NOCTILUCA

Two hundred bioluminescent organisms drifting through dark water, and the sound they make while doing it.

Open index.html. That's the whole thing — no build step, no dependencies to install. p5.js and Tone.js load from a CDN.

The aim

This is meant to be restful. Not a demo of the flocking algorithm, not a toy with a score — something you can leave running on a second monitor, glance at, and feel calmer for having glanced at.

Everything in here follows from that, and a few of the decisions are only sensible in that light:

• There is nothing to do and no way to lose. No score, no goals, no menus, no chrome. Moving the cursor into the flock scatters it, which is the only interaction, and it is entirely optional. Watching is a complete use of it.
• Nothing demands attention. No flashing, no hard transients, no sudden contrast. The boids' glow pulses are individually phase-offset so the school shimmers rather than throbbing in unison. Alarm reads as brighter, not as a colour change — agitated dinoflagellates flash, they do not turn red.
• The predator is gentle. Panic ramps from nothing at the edge of its radius to full at contact, so boids peel away smoothly instead of snapping, and it bleeds off afterwards rather than cutting out. Scattering the flock should feel like putting a hand through water, not like scoring a hit.
• The cursor has no marker. You never see the thing the flock is fleeing — only the flock reacting to it.
• Sound is never in a hurry. See below.

What you see

Deep water, lit only by what is in it. Boids are drawn additively as three concentric layers — two soft halos and a hot near-white core — so overlapping glows brighten each other instead of painting over. The core is elongated along the heading rather than round, because orientation is what makes the flock's turning legible.

Each organism gets a random depth that scales its size and opacity, so two hundred identical marks read as a volume of water rather than a flat sheet. Trails taper to nothing in both width and opacity. Marine snow sinks slowly in the background. The canvas wraps at the edges, so the flock has nowhere it can be cornered.

What you hear

The reference is Austin Wintory's score for Journey, and specifically one principle from it: instruments represent entities in the world, not mood. The flock should sound like it is making the sound, not like music is playing over it.

So there is no independent clock. No transport, no tempo, no loops, no sequences, no LFOs. Nothing makes a sound that the simulation did not cause, and a still flock is a silent one. Everything you hear is one of two things:

• Wind — a dark, wide bed of filtered noise. Its level follows how fast the flock is moving; its brightness and body follow how tightly it is clustered. Deliberately capped below the point where it starts to hiss.
• Bells — struck pentatonic tones with a 280 ms attack and a long decay, so they swell rather than hit. They are silent by default and ring only on events. Two kinds, and they sound different on purpose: a cohesion chime, low and wide, when the flock pulls tight of its own accord; and an alarm chime, high and tight, when you startle it. They have their own registers and the registers never touch — chimes live in E2–C4, alarms in C5–G5, with a full octave left empty between them. Each rings from where it happened in the stereo field — cohesion from the knot of the flock, alarm from the cursor. Both are detected as a rise rather than as a level crossing, which is what keeps them speaking: a flock that is already dense, or a cursor held still in an already-panicked flock, is not doing anything and stays quiet, while every fresh tightening gets its own bell. Expect one roughly every three seconds, more when you stir the flock.

Audio starts on your first click, because browsers require a gesture before they will make sound. That is what the title screen is for: the flock is already swimming behind it, and clicking through both lifts the scrim and opens the AudioContext. If the audio fails to start the screen still clears and the simulation runs on, silently.

Controls

move the cursor	scatter the flock
click	clear the title screen and start audio
m	mute
c	colour menu — six palettes for the flock, or a custom one; the choice is remembered
f	frame rate
g	sonification readout — raw flock statistics next to the mapped values the audio hears

Files

index.html	the page — title screen and the tags that load everything else
favicon.svg	the mark: one cell, a hot core, two trailing filaments
sketch.js	setup, draw loop, controls
boids.js	the Boid class — flocking rules, panic, rendering
flockstats.js	turns the flock into five numbers the audio can use
flocksound.js	the audio engine
sound-sandbox.html	standalone audition rig — the same synthesis driven by sliders instead of a flock, for trying sounds out without waiting for the sim to do something
tools/	measurement and tests, below

Tuning

The constants in flockstats.js are measured, not guessed, and that matters more than it sounds like it should. Aggregate flock statistics do not behave the way the individual boids do: mean speed here is nearly a flat line because every steering rule pushes toward maxSpeed, and alignment sits above 0.87 essentially always, so neither is much use as an expressive signal. How tightly the flock is clustered is the one continuous quantity with real range, and it does most of the work.

If you change a flocking weight, the boid count or the speed limits, those distributions move and the sound will drift out of tune with the simulation.

jsc tools/flockprobe.js        # print what the statistics actually do, over 150s,
                               # undisturbed and with a cursor sweeping the flock
jsc tools/flockstats.test.js   # correctness tests

jsc ships with macOS at /System/Library/Frameworks/JavaScriptCore.framework/Versions/A/Helpers/jsc; any JS engine with load() and print() works.