Field Sampler — interface walkthrough

1Measuring instruments0.3 s

Before measuring the app, the walkthrough's own pitch estimator (analyze.mjs: FFT peaks of the first 8 harmonics; no code shared with the app's pitch detection) is checked on synthetic tones of known pitch: plucks and decaying harmonic tones, 130–880 Hz, with noise, each searched from a hint 100¢ off.

• ✓Estimator error far below the ±2¢ the key checks allow
  worst 0.0007¢ over 10 tones: pluck 223.78 Hz +0.000¢, pluck 220.69 Hz +0.000¢, pluck 130.97 Hz +0.000¢, pluck 261.58 Hz -0.000¢, pluck 438.36 Hz -0.000¢, harmonic tone 130.81 Hz +0.001¢, harmonic tone 220 Hz +0.000¢, harmonic tone 261.63 Hz -0.000¢, harmonic tone 440 Hz -0.000¢, harmonic tone 880 Hz -0.000¢

2First launch0.9 s

With nothing to restore (no take in the address), the app opens on a blank session with the takes menu open (the demo take under Examples), a prompt in the waveform, and every control in place.

• ✓Cold start opens the takes menu
• ✓The demo take is listed under Examples
  Demo take 0:08 · thuds, knocks and plucked notes over wind
• ✓No Demo button in the header
• ✓Empty-state prompt shown
• ✓Fits the screen without scrolling
  overflow 0 px wide, 0 px tall
• ✓Pads and keys rendered
  12 pad slots, 13 keys

3Demo take0.7 s

The demo take (takes menu → Examples) loads a built-in synthetic recording (knocks, thuds, plucked notes) so the app is playable without a microphone.

• ✓Demo take sliced into hits
  12 hits

4Record from the microphone3.2 s

New + starts a new take and records at once (with the demo take open, Rec would record onto its end: see the next step); the waveform scrolls in live. The Rec button turns into a red outline with a stop square, the time and the time left, at the same width, and nothing else on screen moves; Import and New + wait. Rec stops, analyses and saves the take. (The microphone here is a fake device playing the demo file.)

• ✓Recording moves and resizes nothing (Rec keeps its width)
  Rec 156 px wide, "0:02 4:58 LEFT"
• ✓Recording shows as a red outline, not a filled block
  background rgba(0, 0, 0, 0), border rgb(255, 59, 48), text rgb(255, 59, 48)
• ✓Import waits while recording
• ✓Rec is back to its width after stopping
  156 → 156 px
• ✓Recording sliced into hits
  5 hits in 4:58
• ✓Saved as a take with its own link
  http://localhost:8788/t/uZlRLI449uJ9

5Record into the take10.9 s

Rec records onto the end of the open take, like tape; New + starts a new take and records at once (Import, like New +, makes a new take). The recording appears after the take at the scale the take was shown at, and once it reaches the right edge the view follows it; nothing else on screen moves. Each pass gets a band along the top with its time, and Rec shows the time the take has left (5:00 in all). Each pass is a file of its own, analysed alone, so every earlier hit stays exactly as it was: its start, end and id, its edits. Here a take is made with New +, two passes are recorded onto it with Rec (the fake microphone plays the walkthrough file), the take is reopened, and each pass stored in its own time.

• ✓New + is a flat outlined button beside Rec, without the red dot
• ✓New + starts recording at once
• ✓Import and New + wait while recording
• ✓New + made a take of one pass
  3.06 s
• ✓Rec shows the time the take has left
  REC 4:57 LEFT (296.94 s)
• ✓Recording onto the take moves and resizes nothing
  Rec 156 px wide
• ✓The take stays on screen while recording onto it, at the same scale
  span 3.062 → 3.062, 3.062 s
• ✓Past the right edge the view follows the recording (newest at the right)
  start 0.00 → 1.16 → 2.46 s
• ✓The pass being recorded has its band
  ["Pass 1 · 10:08","Pass 2 · Recording…"]
• ✓Import and New + wait while recording onto the take
• ✓The recording was added to the same take, as its second pass
  L1T32bnV8P72: 1 pass(es) after the first, 2.83 s, 17 candidates
• ✓Every pass-1 slice is bit-identical before and after the append
  5 slices before, 5 of pass 1 after; 5 new
• ✓The take’s own analysis is stored as it was
• ✓The new pass’s hits are sliced, their ids after the take’s
  ids 20, 24, 26, 30, 32
• ✓No slice runs past the seam
• ✓Each pass gets a band with its time ("Pass 2 · 16:10")
  ["Pass 1 · 10:08","Pass 2 · 10:08"]
• ✓On stop the view glides back to the whole take (it was at Fit)
  0.000 + 5.892 s
• ✓Rec’s time left counts the take’s passes
  294.11 s
• ✓A second recording onto the take: the passes before it are bit-identical
  2 passes after the first; bounds 0.000, 3.062, 5.892
• ✓Reopened, the take has every pass and the same slices
  13 slices
• ✓Each pass is stored in its own time, from its start
  17 hits in 2.83 s; 19 hits in 2.06 s

6Import a file1.2 s

Import opens a file (here the walkthrough take: footsteps, knocks, a plucked A3 29 cents sharp, a quiet knock, a wind gust, glass, bird chirps, dropped keys, a bell and a faint knock). Analysis runs off the main thread and the take is saved.

• ✓Analysis under 1 s for 10 s of audio
  sliced in 80 ms
• ✓Every broadband hit found (±40 ms)
  12/12
• iBird chirp at 7.6 s (narrowband, a known weak spot)
  not sliced at the default sensitivity: candidate strength 0.14 vs threshold 0.25; raise sensitivity to get it
• ✓No false hit in the wind gust
  4.5–5.4 s clean

Ground truth: footstep 0.4s · footstep 1s · footstep 1.6s · knock 2.3s · knock 2.55s · pluck 3.2s · quiet knock 5.8s · glass 6.4s · bird chirp 7.6s · keys 8.3s · footstep 9.3s · bell 10.3s · faint knock 12.1s

7Keys arm the most tonal slice0.6 s

After analysis the keys arm the slice that makes the best instrument (pitch confidence × ring length), not slice 1. Its detected root labels the keys.

• ✓Plucked A3 is armed
  KEYS PLAY Slice 6/12 A3 +29¢
• ✓Root label within 2¢ of the true pitch (A3 +29.5¢)
  label 223.72 Hz vs true 223.78 Hz (-0.5¢)
• ✓A clear pitch is not marked approximate (no ≈)
  KEYS PLAY Slice 6/12 A3 +29¢

8Pads play trimmed slices2.5 s

A pad plays its slice from just before the attack to where the sound dies into the background, not to the next hit. Faint regions on the waveform are what gets trimmed.

• ✓Pad 1 plays its own slice (the first footstep)
  sounds like slice 1 (r 0.992; next best slice 4, r 0.163)
• ✓Plays the hit, then stops well before the next one
  played 0.34 s; next hit is 0.60 s after this one
• ✓No DC offset in playback (removed at import)
  mean 0.0005 (the file has +0.010 offset and 6 Hz rumble)

9Normalization5.1 s

With Norm on (default), every slice plays at the same peak level, so a distant knock is as playable as a close one, up to a +18 dB lift: a near-silent slice is not turned into loud hiss. Norm off plays slices at their recorded level.

• ✓The faint knock pad plays the faint knock
  sounds like slice 12 (r 0.931; next best slice 4, r 0.422)
• ✓Norm on: both peak near -1 dBFS
  knock -1.0 dBFS, quiet knock -1.0 dBFS
• ✓Norm off: the quiet knock is ~10 dB softer
  knock -4.7 dBFS, quiet knock -15.4 dBFS (10.6 dB apart)
• ✓Norm lifts the faint knock by 18 dB and no more (peaks at −1 − shortfall)
  Norm off -30.3 dBFS, Norm on -12.3 dBFS: lifted 18.0 dB; −1 dBFS would need 29.3, so it should peak at -12.3

10Waveform: tap, double-tap, arrows3.9 s

Tapping a slice selects and auditions it (Delete acts on the selection). Double-tapping arms it on the keys. The ‹ › arrows step the armed slice. A glass and a bell have no clean harmonic series, so their roots are best guesses, marked ≈ with no cents: the glass by its lowest partial, the bell by its hum (the right note, an octave below the strike).

• ✓Tap auditions the slice
  peak -1.0 dBFS
• ✓The route control (Both / Pads / Keys / Hidden) appears under the selected marker
  under 2.55 s; the knock is at 2.55 s
• ✓Double-tap arms the slice
  KEYS PLAY Slice 8/12 ≈ A5
• ✓Glass root: ≈ and its lowest partial (A5)
  KEYS PLAY Slice 8/12 ≈ A5
• ✓Bell root: ≈ and the prime's note name (A, any octave)
  KEYS PLAY Slice 11/12 ≈ A3 (a period detector alone finds 44 Hz, the common divisor)
• ✓‹ steps back to the pluck
  KEYS PLAY Slice 6/12 A3 +29¢

11Keys: pitch matches the label8.0 s

Each key plays the armed slice at that key's pitch, tuned against the detected root, cents included: the pluck is 29 cents sharp, and its A3 key plays A3, not A3 +29¢. Touching low on a key plays it dry (no smear). Measured: the pitch of what comes out of the speakers, by the walkthrough's own estimator (step 1), to ±2¢.

• ✓Root key (MIDI 57) sounds its label (±2¢)
  heard 220.00 Hz, want 220.00 Hz (-0.01¢; 8 partials)
• ✓Lowest key (MIDI 48) sounds its label (±2¢)
  heard 130.81 Hz, want 130.81 Hz (-0.02¢; 8 partials)
• ✓Highest key (MIDI 60) sounds its label (±2¢)
  heard 261.63 Hz, want 261.63 Hz (-0.00¢; 8 partials)

12Keys: a held note sustains9.5 s

Held past the end of its slice, a pitched note that rings plays its natural decay, then hovers where it has fallen 12 dB: each grain comes from a random in-tune spot there, so there is no fixed loop to hear repeating. Release fades it out. Fast-decaying and unpitched sounds have nothing to sustain: held, they play once and end (see MIDI).

• ✓Still sounding at 3–4 s (slice is ~2 s)
  level -10.2 dB below the peak
• ✓Sustain is steady (no re-strike or pulsing)
  level varies 0.0 dB over 3–3.9 s
• ✓The sustain stays in tune (±2¢)
  heard 220.00 Hz at 3–3.9 s, key is 220.00 Hz (+0.00¢)
• ✓The sustain is a clean tone (≥ 97 % of its power on its harmonics)
  99.6 % on the harmonics of 220.00 Hz (±8 Hz) over 3–3.9 s
• ✓No fixed loop (the held sound never replays itself)
  best 20–500 ms self-match (at whole periods of the 220.00 Hz heard) stands 0.02 above the typical one (the old fixed loop scored 0.41)
• ✓Release fades out within 0.4 s
  silent 0.08 s after release (released at 4.19 s on the audio clock)

13Smear7.1 s

Finger height on a key is the smear: low is dry, high freezes the sound into a grain cloud (live PaulStretch). Smeared notes release more slowly.

• ✓Smeared release is longer (≥ 0.4 s)
  sounds 0.87 s after release (at 2.69 s on the audio clock)
• iCloud level variation 1–2.4 s
  7.1 dB

14Computer keyboard3.6 s

With a hardware keyboard: number row plays pads, A…K plays the keys, ← → step the armed slice, Z X shift octave.

• ✓1 plays pad 1 (slice 1, the first footstep)
  sounds like slice 1 (r 0.992; next best slice 4, r 0.163)
• ✓→ arms the next slice
  KEYS PLAY Slice 6/12 A3 +29¢ → KEYS PLAY Slice 7/12 Unpitched → C3
• ✓X shifts the keys up an octave
  C3 → C4
• ✓A plays the lowest key (MIDI 48) (±2¢)
  heard 130.81 Hz, want 130.81 Hz (-0.01¢; 8 partials)

15MIDI15.3 s

MIDI channel 1 plays pads from C1 (note 36); channels 2–16 play slices 1–15 pitched; velocity sets level; mod wheel sets smear. (Driven through the hook a native iPad wrapper uses: window.fieldSamplerMidi.)

• ✓Ch 1 note 36 plays pad 1 (slice 1)
  sounds like slice 1 (r 0.992; next best slice 4, r 0.163)
• ✓Velocity 64 is ~12 dB softer than 127 (level = velocity²)
  11.9 dB
• ✓Two pads at once stay at or under 0 dBFS (master limiter)
  peak -0.50 dBFS; one hit alone -1.01 dBFS
• ✓Pitched channel plays A4 (±2¢)
  heard 440.00 Hz, want 440.00 Hz (-0.00¢; 8 partials)
• ✓A three-note chord stays at or under 0 dBFS (master limiter)
  peak -2.88 dBFS
• ✓A held knock plays once and ends (no sustain)
  silent 0.20 s after note on, key still held
• ✓Mod wheel smears (long release)
  sounds 0.95 s after note off
• ✓MIDI pad (ch 1, note 36) flashes pad 1 and selects it
  flashing ["Pad 1"], selected Pad 1
• ✓A MIDI note on the armed slice's channel (ch 7) holds its key down until note-off
  held [52], re-armed [52], released []
• ✓The keys follow arming: another slice's held note isn't shown on them
  next slice armed: []; ch 2: []

16Which build this is0.2 s

A tester can be on the production site or the develop one, in a browser or in either TestFlight group's app. The takes menu ends with one line naming the build (in the app: its version and build number, which build it is, the site, the web commit), and every page that isn't production carries a small marker in the header (here LOCAL; DEV on the develop site; nothing on production).

• ✓The header marks a page that isn't production
  ["LOCAL"]
• ✓The takes menu ends with the build: site and web commit
  local · web ef31303

17Delete and undo6.5 s

Delete a transient by flicking its marker down (or with the Delete key). Its audio is dropped; Undo brings it back.

• ✓A flick down removes the hit
  12 → 11
• ✓Delete key removes the selected hit
  11 → 10
• ✓Deletions saved to the take: those two hits
  server has 2 deleted (1.00 s, 2.55 s; deleted on screen: 2.55 s, 1.00 s); the stored take gives 10 hits, the screen 10
• ✓Undo twice restores both
  12 hits
• ✓The undo is saved too
  server has 0 deleted; the stored take gives 12 hits (screen 12)

18Add a transient6.1 s

Long-press on the waveform adds a transient at the finger. It gets its own natural end like any other hit, and its own pad. Added here in the middle of the ringing pluck, its pad starts mid-note: a 2 ms fade-in keeps that from clicking.

• ✓Long-press adds a hit
  12 → 13
• ✓The added hit is saved to the take, where the finger was
  server has 1 added (3.900 s); the stored take gives 13 hits
• ✓Starts without a click (the first 0.1 ms is faded in)
  first 0.1 ms peaks at 0.03 × the RMS 2–7 ms in (-43.3 vs -13.1 dBFS)
• ✓Undo removes it
  12 hits
• ✓The undo is saved: nothing added on the server
  server has 0 added; the stored take gives 12 hits

19Sensitivity7.0 s

The slider beside the waveform filters precomputed candidates: instant, no re-analysis. Short ticks along the bottom show what the next stretch of the slider would add. The setting is saved with the take.

• ✓Higher sensitivity is saved, and the stored take has the hits on screen
  server sensitivity 0.950: 146 hits; screen 146
• ✓Lower sensitivity drops the weakest hits
  146 at 0.95 / 12 at 0.5 / 11 at 0.29 (weakest hit at the default has strength 0.44)
• ✓Lower sensitivity is saved, and the stored take has the hits on screen
  server sensitivity 0.294: 11 hits; screen 11
• ✓Back at the middle restores the count, on screen and on the server
  12 on screen, 12 stored at sensitivity 0.500 (was 12)

20Name, history and link2.8 s

Tap the name to rename. The ▾ menu lists this browser's takes and copies the take's link. Edits autosave.

• ✓Rename saved to the server
  server name: Walkthrough take
• ✓History lists all three takes
  Walkthrough take · take-1016-1008 · demo · Demo take
• ✓Copy link copies the take URL
  http://localhost:8788/t/bvaO82dNkFI6

21Shared link: view only, with a preview card22.8 s

Someone opening the link (another browser) gets the same take and can play it, but not change it. Messages and Slack show a preview card with the name and waveform, in the take’s colors as the app shows them with Color on: each slice’s color on its pad, waveform and marker, the keys tinted, hidden slices muted. The card is drawn again when its colors change (the color map, a hue of its own, routing), after the edits settle; this device’s Color switch doesn’t change it. Read back as a PNG from the stack.

• ✓Opens as view only
• ✓Same name
• ✓Link preview has title and image
  Walkthrough take · Field Sampler · 12 hits · 0:13. Open it to play on pads and keys.
• ✓Preview image served
  200 image/png
• ✓The card’s pads are colored where the app’s are
  app 12 of 12 pads colored; card: 63 91 109 149 120 103 56 61 120 64 103 41 (chroma)
• ✓On the card a warm slice (the footstep) and a cool one (the pluck), on their pads and waveform
  timbre: footstep pad rgb(164,118,101), waveform rgb(167,123,107); pluck pad rgb(103,206,180), waveform rgb(130,246,216)
• ✓The color map changed: the card is drawn again in it, its keys by pitch class
  8 pads changed color; 12 key colors (were 2)
• ✓With Color off on this device the card is still drawn in the take’s colors
  app pads colored: 0; card 12 pads colored, 12 as on the Pitch card (Undo from Dusk)
• ✓Undo back to the timbre map: the card in its colors again
  12 of 12 pads as on the first card; undone: footstep pad rgb(164,118,101), waveform rgb(167,123,107); pluck pad rgb(103,206,180), waveform rgb(130,246,216)
• ✓A hue of its own (blue) on the footstep: its pad and waveform on the card turn cool
  footstep pad rgb(87,160,237), waveform rgb(99,163,233) (was rgb(164,118,101))
• ✓A hidden slice: no pad on the card (later pads move up) and its waveform muted
  hidden slice's waveform rgb(82,78,75); pad 7 rgb(154,173,112), the next slice's was rgb(154,173,112)
• ✓Undone: no settings stored

22Reopen a take from history1.4 s

Choosing a take in the menu opens it (its URL becomes the address), with its own slicing and edits.

• ✓Switched to the recorded take
  http://localhost:8788/t/uZlRLI449uJ9

23Switching takes keeps each take whole10.0 s

Opening another take, or starting a new one, never writes one take's edits into another, never loses an edit made just before the switch, and never lets a slow upload take over the address of the take on screen.

• ✓Opening a take leaves its stored name alone
  the reopened take is named "take-1016-1008" on the server (expected "take-1016-1008")
• ✓Edits stay with their own take
  opened take on the server: "Walkthrough take", sensitivity 0.5, 0 deleted (was "Walkthrough take", 0.5, 0)
• ✓An edit made just before switching is saved
  the take left behind is named "Recorded, renamed" on the server
• ✓Import, Rec and the demo take wait while a take opens
  disabled: Import true, Rec true, Demo take (menu) true
• ✓A slow upload does not take over the open take
  address /t/bvaO82dNkFI6, header "Walkthrough take" (expected /t/bvaO82dNkFI6, "Walkthrough take")
• ✓The opened take is untouched by it
  server name "Walkthrough take"
• ✓The new take is still saved and listed
  1 → 2 demo takes in the menu

24Zoom and scroll the timeline2.8 s

Zoom stretches and squeezes time only: the ←|→ and →|← buttons, a sideways pinch, or Ctrl/⌘ + scroll; the waveform keeps its height. Zoomed in, a sideways drag on the waveform (away from a marker) scrolls it, the overview strip above shows the whole take with the view boxed, and a tap there moves the view. Tap still selects and plays the slice under the finger, through the zoomed mapping. Fit (or opening another take) shows the whole take again. A take opened by its link opens without the takes menu.

• ✓A take link opens without the takes menu
• ✓Opens showing the whole take, no overview
  view 0.00+13.00 s of 13.00 s, overview hidden
• ✓Stretch (←|→) twice shows a quarter of the take; the height is unchanged
  span 13.00 → 3.25 s, plot 149.59375 → 149.59375 px tall
• ✓Zoomed: the overview shows
  visible
• ✓Squeeze (→|←) doubles the span
  3.25 → 6.50 s
• ✓A sideways pinch stretches time by the fingers’ spread, around the point between them
  span 6.50 → 2.17 s (×0.333), time under the fingers 3.45 → 3.45 s
• ✓The pinch leaves the height alone
  plot 149.59375 → 149.59375 px, canvas 150 → 150 px
• ✓A sideways drag on empty waveform scrolls it
  start 2.37 → 2.73 s (expected 2.77)
• ✓Tap selects the slice under the finger when zoomed
  tapped the marker at 3.20 s; selected: 3.20 s
• ✓A tap on the overview moves the view there
  start 2.73 → 0.22 s (expected 0.22)
• ✓Ctrl + scroll squeezes time around the pointer
  span 2.17 → 2.92 s; time under the pointer 0.76 → 0.76 s
• ✓Opening another take shows the whole of it
  view 0.00+2.72 s of 2.72 s
• ✓Fit shows the whole take again
  span 2.72 s, overview hidden

25Move a transient13.1 s

Drag a marker sideways to move it: it snaps to the nearest attack (found in Rust, from the onset detection function), a dashed ghost stays where it was, and a readout shows the time and how far it has come. On release it moves, as one undo step, with a toast to take it back; the slice then plays from there. Here the quiet knock’s marker (5.8 s) is dragged onto the glass (6.4 s, whose own marker was flicked away first): the slice plays the glass. The move is saved with the take (a moved row: the hit keeps its id, so its settings and color go with it), survives a reload, Undo takes it back, and dragging it back onto its own attack removes the row. Zoomed in, a drag on empty waveform still scrolls.

• ✓Before: the slice plays the quiet knock
  sounds like slice 7 (r 0.997; next best slice 5, r 0.480)
• ✓While dragging: the marker snaps to the attack, within 5 ms of the true onset
  readout “6.400 s (+600 ms)”: at 6.4000 s (glass onset 6.4000 s, +0.0 ms), from 5.8000 s
• ✓On release: a toast says how far it moved, with Undo
  “Moved slice 7 by +600 ms” (expected +600 ms)
• ✓Saved: a moved row for the quiet knock’s hit, at the glass’s detected start exactly
  stored moved [[68,6.400022675736961,6.40421768707483]]; the glass was detected at 6.400022675736961 s
• ✓After: the moved slice plays the glass
  sounds like slice 8 (r 0.997; next best slice 11, r 0.273)
• ✓Undo takes the move back, saved
  stored moved []
• ✓After Undo: the slice plays the quiet knock again
  sounds like slice 7 (r 0.997; next best slice 5, r 0.480)
• ✓A reload keeps the move: on the server, on screen and in what plays
  stored moved [[68,6.400022675736961,6.40421768707483]]; the slice selected by tapping there starts at 6.40 s; 11 hits on screen, 11 stored
• ✓After a reload: the moved slice still plays the glass
  sounds like slice 8 (r 1.000; next best slice 11, r 0.290)
• ✓Dragged back onto its own attack: the moved row is removed
  stored moved []
• ✓Zoomed: a drag on empty waveform still scrolls, and moves no marker
  view start 4.35 → 4.89 s; stored moved []

26Edits survive reopening3.7 s

A deletion is part of the saved take: opened again later (here: the page reloaded), the take comes back without the deleted hit, on screen and on the server.

• ✓A flick down removes the hit
  5 → 4
• ✓After reopening, the deleted hit is still gone (screen and server agree)
  4 hits on screen after the reload (5 before the deletion); the stored take gives 4, without the hit at 1.010 s

27A take opens the same every time9.3 s

A take's analysis is stored with it, once: its candidates, their attack peaks, and one natural end per candidate. Opened again, it slices from what is stored and measures only the transients added by hand. A take stored before that (legacy) still opens: its peaks are measured on every open, as before, and never written back. Slices are read from instrument.set, through a stand-in for the iPad app's bridge; takes with a known, deliberately shifted analysis are made through the API to tell stored values from measured ones.

• ✓The imported take was stored with its analysis: the version and one end per candidate
  analysis {"dsp":1}; 146 ends for 146 candidates
• ✓Opened twice, every slice has the same start, peak and end, to the bit
  12 slices, both times
• ✓Each found slice ends where its stored end says (capped at the next onset), at its stored peak
  12 found slices checked against the stored document
• ✓A take with its analysis stored slices from it: the stored peaks and ends, not measured ones
  12 slices at the shifted peaks and ends, the same on a second open
• ✓A legacy take (no stored analysis) opens, with its peaks measured again
  12 slices; every found peak back within 0.026 ms of the audio's own (stored 4.2 ms late)
• ✓Nothing measured is written back to a legacy take
  146 candidates as stored; ends null; analysis null

28Color by sound17.8 s

Every slice is colored by how it sounds (docs/features.md, stage B): five timbre traits per hit, measured in Rust once when the take is made and stored with it, drawn through the take's color map (Timbre, Pitch, or a palette). The map is saved with the take, undoable, and the same for everyone who opens it; Color, beside Norm, turns this device back to monochrome. Colors are read from the pads' styles and the waveform's pixels, the map from the stored take.

• ✓Timbre traits are stored with the take: 5 bytes per candidate
  730 bytes for 146 candidates
• ✓The pads are colored, in several colors
  12 of 12 pads colored, 12 distinct colors
• ✓A low thud (a footstep) is warm and the bright pluck cool, on the timbre map
  footstep brightness 23/255, hue 41°; pluck brightness 183/255, hue 174.6°
• ✓The waveform draws slices in their colors: the thud warm, the pluck cool
  thud rgb(216,165,146), pluck rgb(95,213,184)
• ✓The pitch map is saved with the take
• ✓On the pitch map the pluck takes its note’s hue (A: 115°)
  hue 115°
• ✓Each key takes its own pitch class’s hue
  48:25° 49:235° 50:85° 51:295° 52:145° 53:355° 54:205° 55:55° 56:265° 57:115° 58:325° 59:175° 60:25°
• ✓Another browser opening the link sees the same map and colors
  map Pitch; pluck hue 115°
• ✓Undo goes back to the timbre map, saved as such
• ✓…with the colors it had
  pluck hue 174.6°
• ✓A palette is saved with the take
• ✓Color off: pads and waveform are monochrome again
  0 pads colored; pluck pixel rgb(158,155,150)
• ✓Color off is kept on this device after a reload; the take keeps its map
  0 pads colored
• ✓A take stored before colors opens colored (measured live) as if stored, and nothing is written back
  12 pads colored, the same as the stored take's; stored timbre null
• ✓Its owner opening it: the card is drawn in its colors (once: not again on the next open)
  12 card pads colored; not uploaded again

29Slice settings: hold a pad16.0 s

Holding a pad for ½ s plays it at once and then opens its settings beside it; a shorter tap is an ordinary hit, and a finger that moves first opens nothing (pads can still be swept and drummed). The panel: the name as its title, Color (Auto or one hue), Tune (the root stepper and cents), Gain and Reset. Each change is undoable in one step per gesture and saved with the take: a name shows on its pad, a hue colors the pad (Auto restores it), a gain is heard, and tuning moves which key plays the slice untransposed. Holding the Keys card opens the armed slice’s settings; tapping its root note opens the tuning alone.

• ✓A tap shorter than ½ s plays the pad and opens nothing
  no panel
• ✓A finger that moves first opens nothing
• ✓A ½ s hold plays the pad and opens its settings, named for the pad
  panel open, title placeholder Pad 4
• ✓A name appears on its pad and is saved with the take
  pad shows "Kick"; stored {"id":27,"label":"Kick"}
• ✓Dragging the hue turns Auto off and colors the pad that hue, saved
  Auto off; stored hue 252°, pad hue 252° (was 48.1°)
• ✓Auto gives the pad its color from the sound back
  pad hue 48.1° (Auto was 48.1°); stored {"id":27,"label":"Kick"}
• ✓A gain of −12.0 dB is heard on the pad (±1 dB)
  -1.0 dBFS → -13.0 dBFS (-12.0 dB)
• ✓The gain is saved with the take
  {"id":27,"label":"Kick","gain":-12}
• ✓One Undo takes back the whole slider drag
  stored after one Undo: {"id":27,"label":"Kick"}
• ✓Tapping the root note opens the tuning alone
  Tuning · Pad 6
• ✓The root moves up a semitone, the cents kept: “A♯3 +29¢”, changed from the detected A3
  KEYS PLAY Slice 6/12 A♯3 +29¢; root key 58; “Changed from detected A3 +29¢”
• ✓Retuned: the A♯3 key plays the pluck at A3 (±2¢)
  heard 220.00 Hz, want 220.00 Hz (-0.01¢; 8 partials)
• ✓The tuning is saved with the take (root only: the cents stay detected)
  {"id":36,"root":58}
• ✓Holding the Keys card opens the armed slice’s settings
  panel for slice 36; the pluck is 36
• ✓Reopened: the name on its pad and the tuning on the keys
  pad 4: "Kick"; keys: KEYS PLAY Slice 6/12 A♯3 +29¢

30Slice settings: where slices play (Pads / Keys / Hidden)10.1 s

The control under the selected marker sets where a slice plays: Both, Pads only, Keys only or Hidden. Pads only: the keys’ ‹ › and a double-tap skip it. Keys only: it leaves the pads, and later pads move up. Hidden: muted on pads and keys, while its cut stays (the hits around it don’t change). Two thin lanes under the waveform show the mapping, upper pads and lower keys, the armed slice’s keys lane bold. Saved with the take; Undo puts it back.

• ✓› skips a pads-only slice
  route pads; armed slice 8, the pluck is 6
• ✓A double-tap on it arms nothing and says so
  REC 4:47 LEFT NEW + IMPORT MIDI walk Slice 7 plays on the pads only 0:13 Saved UNDO LOCAL
• ✓Keys only: it leaves the pads
  12 → 11 pads
• ✓…and later pads move up: pad 1 plays slice 2
  sounds like slice 2 (r 0.990; next best slice 4, r 0.172)
• ✓Hidden: no pad, and silent when tapped
  11 → 10 pads; peak -180.0 dBFS
• ✓Hidden: the keys can’t arm it
  REC 4:47 LEFT NEW + IMPORT MIDI walk Slice 5 is hidden 0:13 Saved UNDO LOCAL
• ✓Its cut stays: the same hits
  12 → 12 hits
• ✓Undo puts the knock back: its pad, and it plays
  10 → 11 pads; peak -1.0 dBFS
• ✓The routes are saved with the take
  {"5":"keys","29":"hidden","68":"pads"}
• ✓Reopened: the same pads, and the knock still hidden
  10 pads

31Slice settings: shape (attack, decay, grain)57.7 s

The settings panel’s Attack, Decay and Grain sliders shape a slice on its pad and on the keys: Attack fades it in from the very start over up to 300 ms (Natural: today’s 2 ms de-click fade); Decay fades it out after the attack, 30 dB down at the time set, and ends it 60 dB down (Full: it plays to the slice end, as before); Grain sets the keys’ grain length (Auto: smear sets it). Each slider says its value, shows ↺ once changed, and previews the sound on release. Measured on the captured audio against the same pad or key played without them; back at Natural and Full the pad is the same sound, sample for sample. Saved with the take; Reset pad clears them.

• ✓Attack, Decay and Grain in the panel, at Natural, Full and Auto, no ↺
  Natural, Full, Auto
• ✓Attack at the far right: 300 ms, shown with its ↺, saved
  shows 300 ms; stored {"id":36,"root":58,"attack":0.3}
• ✓A 300 ms attack measured on the pad: linear, half level at 150 ms, full at 300 ms (±15 ms)
  ramp 297 ms (95 % at 280 ms); 0.510 of the natural level at 150 ms; 1.000 at 400 ms
• ✓Attack ↺ is Natural again; Decay dragged to about 100 ms, shown and saved
  shows 100 ms; stored {"id":36,"root":58,"decay":0.1}
• ✓Decay 100 ms measured on the pad: −15 dB at half of it, −30 dB at it (±2 dB), silent after twice it
  -15.8 dB at 50 ms, -30.9 dB at 100 ms; tail 183 ms (the natural pad is at -21 dBFS there)
• ✓Natural and Full play as before shapes: today’s graph (the source through one gain; a shape adds its own), and back at them the same samples as the first hit (residual under −90 dB)
  residual -300.0 dB against the first hit; gain nodes per hit: 1 natural, 2 with an attack, 1 natural again
• ✓The 300 ms attack on a key: about a quarter at 75 ms, half at 150 ms, whole by 450 ms (±25 %)
  0.28, 0.52, 0.98 of the natural key
• ✓The decay on a held key: −30 dB at its length (±4 dB), silent after twice it though the key is held
  -30.9 dB at 100 ms; tail 179 ms
• ✓Grain at the far right: 500 ms, saved, and a key’s note carries it to the grain engine
  shows 500 ms; stored {"id":36,"root":58,"grain":0.5}; note grain 22050 samples at 44100 Hz
• ✓Reopened: the attack and grain as set, the decay Full
  75 ms, Full, 500 ms
• ✓Reset pad clears them: Natural, Full, Auto, and nothing stored
  Natural, Full, Auto; stored null

32Inside the iPad app: the native engine plays7.6 s

The same page with a stand-in for the iPad app's bridge (window.fieldSamplerNative and a message handler that records what it is sent). The page must hand the take to the native engine (instrument.set: its id, slice table, roots and Norm gains, armed slice), send pads and keys as messages, and make no sound of its own. The real bridge is tested in the iPad simulator (ios/scripts/build.sh --self-test-web).

• ✓instrument.set names the take by id
  bvaO82dNkFI6
• ✓Slice table: start < end, attack and Norm gain set, every slice
  12 slices; armed 5; roots 48.0 38.1 38.0 54.6 54.5 57.3 48.0 81.0 90.9 48.0 57.0 48.0
• ✓The armed slice is pitched, with a period
  root 57.29, period 4.47 ms
• ✓Saved take: no PCM sent
• ✓Key 1 posts pad 1 (slice 0, with its id)
  {"v":1,"type":"pad","slice":0,"id":5,"gain":1}
• ✓Key A posts note.on on the armed slice (index and id), and note.off
  {"v":1,"type":"note.on","key":1,"slice":5,"id":36,"note":48,"smear":0,"gain":0.8}
• ✓No Rec button in an app without native recording (no "record" feature)
• ✓engine.status stopped shows in the header; midi.status on the MIDI button
  IMPORT MIDI 1 Walkthrough take Audio engine stopped: interrupted 0:13 View only UNDO LOCAL
• ✓midi.notes pad: pad 1 flashes and is selected
  ["Pad 1"]
• ✓midi.notes on/off: the key is held down, then released (a note on and off in one batch still shows, briefly)
  held [50], released [], brief [53] → []
• ✓midi.notes: slices found by id; all notes off per channel; lost events or a reset let every key go
  by id [55], channel off [], dropped [], reset []
• ✓MIDI feedback sends nothing back to native (no echo, no second sound)
  []
• ✓The takes menu names the app's build: version, build number, develop build, site, web commit
  App 0.1.0 (202610151840) · develop build · local · web ef31303
• ✓No Web Audio at all
  no AudioContext created
• ✓Unsaved take: PCM sent once, then the id under the same ref
  {"pcm":[{"ref":"mvat6mwe-2","frames":352800,"bytes":1411200}],"first":{"id":null,"local":null,"ref":"mvat6mwe-2","api":"http://localhost:8788","duration":8,"sampleRate":44100,"channels":2},"last":{"id":"OlADdKgBieHQ","local":null,"ref":"mvat6mwe-2","api":"http://localhost:8788","duration":8,"sampleRate":44100,"channels":2},"history":2}
• ✓The take list goes to native (takes.history)
  2 entries
• ✓Plugin: the first instrument.set is the restored one, complete
  1 sent; first: 12 slices, armed 36 (saved 36)
• ✓Plugin: the project's restored edits play, but aren't saved over the take until edited here
  0 PATCH; take still "Walkthrough take"

33Inside the iPad app: native recording and import3.6 s

The app records and imports natively (features record and import): Rec is back in the app, the recording's live peaks come from native, and on stop the page reads the file through the bridge (capture.read, in slices) and saves it as a take on the server, exactly like an import. Import opens the app's document picker instead of the web file input. A refused microphone shows a message. Nothing uses getUserMedia or Web Audio output. Stand-in bridge here; the real one is tested in the simulator (ios/scripts/build.sh --self-test-record).

• ✓Rec is shown in the app
• ✓A refused microphone says how to fix it
  REC 5:00 LEFT NEW + IMPORT MIDI 0 Microphone access is off for Field Sampler. UNDO LOCAL
• ✓Recording shows the time from native
  {"on":true,"text":"0:01 4:59 left"}
• ✓The recording is read in slices and saved as a take on the server
  /t/7mvELvaUZ0eE: audio/wav, 13.00 s (fixture 13.00 s), 10 capture.read calls
• ✓The recording is sliced
  12 hits
• ✓Import uses the app's document picker and saves the file as a take
  "Field import", audio/wav
• ✓No WebKit capture, no web file input, no Web Audio output
  getUserMedia 0, file input 0, AudioContext 0

34Takes are at most 5:008.6 s

A take is at most 5:00 long (the length whose 16-bit WAV fits the 64 MiB upload limit, with headroom). Recording stops by itself at the cap and the take is kept exactly as if Rec had been pressed; in its last 30 s the Rec button counts down. An import longer than the cap keeps its first part and the header says so. Here the cap is shortened to 4 s for the test (window.__fieldSamplerMaxTakeSeconds), and the 13 s walkthrough take is imported twice: as a WAV, which is cut before decoding, and with a 1.5 MB chunk before its audio, which is decoded whole and then cut. In the app (stand-in bridge), a recording the app ends at the cap is stopped and saved, and the app's own cut of a long import is announced.

• ✓While recording the Rec button shows the time left
  "0:03 left"
• ✓The recording stopped by itself at the cap and was saved, as with Rec
  4.000 s saved, cap 4 s
• ✓The header says why it stopped
  REC 0:04 LEFT NEW + IMPORT MIDI take-1016-1012 Recording stopped: it reached 0:04, the longest a take can be 0:04 · sliced in 77 ms Saved UNDO LOCAL
• ✓A longer WAV keeps its first part, in its own format
  4.000 s, 768044 bytes (16-bit stereo 48000 Hz: 768044) of a 2496044-byte file
• ✓The header says what was kept
  REC 0:04 LEFT NEW + IMPORT MIDI walk Kept the first 0:04 of a 0:13 file: takes are at most 0:04 0:03 · sliced in 38 ms Saved UNDO LOCAL
• ✓No slice past the cap
  6 slices
• ✓A file decoded whole is cut after decoding, and saved as the part kept
  4.000 s, 705644 bytes of a 3996052-byte file
• ✓and says so
  REC 0:04 LEFT NEW + IMPORT MIDI Long header Kept the first 0:04 of a 0:13 file: takes are at most 0:04 0:04 · sliced in 44 ms Saved UNDO LOCAL
• ✓A recording the app ends at the cap is stopped and saved, and the header says why
  REC 5:00 LEFT NEW + IMPORT MIDI 0 take-1016-1012 Recording stopped: it reached 5:00, the longest a take can be 0:13 · sliced in 178 ms Saved UNDO LOCAL
• ✓The app's cut of a long import is announced
  REC 5:00 LEFT NEW + IMPORT MIDI 0 Long walk Kept the first 5:00 of a 12:34 file: takes are at most 5:00 0:13 · sliced in 110 ms Saved UNDO LOCAL

35Inside the iPad app: recordings are never lost2.9 s

The app keeps every recording and import until the page confirms it saved as a take (feature recover). A page that loads while the app is still recording (it reloaded, or its process ended) picks the recording up and stops it. Opening another take while a recording is being read leaves the recording on the list of recovered recordings, from which it opens and saves; one can be discarded. While a capture is being read, Import and the takes menu (the demo take too) wait. Stand-in bridge here; the real one is tested in the simulator (ios/scripts/build.sh --self-test-recover).

• ✓A page that loads while the app records picks the recording up (time running)
  {"on":true,"text":"0:03 4:57 left"}
• ✓An unsaved import from before is announced in the header
  0:04 4:56 LEFT NEW + IMPORT MIDI 0 1 recording not saved — see Takes UNDO LOCAL
• ✓While the recording is read, Import, the demo take and the takes menu wait
  {"imp":true,"demo":true,"rowsDisabled":true}
• ✓Opening another take while a recording is read keeps the recording (offered back, not saved)
  REC 5:00 LEFT NEW + IMPORT MIDI 0 Walkthrough take 2 recordings not saved — see Takes 0:13 View only UNDO LOCAL; saved []
• ✓The takes menu lists the recovered recordings
  COPY LINK COLORS Timbre Pitch Dusk Field Mineral RECOVERED RECORDINGS Not saved as takes yet: open one to save it, or delete it. take-1016-1012 0:06 · Oct 16 · recording × Old import 0:06 · Oct 16 · import × YOUR TAKES Walkthrough take 0:13 · Oct 16 · view only × EXAMPLES Demo take 0:08 · thuds, knocks and plucked notes over wind App ? · local · web ef31303
• ✓A recovered recording opens, saves as a take, and the app is told it is saved
  /t/8KeqOrCZ3dSj "take-1016-1012"; capture.saved ["BBBBBBBB-0000-4000-8000-000000000002"]
• ✓A recovered import can be discarded; then nothing is left to recover
  ["BBBBBBBB-0000-4000-8000-000000000001"]

36Inside the iPad app: with the library, a recording is a take before the page reads it3.4 s

With the take library, the app makes a recording (or an import) a take in iCloud Drive as soon as it stops, before the page reads a byte of it; the page then opens that take and analyses it, storing the analysis. Analysis failing (here, a recording the page can't read at first) loses nothing: the take is listed and opens later. Stand-in bridge here; the real one in the simulator (ios/scripts/build.sh --self-test-library).

• ✓The app makes the recording a take at stop; analysis failing loses nothing
  {"takes":["take-1016-1012"],"stop":{"v":1,"type":"record.stop","library":{"name":"take-1016-1012"}},"reads":0,"header":"REC 5:00 LEFT NEW + IMPORT MIDI 0 Couldn't open that take: not downloaded yet UNDO LOCAL"}
• ✓Opened later, the take is analysed and its analysis stored once: candidates, one end each, the version
  {"candidates":160,"ends":160,"analysis":{"dsp":1}}
• ✓Analysed elsewhere first: this page's analysis is refused, and the take opens with the stored one
  {"writes":1,"refused":1,"opens":4,"header":"REC 5:00 LEFT NEW + IMPORT MIDI 0 take-1016-1012 “take-1016-1012” was analysed on another device first: opened with that analysis Opening take… UNDO LOCAL"}

37Inside the iPad app: a recovered recording longer than a take can be is kept whole3.3 s

A recording left unsaved by an app build from before the take length cap can be longer than a take may now be. Opening it never cuts it: with the take library it is saved whole to the library first (as "Save to library" does; the app deletes the capture only once the library has its copy) and opens like any older take over the cap; without the library its first part opens and the capture is not confirmed saved, so the whole file stays under "Recovered recordings". Here the cap is shortened to 4 s (window.__fieldSamplerMaxTakeSeconds) and the recovered capture is the 13 s walkthrough take. Stand-in bridge.

• ✓With the library, an over-cap recovered recording is saved to the library whole, not cut
  {"created":["DDDDDDDD-0000-4000-8000-000000000001"],"reads":0,"saved":[],"bytes":2496044,"last":12.992729166666667,"header":"REC 0:04 LEFT NEW + IMPORT MIDI 0 take-1015-1012 0:13 Saved UNDO LOCAL"}
• ✓and the page opens all of it (slices past the cap), with nothing left to recover
  last candidate at 12.99 s; REC 0:04 LEFT NEW + IMPORT MIDI 0 take-1015-1012 0:13 Saved UNDO LOCAL
• ✓Without the library it opens cut, but the capture is not confirmed saved: it is offered back whole
  4.000 s take; {"saved":[],"library":[],"header":"REC 0:04 LEFT NEW + IMPORT MIDI 0 take-1015-1012 1 recording not saved — see Takes Kept the first 0:04 of a 0:13 file: takes are at most 0:04. The whole file stays under Recovered recordings 0:04 · sliced in 77 ms Saved UNDO LOCAL"}

38Inside the iPad app: takes in iCloud Drive15.3 s

With the take library (feature library), new takes go to the person's iCloud Drive through the bridge, never to the server; the menu lists them, then the older server takes, marked by where they live. Switching takes keeps each take whole here too. Sharing uploads that one take on demand and keeps its link current. Stand-in library here; the real one is tested in the simulator (ios/scripts/build.sh --self-test-library).

• ✓New takes go to the library, not the server
  2 library takes with their analysis (candidates, one end each, the version); 0 server creates; address /
• ✓instrument.set names the library take (take.local; the demo went as PCM before it was saved)
  {"id":null,"local":"70E6712A-24A9-4FF6-9DAC-74A3301B926D","ref":"mvat77t1-2","api":"http://localhost:8788","duration":8,"sampleRate":44100,"channels":2}
• ✓The menu lists the iCloud takes, then the server takes
  SHARE LINK Uploads this take so anyone with the link can play it. COLORS Timbre Pitch Dusk Field Mineral This version of the app can’t save a take’s colors: update it to change them. IN ICLOUD DRIVE demo 0:08 · Oct 16 × demo 0:08 · Oct 16 × ON THE SERVER (OLDER TAKES) Walkthrough take 0:10 · Oct 15 · view only × EXAMPLES Demo take 0:08 · thuds, knocks and plucked notes over wind App ? · local · web ef31303
• ✓Library: an edit made just before switching is saved to its own take
  renamed take "Library renamed", other "demo"; updates [["70E6","Library renamed"]]
• ✓Library: the opened take shows its own name
  demo
• ✓Library: an older server take still opens
  http://localhost:8788/t/bvaO82dNkFI6
• ✓Library: back to the library take, off the server address
  http://localhost:8788/
• ✓Sharing uploads the take and stores its link in the library
  http://localhost:8788/t/52SzhqRKfaFT; 1 server create
• ✓A shared take's edits reach its link
  server name "Renamed after sharing"
• ✓Sharing writes the server take's key to take.json (pending) before uploading the audio
  pending, complete
• ✓A failed link update shows in the header
  IMPORT MIDI 0 Link is down 0:08 Saved Link not updated: down UNDO LOCAL
• ✓Edits are saved as soon as the page is hidden
  1 saves within 150 ms
• ✓A failed save is retried until it goes through
  IMPORT MIDI 0 Saved on retry 0:08 Saved UNDO LOCAL
• ✓An edit made elsewhere meanwhile is not overwritten: the take reopens as it is, and the header says so
  {"name":"Edited on the other iPad","conflicts":1,"header":"IMPORT MIDI 0 Edited on the other iPad “Edited on the other iPad” was changed on another device: opened its latest version (your last change here wasn’t applied) Opening take… UNDO LOCAL"}
• ✓Library take in an app without library.duplicate: no ⧉, no ▾ → Duplicate
  {"button":0,"menu":0,"copyLink":1}
• ✓Deleting removes the take from the library
  1 left

39Duplicate a take6.4 s

The ⧉ button beside the name (or ▾ → Duplicate) copies the open take with its edits as a new take, which opens at once; a toast says so, with Undo. The copy is made on the server (audio, analysis and preview copied there, nothing re-uploaded) and has its own link and key. Editing the copy leaves the original as it was; Undo deletes the copy and goes back to the original.

• ✓⧉ shows beside the name of a take this browser owns
• ✓The copy opens at once, named "<name> N"
  /t/JLKRJkYpb92v "Walkthrough take 2" (original /t/bvaO82dNkFI6 "Walkthrough take")
• ✓The toast says where you are and that the original is kept, with Undo
  Duplicated. You’re now in “Walkthrough take 2”; “Walkthrough take” is kept as it was. UNDO
• ✓The copy holds the same analysis and edits (stored on the server)
  candidates 146/146, ends true, analysis {"dsp":1}, deleted [], sensitivity 0.5
• ✓The copy shows the same hits
  12 (original 12)
• ✓The copy has the same audio bytes
  2496044 bytes (original 2496044)
• ✓An edit to the copy is saved to the copy
  copy: 1 deleted (original 0); 11 hits on screen
• ✓The original, reopened from ▾, is unchanged (screen and server)
  "Walkthrough take", 0 deleted (was 0), updatedAt unchanged, 12 hits (was 12)
• ✓▾ offers Duplicate
• ✓A second copy takes the next free number
  "Walkthrough take 3"
• ✓Undo deletes the copy (server and list) and returns to the original
  copy 200 before Undo, 404 after; listed: Walkthrough take | Walkthrough take 2 | walk | demo | Before colors | Recorded, renamed | walk | demo | demo | Demo take; on /t/bvaO82dNkFI6
• ✓The original is still unchanged after Undo
  "Walkthrough take", 0 deleted

40Portrait1.9 s

Portrait works too (secondary): the same single screen, stacked.

• ✓Fits the screen: no scrolling, keys fully visible
  overflow 0 px wide, 0 px tall; keys end 14 px above the bottom

41Console0.0 s

Errors the page logged during the whole walkthrough.

• ✓No page errors
  none