Three ways to load a clip: search the sample library (results list under the field), 🎲 Random (pulls a random match for the query, or any random sample when the query is empty), or Upload a local file.

The Clip panel shows the decoded waveform with ▶ Audition and a zoom slider. Dragging across this waveform carves a dry phrase straight from the raw audio — see the Phrases section below.

Uploads can’t be interpreted (the brain needs a library media id), but everything client-side works on them: audition, zoom, and dry phrase carving + export.

INTERPRET sends the clip through a RAVE model (a “brain”) that has only ever heard one kind of sound and re-hears your clip through it. The job runs server-side (submit → queue → run → fetch); one job at a time.

The knobs shape the pass in latent space: Temperature (drift from a faithful re-hearing — 1.0 honest, 1.5+ hallucinating), Noise (random energy injected before decoding), Wet/Dry (blend with the original), Shuffle (chops the latent timeline into chunks and deals them at random — the value is chunk size), Quantize (bit-crush in latent space), Dims (touch only specific latent dimensions, e.g. 0,2,6; blank = all), and Reverse (flips the latent trajectory while the audio plays forward). The what do these knobs do? glossary in the panel has the same notes inline.

When an interpretation arrives it is auto-carved: onset detection finds the transients, every cut snaps to a zero crossing (no clicks), and each slice is classified by brightness (zero-crossing rate) and length into a slot:

Sensitivity re-tunes the onset detector (higher = more, smaller slices) — hit Re-carve to apply. Each slice is capped at 2 s.

Long clips carve hundreds of hits, so each slot auto-keeps only the strongest (the Keep slider, default 12/slot); the rest sit in a per-slot “+ N more” bench drawer, drawn dimmer on the waveform. + promotes a benched hit into the kit, − demotes. Raising the slider fills with the strongest benched; lowering trims the weakest kept — manual promotions survive.

On the waveform: click selects the nearest hit for editing; the colored ticks along the top are slice boundaries in each slot’s color. Every hit row has audition (▶), a merge checkbox (see Phrases), a slot dropdown to reassign, and ✕ to drop.

Click a hit’s name (or its carve on the waveform) to open the editor. Everything auditions and exports rendered — what you hear is what you get.

• Trim handles — drag the edges on the mini waveform; both ends snap to zero crossings.
• Shape — Reverse, Normalize, Gain, and Pitch (baked into the render).
• Envelope — Attack / Decay edge fades with a Curve control. On a one-shot that’s a percussive gate; on a phrase it’s a gentle de-click.
• FX — Filter (cutoff/reso) + peaking EQ (freq/gain/Q), Delay (time/feedback/mix), Reverb (size/mix).

Each percussion slot header has a ↻ button: a minimal single-track loop that triggers that slot’s kept hits in rhythmic context, rotating through them round-robin. Deliberately not a sequencer — Litany owns sequencing; phrases never loop.

The default pattern is invisible and diagnostic: it seeds steps from the slot’s own onset timing, so you hear the rhythm the brain actually found. BPM / Rate / Length re-seed that default; touching Probability switches to a generated pattern.

A phrase is any long material — a whole word, a spoken phrase, a melodic run, a texture pull — up to 15 seconds. Phrases are never auto-detected and never loop: they only exist because you carved one. They live in their own panel at the bottom of the page.

Two gestures

• Drag-select — drag across either waveform to paint a selection (ends snap to zero crossings; under 100 ms is discarded; over 15 s clamps with a hint). The bar that appears lets you audition the region, Carve as phrase, or clear. A plain click still selects the nearest hit.
• Merge slices — tick the checkboxes on two or more adjacent carved slices and hit Merge N slices → phrase. Merging consumes the source hits; Re-carve is the recovery path.

Dry vs wet

A phrase carved from the Clip waveform is dry — no brain touched it, it survives re-interpretation, and it indexes without model lineage (tagged dry). One carved from the interpreted waveform is wet (tagged interpreted) and dies with its interpretation — a confirm guard warns first.

Each phrase row shows its origin badge and duration, and has an optional name field — a named phrase exports as that-name.wav instead of kit_phrase_2.wav.

Two destinations, from either export row. Index to library renders each hit and uploads it to samples-bored, where Litany and search find it. Download ZIP bundles the same renders locally (phrases nest in a phrases/ folder).

The kit buttons (bottom of the carve panel) export everything kept; the phrase buttons (Phrases panel) export phrases only — so drums-only, phrases-only, and both are all one-click workflows. The optional kit name becomes the filename prefix and the kit: tag.

Tags on every indexed sample