation” (Bartram, Ware & Calvert 2003). It reaches bystanders’ eyes too.
  • Trend indicators need hysteresis and damping, and damping adds lag (AC 25-11B A.4.5); Sukrara’s trend already settles that balance (docs/34).
  • Blinking: 0.8–4 Hz, cancellable after ~10 s (FAA); medical standards discourage flashing text. Photosensitivity: ≤ 3 flashes a second, and a flashing area under 0.006 sr: the whole face at 28 cm is 0.0093 sr, so a full-face flash fails (WCAG 2.3.1).
  • Reduced motion: non-essential animation must be possible to turn off (WCAG 2.3.3).
  • Case’s calm-technology principles: the smallest possible amount of attention; inform and create calm; use the periphery; work even when it fails (default to a usable state); respect social norms (calmtech.com).
  • Ambient-display heuristics: “sufficient information design” (just enough); consistent, intuitive mapping; visible state with noticeable transitions; easy path to more detail; unobtrusive until needed; notice it “because of a change in the data… not because its design clashes”. In the field, people thought one ambient display “was broken” (Mankoff 2003): a calm face must still look alive.
  • Notification levels run user poll → change blind → make aware → interrupt → demand attention; ambient displays by definition carry “important but not critical” information (Pousman & Stasko 2006). A glucose face spans the whole scale, ambient in range and demanding at a very low.

6. Principles from watch platforms and watchmaking

Section titled “6. Principles from watch platforms and watchmaking”

Each with its reason; the face decides how far to follow.

  • Rank by the precision and urgency each datum needs, not by convention. The B-Uhr gives the outer ring to the minutes because navigators needed them; the regulator moves the hours to a subdial “where imprecision does not matter much” (Quill & Pad). The longest radius, the biggest size or the strongest contrast goes to what needs it most.
  • Contrast separates value from context, not only size: Apple’s heart-rate slot draws the value high-contrast and the graph and labels low-contrast grey (HIG Complications).
  • “The great dials say remarkably little” (The Horologue). Every glyph competes with the data; a fourth central hand or a third subdial makes people “stop and look for longer” (Fratello). Braun used one typeface and “black and white and a bit of metal” (Dezeen).
  • One message vs many slots are both official: Android Wear (2014) said “stick to one message” and “a watch dial is a timeline; add data to it” (AOSP doc); Infograph and Wear OS allow 8 slots. The owner’s list is already 8 data; each extra slot spends glance time.
  • A slot earns its place by staying current: “a static complication… may be less likely to remain in a prominent position” (HIG).
  • Digits for “what is the value”, position for “how far from a line / how long until”. Digital time reads 3.5–4× faster with 10× fewer errors for absolute time (1971 review; a 2023 study agrees even for people who prefer dials, Springer); analogue reads faster for time differences (MDPI 2024). Glucose needs both questions: the value, and the distance to a threshold.
  • Precision is a design variable. A single-hand watch reads “almost to the nearest minute” and its buyers don’t care whether it is 13:29 or 13:31 (MeisterSinger); word clocks round to 5 min and add corner dots for the rest (QlockTwo). Layered precision lets the glance stop early; rounding needs a rule for which way it goes.
  • One unmistakable anchor: a round face without numerals hangs on one landmark (the B-Uhr’s triangle, Grand Seiko’s double index, Movado’s single dot) and the wrist rotates. On the Descent the printed bezel ring is a second frame of reference (§10).
  • Pointers must stay distinguishable in the worst light: ISO 6425 requires the minute hand to be told from the hour hand at 25 cm in the dark, and a visible sign that the watch is running (ISO 6425 sample). “Is it alive?” has a precedent in a standard.
  • One accent, one role, borrowed meaning: Braun’s alarm clocks all had a yellow seconds hand; Hilfiker’s red seconds hand is shaped like the dispatcher’s signal baton (Wikipedia). An accent copied from a real-world signal arrives with that signal’s meaning, and “avoid colours used in places they may have a pre-conceived meaning which does not apply” (Pebble).
  • Zones on the scale, a neutral pointer: aviation airspeed arcs (white, green, yellow, red line) put the meaning on the scale, not on the needle (Wikipedia / FAA). The pointer’s colour then stays calm.
  • Black is the ground on emissive screens: it saves power and merges with the bezel, which “will provide visual padding” (WWDC15); “every pixel uses power”; Garmin: light-on-dark, “dark-on-light… will look out of place on an AMOLED display” (Garmin UX). Avoid full-screen colour on views that stay on screen (HIG Color).
  • Night: Apple Ultra’s Night Mode turns faces red automatically from the light sensor, down to 1 nit “so as not to disturb people close by” (Apple); green lume (~515 nm) looks brightest in the dark, blue dimmer (Aquastar). The Descent has Lume (turquoise glow from the light sensor) and Red Shift natively.
  • Hand angles carry mood: 10:10 reads as a smile and raised liking modestly; 8:20 reads sad (Karim 2017, small samples). A data pointer’s resting angle may be read as an emotion: a falling trend drawn as a frown is a judgement.
  • One family, few sizes, no thin weights: Apple: avoid Ultralight, Thin and Light; minimum 12 pt, lines ≥ 2 pt “especially when the wearer is in motion” (HIG Typography, Complications); Wear OS: 12 sp essential, 10 sp other (quality); Samsung: one font, no drastic size jumps (typography). In physical units for the Descent: §5.
  • A numeral’s shape is legibility and identity at once: railways banned Roman numerals for big Arabic ones; Max Bill drew his own; the Tangente’s narrow numerals structure an empty dial (Monochrome, Gear Patrol). Display-only numeral fonts are cheap (TTMM shipped fonts of ten glyphs: TTMM).
  • Size for the widest value (Zepp sizes every field for “000”/“00000”: Zepp): 400, 22.2, HI, --, Spanish weekday and month words, a 3-digit heart rate.
  • Stay inside the bezel and away from the edge in always-on: the system shifts the picture a few pixels to protect the panel, so edge elements can be pushed off (Wear OS, Samsung); Garmin asks faces to move elements up to 4 px a minute.
  • Thin rings at the very edge fail on round screens; use thick rings or inset them; each text row has a different width, so don’t reflow text (Pebble round). “Be open to creating a new UI” for round screens instead of porting a rectangular one.
  • The edge can carry a reading: Apple’s Modular Ultra uses “the outermost edge to present real-time data, including seconds, altitude, or depth”; text can curve along nearly 180° of the bezel (Apple).
  • Items along the curve read as equal or as flow; motion out from the centre reads as expansion (Samsung circular UX).
  • Not a dimmed copy: “Do not simply lower the brightness level of an active state” (Samsung); replace large bright areas with outlines and dim interiors (Apple’s Numerals Duo); avoid much white or blue, prefer light grey and thin weights (Garmin).
  • Keep the layout and let motion settle between modes: “Avoid making distracting interface changes when Always On begins or ends”; “Gracefully transition motion to a resting state” (HIG Always On). Apple’s own faces still change shape continuously (Metropolitan’s numerals “rotate to become pills when your wrist is down”).
  • Budgets: 10–20 % of pixels or luminance lit, once-a-minute updates, usually no seconds (table in §10). On the Descent, always-on cuts battery from 10 to 4 days.
  • Motion means something: Spring Drive’s glide “the natural flow of time”, deadbeat seconds were built so observatories could read the exact second, Hilfiker’s hand pauses ~1.5 s at 12 waiting for the master clock (Galaxus). Flowing, measured or synchronised with something outside are three different messages.
  • A change of rhythm can be the alarm: Citizen and Seiko seconds hands jump in 2-s steps when the charge is low (Citizen) — a second datum and a warning without one more pixel, readable only by those who know.
  • Motion only when the data needs it: Samsung’s example blinks only when heart rate is outside 60–100 (Samsung); Pebble animates “to draw the eye to updated or changing information” and lets users turn it off. “Make motion optional… avoid using it as the only way to communicate” (HIG Motion).
  • A transition can mark the update: ClockClock’s 24 clocks spin and line up into digits once a minute; Apple’s Flux numerals “snap into new forms when the minute changes”. A new glucose reading is a moment a face can mark.
  • People expect to choose: data slots, colour themes, hands/numeral styles; Wear OS listings must show more than one permutation (quality WO-G6). The most loved Garmin faces sell exactly this (Crystal: data fields, indicators, meters, a dozen themes; its 2026 rebuild added full-colour always-on, per-element always-on toggles, on-watch settings “a million times easier”, custom colours: the5krunner).
  • On the watch: binary choices and short lists; the rest on the phone (Android Wear 2014). Settings that only exist on the phone confuse people (Segment34 FAQ).
  • A slot must handle everything a source can say: no data, the widest value, a unit, an error. On Wear OS the source sends data and the face decides how it looks (“the data source plays no part in determining how the data should be represented”: WFF); a slot that shows “null null” becomes a 53-comment issue (xDrip #303).
  • Previews are judged first: store previews standardise 10:09 and default data (Zepp); the face’s default state is its first impression.

A catalogue to widen the space, from watchmaking, conceptual watches, smartwatch faces and instruments (full catalogue with every source). Glance: L one look, M a second look or a learned key, H decoding or a press. Discreet: P public (anyone reads it), C coded (looks like watch furniture; needs a key), H hidden until asked. Entries marked † are invented, untested.

Encoding Example Precision · glance Discreet Watch out
Digits, any size relation (hours big / minutes small, stacked, side by side) Most digital faces 1 min · L P The hierarchy between H and M is a choice
Segment / dot-matrix digits Pulsar 1972 (27 dots a digit), Segment34 1 min · L P Lit-pixel budget
Mechanical-disc digits (jumping) Lange Zeitwerk, IWC Pallweber 1 min · L P Big area
Words, with dots for the remainder QlockTwo, Pebble Fuzzy Text 5 min (1 with dots) · L–M P One language
Wandering numeral travelling a minute arc AP Starwheel, Urwerk 1 min · M C Uses part of the dial
Jumping hour + retrograde minute arc Chanel Monsieur, VCA Pont des Amoureux 1 min · L P The snap-back draws the eye
Hands (two, one, 24 h) universal; MeisterSinger single hand; Solar Dial 1–15 min · L P The owner asked for digital time; listed for other faces
Minutes outer, hours inner B-Uhr Type B 1 min · M P Hour demoted
Discs turning under a fixed mark Ressence, Ziiiro swirls 1 min to “about quarter to” · L–M C Approximate
Count of lit elements (rings, dots, squares of ten) Tokyoflash RPM, Nooka Zot, TTMM 9TTMM 1 min · M→L with practice C Puzzle fatigue (“I just want to know the time”)
Morphing / highlighted numerals Apple Metropolitan, Contour, Flux 1 min · L P Power and animation cost
Clock of clocks forming digits once a minute ClockClock 24 1 min · L when formed P Unreadable mid-transition
Hidden until asked Hermès Le Temps Suspendu, Pulsar press — · H H No passive glance
Encoding Example Precision · glance Discreet Watch out
Sweeping or ticking element round the perimeter seconds hands; Apple Modular Ultra’s edge count-up 1 s · L P A full ring lights many pixels
Stop-to-go: a lap, then a pause Hilfiker ~1 s · L P The pause carries its own meaning
Retrograde arc Moser, 30-s arc 1 s · L P The snap-back is an attention event
Jumping digits F.P. Journe Vagabondage III 1 s · L P Constant change pulls the eye
Fill or gradient that moves with the second Apple Flux (fills the face), Palette ~5 s · L C Large lit area
Rhythm of something else (a blinking colon, a pulse, a character’s step) Mickey’s tapping foot liveness · L P Whimsy, motion cost
A gap travelling the minute track (one missing tick) † — 1 s · M C An absence pops out less than a presence
Running indicator only (proves “alive”, not which second) ISO 6425 — · L P
Seconds on demand (press, tap, wrist raise) Pulsar hold; Exactograph’s dial “expands on tap” 1 s · H H
Encoding Example Precision · glance Discreet Watch out
Window or text (abbreviated / full word) Day-Date exact · L P Language, width
Outsize date Lange 1 exact · L P Space
Pointer on a 1–31 track, or a marked numeral on a ring Oris pointer date; Nomos coloured apertures exact · M / L P One more pointer
Retrograde arc (7 steps or 31) weekday snail cam; Hermès date arc exact · L P
One glyph per weekday Seiko kanji wheel (月 火 水 木 金 土 日) exact · L once learned C
A dot on a spiral of 31 Ochs und Junior (indices double as 5-day marks) exact · L after learning C Needs the key
Month as the hour index it equals (1 = January) Ochs und Junior perpetual exact · M C
Weekday as an accented index 1–7 † — exact · L after learning C Clashes with the hours
Weekend or not by a change of the whole face Snoopy’s “Sunday Surprise” weekend · L C
Season as day length Solar Dial weeks · L C Depends on location
Encoding Example Precision · glance Discreet Watch out
Arc between full and empty Lange “Auf/Ab” power reserve (1879) continuous · L P
Days-left number Segment34’s battery-days field 1 day · L P
A drawn element that erodes with the days (a chapter ring shortening) † — days · L C Can read as decoration, or as a bug
Rhythm change when low (2-s steps) Citizen, Seiko threshold · L if known C Only while awake
Countdown that appears only near the end Hermès L’heure impatiente 1 min · L P

Glucose: the value and its zone (a quantity with a good zone)

Section titled “Glucose: the value and its zone (a quantity with a good zone)”
Encoding Example Precision · glance Discreet Watch out
Digits every CGM face exact · L P The most public form
Zones on a scale, neutral pointer FAA airspeed arcs zone + position · L P Zone limits must be the user’s
Fine scale where decisions happen (magnified low band) Blancpain X Fathoms: a 0–15 m hand with the 3–6 m band in orange high where it matters · L–M P/C Two pointers to tell apart
Memory of the extreme (the night’s lowest stays until morning) X Fathoms max-depth hand value · L C Needs a reset rule
Zone only: a colour, a shape, a word, a mood GDH’s “very big and coloured”; Nightscout Clock creatures; a colour band zone · L P→C The exact number one action away (§4)
Fuzzy words (“about 6, steady”) † after Fuzzy Text coarse · L P Words are public
A fourth hand pointing at mmol/L on the hour numerals, posing as a GMT hand † GMT hand, pointer date ~0.2 mmol/L · L C Wraps above 12; risk of reading it as time
The target range as a sector band, in or out † 1930s sector dials zone · L C
Level of a liquid (height = value, tilt = trend) † HYT, Flux coarse · L C Large lit area
Range as daylight: the value as a sun inside its “day” arc † Solar Dial zone · L C Collides with real day and night
Masked: shown on raise or tap, otherwise hidden † Hermès L’heure masquée exact · H H Hides when it matters unless the face unmasks itself on an alert
Encoding Example Precision · glance Discreet Watch out
Arrow glyph, 5 or 7 states Dexcom, Libre class · L P Each system’s arrows mean different speeds
Continuously turning arrow Sukrara: angle = clamp(rate / 3) × 90° continuous · L P Pre-attentive (orientation pops out)
Tilt of another element (an index, an anchor, the whole value) † B-Uhr anchor tilted 5–7 classes · L C The anchor stops anchoring
Shape of a tip (arrowhead, flat, chevron) † pointer-date tips 3–5 classes · M C Small
Inflection: rising-but-slowing, falling-but-turning Casio Pro Trek barometer’s four arrows classes · L P/C Needs a defined second derivative
Delta number every DIY face 1 mg/dL · L P
Signed offset from the last reading Casio’s differential pointer coarse · M C
History graph with the newest point marked Casio pressure graph; Sukrara’s 12 h chart shape · M P Small; needs scale cues
The hour track as history (each hour index’s length = that hour’s glucose) † radial sparkline profile · M C Radial comparison is slow (Blascheck)
A marker where the value will be in 15–30 min † Loop’s eventual glucose zone · L C Planned in Sukrara
Speed of a motion (a whirl for fast, a twitch for steady) Weiser’s Dangling String coarse · L, peripheral C Motion is noticed by others too
Age: text (now, 3m), struck-through or greyed value, a ring that drains between readings, an element that stops moving Sukrara freshness, xDrip+, Nightscout Clock’s stopped car steps · L P/C Must reach “old” unmistakably
Seconds that wait for the next reading (stop-to-go synced to the sensor) † Hilfiker fresh/due/late · L C Drifts from true seconds; bounded wait
The “running” element stops or dims when data is old † ISO 6425 running indicator binary/graded · L C
Encoding Example Precision · glance Discreet Watch out
Time since as text (1h 20m), amount beside it Sukrara’s last-treatment slot min · L P
A mark on the time dial at the moment it happened (the dial as a timeline) † Android Wear’s “a watch dial is a timeline” position · L C Wraps after 12 h
Quiet window as an arc or countdown that empties Sukrara’s 10m left badge; Hermès L’heure impatiente min · L C
Sensor life as a disc that drains (moon-phase-like); countdown ring in the last hours Sukrara’s sensor disc and countdown ring days/hours · L C
Warm-up as something filling Sukrara’s filling disc min · L C Never a figure that passes for glucose

8. Discretion: readable by the wearer only

Section titled “8. Discretion: readable by the wearer only”

The owner’s brief in one line. Ways to get there, with their costs (evidence):

  1. Unlabelled positions read with a private key (Ochs und Junior dots, Ziiiro swirls, a single dot): onlookers see ornament. Costs learning time.
  2. Disguise as a familiar complication (GMT hand, power reserve, sector ring, moon phase): looks like watch furniture. Risk: the wearer misreads it too, e.g. glucose as time.
  3. Hidden until asked (Le Temps Suspendu, Pulsar, Tokyoflash’s tilt-to-reveal): loses the zero-effort glance; must unmask itself when action is needed.
  4. Behaviour codes: a rhythm, a jump size, a pause; invisible to anyone who doesn’t know.
  5. Glyph systems others don’t share: kanji weekdays, planetary symbols, a personal icon set.
  6. Coarseness: zones or words instead of numbers reveal less, to the wearer too; the exact value one action away.
  7. Familiar context (reasoned, untested): a face full of ordinary data (weather, steps, heart rate) makes one more number unremarkable; a face where glucose is the only datum makes it the subject.

And what works against it:

  • Pre-attentive channels (hue jumps, flicker, motion, sudden size change) reach everyone’s peripheral vision, not only the wearer’s (Healey). Red and green zones are a public code.
  • Medical iconography: drops, syringes, the word glucose, mg/dL printed beside a number.
  • Discretion against safety: Dexcom keeps urgent low and failures un-silenceable on purpose. A face’s discretion can yield when action is needed (it is Sukrara’s alert, not the face, that decides that), and §4 rule 4 always holds. What is known about onlookers:
  • Shoulder surfing is opportunistic: curiosity (34 %), boredom (34 %), “device in line of sight”, “reaction to stimulus”; text is what gets read (47 %); people react by switching to something harmless (Eiband 2017, phones). A change, a flash or a bright patch is the stimulus that draws a stranger’s eye.
  • Obfuscation works on casual readers: handwritten-style fonts deterred them (CHI EA 2016).
  • Reading as medical isn’t automatically stigmatising: a device labelled “medical” was rated more acceptable than the same device as a “brain fitness tool” (d = 0.78, N = 1,387: Kelly & Gilbert 2018); some people with diabetes treat visibility as identity (PMC13154725). The owner’s “doesn’t scream diabetes” is a personal preference these studies neither confirm nor refute; other wearers may want the opposite, so discretion is a setting of the person, not a law of the face.
  • Fashion and social acceptability are separate: a watch must be “a traditional watch” and personal expression (Visuri 2017, Ouverson 2017).
  • No study yet of what bystanders notice on a watch: a design round can only reason about it, or test it with people.

What the face could react to, and the evidence on how much to change.

Trigger Known from Example reaction (space of options)
Zone change, distance to a threshold Sukrara Emphasis grows with distance and urgency; nothing changes while in range
Fast move, inflection Sukrara trend A quiet sign of speed; a turning point
Reading age Watch clock + reading time The value ages visibly (§3)
A treatment logged Sukrara “Handled” state with its remaining time; the time since it
An alert needed / muted / quiet Sukrara alert status The need in words; the quiet with its end
Sensor warming / ending / ended Sukrara Countdown, filling, flag
Time of day, sleep schedule Watch Night rendering; fewer data at night
Activity in progress, heart rate Watch The zone bigger, the rest smaller (sport)
Water / dive Watch (native) Gaps expected; pre-dive checklist
Phone or link lost Watch Which link, since when
Watch battery low Watch Battery more prominent
Wrist raised vs always-on Watch Detail when awake, essentials at rest
Time-zone change Watch Travel aid (insulin timing)

How much to change:

  • Calm by default: Weiser and Brown’s calm technology moves information between the periphery and the centre, with the person in charge (Weiser & Brown 1995). The face stays in the periphery while nothing needs the wearer.
  • Platforms disagree: Android Wear asked faces to “react to [data] by changing styles and colors”; Apple’s always-on rules ask for “a consistent layout” and “infrequent, subtle updates” (§6). A reaction that keeps the layout (emphasis, contrast, size within its place) reconciles them.
  • Gate the change on the data (Samsung’s blink only out of 60–100 bpm): reactions cost attention and power, so they appear only when the condition holds.
  • Dark cockpit: Airbus panels light up only “to indicate where an action is required”: “if all the lights are out, then the aircraft is ready to fly” (Airbus). But dark isn’t proof of OK, so pilots cross-check; a face that goes quiet in range still needs a sign that it is alive and current (Mankoff’s “was broken”).
  • Escalation is a ladder with few, redundant steps: medical alarms (IEC 60601-1-8) have three priorities, each coded at once by colour (cyan or yellow steady, yellow 0.4–0.8 Hz, red 1.4–2.8 Hz), sound and glyph count (!, !!, !!!); information must look different from alarms (IEC text). Shearwater dive computers make the level readable without colour: warnings solid inverted, critical flashing between inverted and normal (Perdix manual).
  • Persistence before reacting: the Descent’s “ascending too fast” fires after > 9.1 m/min for more than 5 s; Sukrara’s fast alerts need the rate held 5 min. A face’s reaction to a fast move can follow the same rule.
  • Show where it’s heading, not only where it is: an airliner’s speed trend vector is an arrow to the speed 10 s ahead, drawn only above a dead band, and on Gulfstreams the digits turn amber when the vector reaches the low-speed band: a predictive alert (FAA report). Up means more, so rising energy moves upward.
  • Bad data must be unusable, not just marked: lost or wrong primary data is removed or crossed with an X, and the detector is independent of the source (AC 25-11B §4.6). Stale glucose is the watch’s failed sensor (§4 rule 1).
  • A reaction explains itself: what happened, why the face looks different, how it ends (CLAUDE.md Make it juicy: “motion that explains what happened”).
  • Context is weak evidence: heart rate can’t tell a low (1.4); “exercising” doesn’t say which way glucose goes; a reaction inferred from context should be mild and reversible.

Facts that a design can use or must know; limits here are not design rules (How to use). Details and sources: evidence: Garmin, evidence: platforms.

The owner’s watch: Garmin Descent Mk3i 43 mm

Section titled “The owner’s watch: Garmin Descent Mk3i 43 mm”

There is no 42 mm Descent Mk3: the owner’s “42 mm” is the Mk3i 43 mm (descentmk343mm, Connect IQ API 5.1).

Display 1.2″ (30.4 mm) round AMOLED, 390 × 390 px, 325 ppi, ≈ 12.8 px/mm (1 mm ≈ 13 px), 16-bit colour; sapphire
Case 43 × 43 × 14.15 mm, titanium bezel (carbon-grey DLC or bronze PVD) with four exposed screws at 1:30, 4:30, 7:30, 10:30
Bezel A printed 24-hour ring outside the pixels: 24 at the top, 06 at 3, 18 at 9, “GARMIN” at 6, red “NO-FLY TIME” and cyan “ALT TZ” near 2. Always visible, even with the screen black; every edge pixel reads against it. Garmin’s own face points at it for no-fly time and a second time zone
Buttons 5: LIGHT (top left), UP·MENU (mid left), DOWN (bottom left), START·STOP (top right), BACK·LAP (bottom right); all belong to the system on a face
Touch Yes; a face gets touch-and-hold with coordinates (e.g. hold a slot to open its app); no tap on this model
Modes Wrist raise wakes the face for ~10 s (every second, animation possible); otherwise once a minute. Always-on optional: ≤ 10 % luminance or the face goes black, no seconds; battery 10 days → 4 days
Night Red Shift (red/green/orange tint for night vision, can turn on at bedtime or sunset; a face can’t detect it, the system re-tints), Lume (native faces only: turquoise glow from the ambient-light sensor), Night Vision mode, Large Fonts (a face can detect it)
Data a face can read Battery % and days left, steps, floors, intensity minutes, calories, heart rate (and 6 h history), body battery, stress, respiration, SpO2, recovery time, altitude, pressure, temperature, compass, weather and forecast, sunrise/sunset, notifications, alarms, do-not-disturb, phone connected, user’s sleep and wake times, 12/24 h, first day of week; 42 system complications
Not readable by a face Dive data (surface interval, no-fly, tissue load), ambient light, Red Shift state, HRV, training readiness, sleep state
Other apps’ data Complications published by other apps (up to 4 per app, with value, labels, unit, icon and ranges), also usable in Garmin’s own photo face (Face It)
Attention A face can’t vibrate or sound by itself; alerts come from Sukrara (phone notifications mirrored by Garmin, or a Sukrara watch app)
Fonts System vector fonts at any size (Bionic numbers, Roboto Condensed, Swiss 721), text along a circle, custom fonts; built-in heights 19–80 px
Settings On the watch (hold MENU › Watch Face › Customize: menus) and in the Connect IQ phone app
Mockups SDK simulator image descentmk343mm.png 638 × 830 px with the display at x 131, y 211, 390 × 390 (copy); installed SDK: ~/Library/Application Support/Garmin/ConnectIQ/Devices/descentmk343mm/; product photos: Garmin 852217
Today’s data path The watch polls every 5 min (docs/08); push from the phone, a companion app or BLE are faster paths. A limit to solve, not to design for
Platform Lit budget Update Notes
Garmin AMOLED (API ≥ 5) ≤ 10 % of total luminance 1/min Above it the face goes black; move elements ≤ 4 px a minute
Wear OS ≤ 15 % average luminance (Play review) ~1/min ≥ 85 % black; outlines not fills; time always shown
Samsung ≤ 15 % on-pixel ratio — Red counts 33 %; no dimmed copy of the active face
Fitbit ≤ 20 % 1/min No numeric stats in AOD (stale data)
Zepp ≤ 10 % of pixels — No second hand
Apple No quota; the system dims by lit ratio Down to 1/min (1/s on Series 10) Outlines and dim interiors; 1 nit night mode

Before showing a design to the owner, check it against:

  • A 2-second glance answers: the time; the value or its zone; the direction; whether it is current.
  • Every state of §3 is drawn, including stale, no readings, LO/HI, no trend, warm-up, ended, quiet, muted, no sensor; and mg/dL and mmol/L at their widest.
  • The §4 rules hold, in colour, in grey (Red Shift / tinted) and in always-on.
  • The discretion level is deliberate: what a stranger at 1 m sees in each state; nothing medical-looking unless chosen (§8).
  • What draws the eye pre-attentively is only what should: nothing moves, flashes or turns saturated while in range.
  • Always-on is designed (not dimmed), within budget, layout kept, and still answers the 2-second question: glucose is in it, not only the time.
  • Stale and no-data marks never cover the time, and say since when (and why, when known).
  • Night: brightness, red rendering, nothing that lights a bedroom or a cinema.
  • Secondary data stays legible for older eyes (no small dark-grey text).
  • Sizes in mm at wrist distance meet §5; no thin weights; distinguishable digits.
  • Each reaction (§9) is gated on data, explains itself and has a visible end.
  • User-chosen slots handle no data, widest value, unit and errors; settings are reachable on the watch.
  • Words exist in English and Spanish (weekdays, months, states) and fit.
  • Motion has a meaning, settles, and the face works without it.
  • Mocked up on the real frame: the Descent’s 24 h bezel ring around the pixels, at 1:1 physical size.

What design rounds teach that applies to any face: the reason or the usability fix, not the design. Add an entry per lesson: date, the face (by name only), what was learned, why, and the evidence (the owner’s words, a test).

  • 2026-11-16 · Sukrara Aura · a line behind a number reads as a strike-through. In Sukrara a struck value means “old”, so any background (lines, rings, lasers) must stop short of every text. A cut-out that follows the glyphs keeps the background whole; boxes look like stickers. Evidence: rounds 2–3, the owner chose the outline.
  • 2026-11-16 · Sukrara Aura · coverage is the signal only if it fades. A background that fills the screen evenly says nothing; fading with distance from the level makes “how far it spreads” readable, but too steep a fade hides it. Evidence: the owner: “now they fill the screen with color almost uniformly”, then “now the isobars fade too fast, just an inbetween”.
  • 2026-11-16 · Sukrara Aura · motion belongs at the rim. A seconds sweep across the dial competed with the value; confined to the outer quarter, faint and optional, it was kept. Lines with visible ends read as objects; drawn edge to edge they read as light (the owner: “infinitely long so the ends are no visible”).
  • 2026-11-16 · Sukrara Aura · empty rows give their space back. When a row (pill, reading time) is absent, the rest re-centres and grows; fixed slots left the face looking broken. Evidence: round 3, the owner asked for it.
  • 2026-11-16 · Sukrara Aura · short words keep the numbers big. Every word on the face competes with the time and the value for width; “Take insulin” became “Insulin”, “Sensor ended” became the flag + “Ended”, and the battery’s days and percent go on one line where a stack would touch the value. Evidence: the owner, in round 2 and again for the spec (“shorten the texts so we had more space”).
  • Bystander perception: no study of what others notice on a watch face (number, colour, graph). Testable with people in a design round.
  • Health data in always-on: Apple’s docs both hide and show health data; nobody addresses medical data the wearer acts on.
  • Stale threshold: products disagree (10.5–20 min); Sukrara’s steps (5 / 20 min) are stricter than most.
  • Should the wrist alert? Split ecosystem (1.6); the DVLA treats dependence on alarms as a reason to stop driving.
  • Trend vocabulary: continuous angle (Sukrara) vs 5/7 classes vs inflection (Casio); no study of reading a trend arrow at a glance.
  • Rings vs bars: concentric radial bars were ~6× slower to compare (2019), yet donuts beat bars (2019) and bars beat rings by ~80 ms on proportions (2023), where rings won on beauty; reading one live value or zone is untested.
  • Text vs graphics for one value: text read slowest on proportions, yet it is what faces use most, and aviation pairs every readout with a scale because a number hides the margin.
  • Polarity and weight at night: legibility research favours dark-on-light for small text in the dark; dark adaptation, always-on budgets, burn-in and battery all push to sparse light-on-dark. Lighter weights may read better (Dobres) but lose contrast first in sun and at an angle. Untested on a 1.2″ emissive screen.
  • Red/green vs colour blindness: aviation, the Descent and Shearwater all use green/yellow/red, which ~8 % of men can’t rely on and Red Shift erases; redundant coding (inversion, flashing, shape) is the documented answer.
  • Subtle change vs change blindness: ambient theory asks for subtle changes, but a change between glances is invisible unless the face encodes it. With the wrist down the screen is out of sight, so peripheral-motion research mostly applies to the moment the wrist comes up.
  • Flash rules disagree: 1.4–2.8 Hz (IEC high priority), 0.8–4 Hz cancellable (FAA), ≤ 3/s and an area limit (WCAG) that a full-face flash at wrist distance exceeds.
  • Garmin: the always-on rule version on this firmware, the exact awake window on the Descent, whether Red Shift re-tints Connect IQ faces, whether a face’s notification vibrates.
  • Gaps in this research: Reddit was blocked to every tool used (web fetch, and Claude in Chrome refuses the site itself, even in a tab the owner opened), so people’s voices come from Garmin store reviews, Play reviews, Diabetes UK and other forums, GitHub issues and qualitative studies; the App Store’s reviews weren’t readable; Hodinkee couldn’t be fetched; a few numbers come from secondary summaries (marked in the evidence files).