The Hedgehog's Dilemma

An interactive audiovisual installation for the browser.

A white room. A small wooden chair at the centre. Writing on the walls — loneliness, affection, misunderstanding, a sandwich eaten slowly one afternoon. The room listens to how you move, what you look at, how close you come to the chair and how long you stay still, and it slowly changes in response. It never explains itself.

This work grew from a physical diorama made in childhood after watching Neon Genesis Evangelion: an enclosed white room, one dollhouse chair, words and phrases written across every wall. The digital room keeps that diorama's minimalism and tries to deepen it rather than replace it.

Running it

Requirements: Node 18+ and a WebGL2 browser (desktop Chrome, Firefox, Safari, Edge; recent mobile browsers work at lower quality).

npm install
npm run dev          # http://localhost:5173

Production build:

npm run build        # typecheck + bundle into dist/
npm run preview      # serve dist/ at http://localhost:4173

Tests:

npm test             # unit tests (Vitest): emotional engine, selection, scheduling, layout, interpreter
npm run build && npx vite preview --port 4173 &
npm run smoke        # browser smoke test (Playwright + Chromium): full visit, screenshots in test-output/
npm run perf         # startup, frame pacing, shader programs, GPU memory, fixed-state frames (BASE, OUT)
npm run style        # each look from fixed places and through the opening, beside the reference image (BASE, OUT)

Playwright is a pinned devDependency (1.56.1, Chromium 1194). It downloads no browser on install; point it at one with PLAYWRIGHT_BROWSERS_PATH or CHROMIUM_PATH (perf).

Tuning. Add ?tune to the URL for a small panel (loaded only then): the room's light (key, room, sky, glow, cone), the look (original, subtle, reference, and each of its values: fill, area, key, cone, glow, warmth, ceiling, exposure, vignette and its shape, grain, plaster relief, the chair's contact and shadow softness, stains, the key's aim, the skylight's brightness), the memory's light, a button that calls the memory now, and copy values. It changes how things look, never what the Director decides, and stores nothing.

Headphones are recommended. Sound starts only after you press go in.

Being in the room

Walk	W A S D, or ↑ ↓; Shift to go slower
Look	mouse (click the room to capture the mouse; Esc releases it), or ← → / Q E
Go toward what you are looking at	F (a gentle walk; moving yourself cancels it)
Turn to the next piece of writing	Tab / Shift+Tab
Turn to the chair	C
Kneel by the chair	Space, Enter, K, or click it when close
Leave something on the wall	T, or the ✎ button
Pause, settings, leave	Esc, or the ··· button
Touch	drag on the left half to walk/turn, the right half to look; tap writing to read it closely, tap the chair to go to it and again to kneel

Or do nothing. Standing still is one way of being here among many; the room answers looking, reading, coming back to a line, touching, crossing the room, and pausing just as much.

For inspection: add ?debug to the URL for a live overlay of the engine's state, or ?seed=N to fix the variation seed.

The room itself

Minimal, but authored. The emptiness is meant to look chosen, not default.

• Surfaces are painted once at load (src/world/surfaces.ts), with the fine structure of real material scans folded in (CC0, Poly Haven; public/textures/scan, made by scripts/scan-detail.py and loaded by src/world/scans.ts). A scan gives only its texture: the room's colours, warmth and marks are its own, and how strongly each surface takes its scan was measured against the artist's reference. Phones take half, and without the scans the room bakes as before.
  • off-white, hand-trowelled plaster: broad gentle mottling, faint horizontal passes, a fine tooth a few centimetres across, specks, small pits and hairline scratches;
  • sealed concrete, a shade darker than the walls: cloudy, with a soft sheen where people stand by the chair, slightly smoother along the path people walk in, scuffed in arcs around the chair, duller and greyer further from it;
  • a ceiling a little cooler than the walls, with the faintest board joints;
  • a grey painted skirting standing 1.5 cm proud, its top catching the light; and every junction drawn as the reference draws it, a crisp line about a centimetre wide fading out by three: under the ceiling, down the corners, above the skirting, and where the skirting stands on the floor.
  • the writer's hand: graphite dragged along under the lines on the walls and a thumb at their ends; a few smudges elsewhere, an eraser's paler halo.
• Physical history, faint, but there to be found from a metre or two:
  • writing that was rubbed out: a single "you" in the empty space behind the chair, an apology lower and off to one side, and not only confessions (somewhere, a shopping list);
  • masking-tape residue around paler rectangles where something once hung;
  • pencil height marks with an initial near one corner, and a few days counted in fives low on the wall behind you;
  • scuffs along the skirting, graphite dust settled above it under the busiest writing, soft occlusion in every corner;
  • on the floor, a little to one side, the pressed marks of where the chair used to stand.
• Light comes almost entirely from one skylight north of the chair: a square opening in the ceiling, its frosted glass set back 25 cm up a well whose sides are lit by it, brightest where they meet the glass (src/world/Room.ts). The glass is the one near-white thing in the frame, and the well dims and cools with the room's light.
  • The floor is brightest around the chair. The wall you read is the brightest wall, washed from its top centre by the skylight's glow; the side walls, the corners and the ceiling fall away from it, the ceiling brightest round the opening.
  • The light is cream rather than white. The warmth is in the light alone, never the plaster or the grade, so when the room cools (someone too close) it cools as far as it ever did.

What the room does

All of this is implemented and running; none of it is shown to the visitor as numbers or goals.

• The opening. One authored score (src/state/OpeningScore.ts), first visit:

  • 0 s. You tap the black threshold, which is also the title card; that tap lets the room be heard.
  • 0.3–5 s. "The Hedgehog's Dilemma" holds on black and fades.
  • In the dark. A chair creaks, then a switch clicks by the door.
  • Where you are. Low (a child's eye height, 1.4 m), centred on the chair from the back of the room: the room as a box around it, its light at the top. Your first deliberate step stands you up, in one slow movement; looking around doesn't. With reduced motion you start standing.
  • The chair. A wide, soft throw of light finds it alone, and the walls are barely there. The fixture stays dim, the source rather than the subject. You can look around from here.
  • The room. It comes up nearly empty: plaster with old erased marks, and small writing on the floor by the chair. You can walk from here.
  • A hand writes. "I wanted you to", in pen, in silence, then stops. The stop is held for two seconds.
  • Laughter. It comes from the wall to the right of the chair, and "You've got crumbs on you." writes itself there as it is heard, pausing through the laugh. Its caption follows the writing rather than giving it away.
  • Hand-off. Once the voice and its writing are done, at about 23 s on a phone and 24 s on a wider screen, the room is yours. Nothing of the room's own life starts before then, nothing is owed afterwards, and nothing answers the unfinished line until you do.

  Returning visitors skip the title and are in the room in about 9 s. A tap advances the title; Skip, Esc or a hidden tab ends the opening at once, easing the light up rather than cutting. Why this and not another opening is recorded in docs/reviews/2026-10-cinematic-opening.md.

• Accumulation. The room fills over the visit:

  • A few lines soon after the hand-off. Ordinary ones first (cereal for dinner, half a sandwich), written on the wall you face, wherever they first belonged.
  • Then more slowly. Lines held in view are answered by a related line written near them: something they echo or contradict.
  • Some spoken lines write themselves, about one in three, onto their place on the wall as they are heard. Most leave nothing.
  • The unfinished line is finished when someone stays with it: a long look at it, or being still at the chair, facing it, for a while. A voice from the line says the whole sentence, the hand writes "stay." as the word is heard, and the room is quiet for a while afterwards. Until then, nothing answers it.

The walls are not partitioned

Each wall keeps a dominant character — north holds the dilemma itself, east the ordinary warm past, west being misread, south almost nothing — but the emotional spectrum leaks.

• Some lines begin out of place:
  • "You made me laugh so hard I couldn't breathe." is in amber crayon among the shame on the west wall.
  • "I didn't mean it. I meant some of it." is among the warm memories.
  • "I kept your jumper. I don't wear it." is on the dilemma wall.
  • The most intimate line in the room, "You fell asleep on my shoulder and my arm went dead and I didn't move it.", is alone on the wall behind you when you arrive.
• As the visit deepens, material migrates. From a little after two minutes, at most two lines while settling in, four once the room starts contradicting itself and seven late in the visit.
  • A line is written, unchanged, on a wall where it doesn't belong, beside the line that most belongs there. For example, "You always took the last biscuit and I let you." beside "sorry", or "I love you. Go away." beside "That was a good day."
  • It happens on a wall you aren't looking at, so it is found rather than watched.
  • Each line that hosts a stranger does so once, and floor and ceiling writing never wanders.
• Late in the visit, the line you held longest is written small on the empty wall behind you.

Walls' characters are in src/content/walls.ts; migration is ruleContaminate in src/state/Director.ts.

Meeting the same words again

Meaning in the room accumulates through repeated context, not through more content. Every line has a history of being met: a settled look is a meeting, and looking away for a while and coming back is another. Each time, the line is the same. What happens around it is not.

"I wanted you to stay." is first just written there, with no sound. The second time, it is said quietly from behind you, and beside it, in lighter pencil, someone has written "I didn't know how." The third time, nothing new appears; somewhere else in the house a door closes. After that it is only itself again.

About fifteen central lines have authored arcs like this (src/content/encounters.ts). Every other line with relations develops through them, restrained: alone; then one relation; then a different one, or its own words in a voice they weren't written in; then rest. Most meetings are deliberately empty, and empty ones are never delayed, so the visitor never learns that looking makes something happen. The same words met elsewhere (re-written on another wall, remembered on the floor by the chair) count as meeting them again.

The door, a cup set down next door and the chair creaking are recorded textures in public/voices/ (generated offline like the others), with synthesised fallbacks.

Situations

Underneath the room's general behaviour there is a second layer: a few authored episodes that bind together a sound, a place, a piece of writing and a moment later in the visit. Nothing announces them, and each visit holds a different few.

• The next room. Laughter through a wall. If you go to it, something happy is written there. Much later, at the chair, the same words appear small on the floor, and where the laughter was there is quiet. If you don't go, it calls once more from somewhere else, then lets go.
• Someone humming. While you stand still, someone hums beyond a wall. Walk toward it and it stops mid-phrase. Later the piano plays the same tune, and stops where the humming stopped; if you let them finish, it plays it through.
• Listening. Two voices next door. Go right up to the wall and stand still to hear them, and they stop. Much later, on another wall, a question.
• Residue. Kneel by the chair a while. When you leave, a faint warmth stays on the floor where you were, not at the chair. Come back later and the chair has turned toward that place.
• Erosion. Some minutes after you write something, while you are elsewhere, it is rubbed out except for one word. Sometimes the room's own writing shares that word and appears faintly beside it. Only that word is kept by the engine, never your whole sentence, and "forget" removes it.

Situations live in src/state/situations/: Situation.ts holds the runner, and library.ts holds the five episodes as small state machines. A situation claims the channels it uses (sound beyond the walls, voices at the chair, music), and the general rules keep out of its way. Any situation can be forced for inspection with ?situations=next-room,humming.

Vignettes

Small remembered fragments, heard rather than seen. Each is one or two voices and the pauses between them: someone asking for a charger on their way out; a question asked in a car at night and answered too quickly; a phone call rehearsed and never made; a child asking someone to stay until they fall asleep; a shop assistant's ordinary kindness; an invitation that will never be taken up; the last biscuit; a party where nobody notices you've gone quiet. Most are sad, two bittersweet, one warm.

They open beside a line on the wall (read it, or touch it), or from stillness at the chair. They keep their distance from the other memories, and visitors meet them only once their recordings have been kept. Some come back differently: on a return visit, or when the visitor has been keeping their distance. A few leave something:

• a word goes from a line;
• a line becomes harder to read the closer one comes;
• a single word is found later on another wall.

The room stays plain while they play: no figures, no cutscene. Turning away never ends one; walking off lets the sentence finish. The catalogue is in docs/vignettes.md and the decisions in docs/reviews/2026-10-weight-of-absence.md. ?vignette=<id>&variant=<name> previews one.

Architecture

src/
  content/      phrases.ts (119 authored phrases), relations.ts (re-contextualisations,
                voice fragments, interpretation lexicon, departure lines), types.ts
  state/        VisitorModel  observable behaviour: distance, stillness, approach/withdraw
                              cycles with hysteresis, attention and reading, surfaces visited
                EmotionalField latent "weather": warmth, exposure, solitude, relief, tension,
                              reminiscence, hush, dissonance — competing needs with inertia
                Memory        motifs and warm memories (session only)
                Engagement    what the visitor does, as signals (read, return, tap, approach,
                              kneel, look around, cross the room, pause, write) → depth
                Pacing        rhythm: recovery / chosen quiet / stalled; repetition
                PhaseTracker  arrival → familiar → contradiction → transform, by depth
                OpeningScore  the opening as data: light, image, sound, permissions, events
                PhraseSelector deterministic weighted choice: field affinity, contrast,
                              motif bias, relation boosts, recency
                Director      the composer: rules that turn all of the above into cues,
                              plus continuous environment targets (light, scale, sound)
                cues.ts       the only vocabulary between the composer and the world
  core/         seeded RNG, event scheduler (spacing of major events), maths
  world/        Room (procedural plaster/floor with scanned detail, skirting, seams, the
                skylight's well), Chair (square ladder-back, wood atlas with scanned grain),
                scans (CC0 material detail), Style (the looks), Lighting (area light + shadow-casting
                spot that can drift), WallText (canvas handwriting on lit planes, dissolve
                shader, annotations, links, strikes, shy/near legibility, attention),
                layout (authored regions + deterministic packing), CameraRig, Post (grade),
                handwriting.ts (the hand's pace; following speech), strokes.ts (stroke order)
  audio/        AudioEngine (buses, limiter, procedural reverbs, ducking), Ambience (hum,
                ventilation, outside, unresolved drone), Music (generative piano-like
                phrases that sometimes stop before resolving), Voices (recorded assets →
                synthesised formant whispers/murmurs/laughter/humming → optional browser speech)
  interpret/    schema.ts (validated Interpretation), LocalInterpreter, Interpreter interface
  ui/           preferences, captions
  App.ts        wiring: input, cues → world/audio, UI, departure/return

Design rules the code follows:

• The engine is deterministic and inspectable. Given a seed and a sequence of behaviour, the room produces the same cues (tested). window.__hedgehog exposes the live engine; ?debug shows it.
• Cues are data. The world and sound only ever receive typed cues and continuous environment targets; nothing else reaches them.
• Interpretation is structured and validated. A contribution becomes an Interpretation (dimensions, related/contradicting phrase ids, a whitelisted set of gestures). validateInterpretation drops unknown ids, unknown gestures and any extra fields. A future model-backed interpreter would implement the Interpreter interface (none is included or wired) and could only ever return this shape — never code, styles or audio commands.

Sound and voices

• Everything in the ambience and music is synthesised live with the Web Audio API.
• Positional sound uses HRTF panners: the hum comes from the ceiling panel, ventilation from a high corner, voices from the chair, from a wall, from behind you, from beyond a wall (muffled), or from nowhere (drifting, unpositioned).
• public/voices/ holds recorded textures (laughter through a wall, breathing, humming, murmured conversation, whispers) generated once, offline, with the ElevenLabs sound-effects API via scripts/generate-voices.mjs. They are loudness-normalised at load time. The app makes no network calls to any service at runtime.
• Spoken lines — 23 phrases have recorded, intelligible performances (38 in the room, plus 6 retired takes kept for comparison), in public/voices/dialogue/ (see Recorded voices below). The room chooses which performance of a line to use from its emotional field, and plays it from a place in the room. Lines without a recording keep the earlier unintelligible textures; captions carry the words either way.
• Where a voice comes from, in order: the recorded performance chosen for this moment → another recorded performance of the same phrase → the browser's own speech, only if the visitor turns that setting on → a recorded texture or synthesis. A recording always wins over browser speech (src/audio/selection.ts).
• The optional browser speech setting speaks unrecorded lines intelligibly with the system voice. It is off by default and non-spatial.
• Safety: a limiter before the output, all level changes ramped, at most four voices at once, generous headroom (measured peaks for every voice type and the music were between −18 and −35 dBFS in headless Chromium).

To regenerate the recorded textures: ELEVENLABS_API_KEY=sk_... node scripts/generate-voices.mjs (the key is read from the environment and never stored). Sound-effect output from ElevenLabs is subject to their terms for your plan.

Recorded voices

The voices are meant to sound like fragments of somebody's life, not a narrator reading the walls: people who laugh, bristle, get embarrassed, lose the thread, and break. They are generated offline with ElevenLabs text-to-speech. The browser never contacts ElevenLabs; it only plays files listed in public/voices/manifest.json.

Why the first takes were flat, and what changed. The first batch (eleven_v3, October 2026) gave each short line one generic tag ("[softly]"). Measured, those takes are short (1.0–1.4 s), unpaused and narrow in pitch (3–6 semitones): technically valid, emotionally inert. The current takes are directed:

• Model: eleven_v4 (available on this key; probed with three-character requests).
• Stability: 0.2–0.3 rather than 0.5, so delivery is allowed to vary.
• Tags placed where the voice should change, not one at the front: [holding back tears] I wanted you… [voice cracking] to stay., [giggling] I love you. [laughs, pushing him away] Go AWAY. Ellipses and capitals shape rhythm and emphasis. The authored words never change: validation rejects any input whose words differ from the line, and speech-to-text flags any added word.
• Unspoken scene context. v4 accepts previous_text / next_text. Each take carries the other person's line before and after ("It's two days. You're being dramatic about this."), or a line of narration (He found the photograph in a box he hadn't opened in twenty years…). It shapes how the words are said, is never heard, and was not billed. A short isolated sentence is hard to perform; context gives it somewhere to come from.

The lab (scripts/voice-lab.json, node scripts/generate-dialogue.mjs --lab, files in audition/lab/) varies one thing at a time on the same line and seed: model alone; direction; stability; scene as narration vs as dialogue; another voice; and a spoken lead-in cut away afterwards using Scribe's word timestamps. Its measurements (below) chose the method; only listening can choose the takes. (Strictly, the order should have been lab → listening → batch; the main batch followed the lab without anyone listening in between, so every take is a candidate, not a choice.)

Readings. Three clearly different readings of each of the three central lines, in different voices:

Line	Readings
I wanted you to stay.	nearly breaking (Jessica) · frustrated, abandoned (Liam) · affectionate, a quiet request (Sarah)
I love you. Go away.	exhausted, overwhelmed (Matilda) · teasing, laughing (Jessica) · angry, hurt underneath (Chris)
I think I was happy then.	nostalgic, an unexpected smile (Liam) · unsure, trying to remember (Sarah) · aware of what was lost (Bill)

…and, across other lines: laughing happiness ("You've got crumbs on you."), joy ("You made me laugh so hard I couldn't breathe."), irritation, embarrassment, confusion, tenderness, grief ("Do you remember me?", at a mother's bedside), and plain happiness ("That was a good day."). Earlier takes of the three re-directed lines are retired: kept on disk and in the ledger, out of the room unless someone approves them.

Choosing a reading. The Director picks a performance with rankPerformances (src/state/performance.ts) from the existing EmotionalField, as a weighted mean (so a reading directed for several feelings at once doesn't win just by listing more): warmth → the affectionate request; tension and dissonance → the abandoned one; solitude and exposure → the breaking one. Each of the three readings of each central line wins in some state of the room (tested). A line heard a second time is given the reading furthest from the first.

Leaner alternatives (audition: true): an external review warned that stacked tags ("[holding back tears] … [voice cracking]", "[angry, jaw tight] … [voice breaking]") and synthetic laughter risk swapping flat for acted. So the audition sheet also has leaner readings — the unspoken scene doing the work and at most one non-verbal tag — and re-situated ones ("Do you remember me?" as an old school friend at a bus stop rather than a bedside). They stay out of the room unless approved. Five more lines that visitors heard most often in the old v3 voice ("If I come closer, will you stay?", "You understood me once…", "I'm glad you came.", "Please don't come any closer.", "I was afraid you'd leave.") have new v4 readings in the room.

Playback (src/audio/Voices.ts, AudioEngine.ts), after an audio review:

• Recordings are only half-normalised, with an authored levelDb per take, so a whispered take stays quieter than a raised one (full normalisation had squeezed a 12 LU spread into 2).
• A 5 ms fade-in and 30 ms fade-out on every recording; four takes whose files ended mid-sound were re-cut or faded (recorded in the ledger as edit).
• The output stage is a true safety limiter (it had been a heavy compressor with automatic makeup), with volume after it.
• Through a wall, recordings are muffled gently (2.2 kHz, quieter) instead of at 750 Hz, which had removed the breath and catches that were directed.
• The hush before a voice falls fast enough to be felt and never lifts a deeper silence already under way; it belongs to the voice cue, so it can't happen without the voice.
• Voices from nowhere drift from a random side, and less widely.

Auditioning. npm run audition builds audition/index.html: every take grouped by line, side by side (with "play in turn"), each with its reading, voice, model, input, unspoken scene, transcript, detected laughs and sighs, and measurements; mark each keep / not this / unsure, then apply the generated command:

npm run audition                      # then open /audition/ from `npx vite` (or any static server at the repo root)
node scripts/generate-dialogue.mjs --approve=stay--breaking,love-go-away--teasing --reject=happy-then--unsure

node scripts/audition.mjs --out=<dir> writes a self-contained copy (page + audio) to share.

Measurements (scripts/voice-analysis.mjs): pitch range and movement, pauses, loudness range, voicing (breath, whisper, laughter), ending contour. They show that takes differ, not that they convince. Median pitch range went from about 6.5 semitones (old takes; nine of them were already above 9) to about 12 (new); most new takes have one or two pauses. Pitch range partly tracks breathiness here (r ≈ −0.6 with voicing), so it is not a clean measure of expressiveness. Speech-to-text detected the directed laugh in most, not all, takes that ask for one (not in this-bit--fond), and some directed sounds ([scoffs], [exasperated sigh], [long exhale]) are faint or untagged. With the same seed, stability 0.2 vs 0.5 changed little: the seed dominates, so several seeds per line, chosen by ear, would be the next step. The audition page hides these numbers until asked, so they don't prime the ear.

Generation (scripts/generate-dialogue.mjs, logic in scripts/dialogue-lib.mjs):

node scripts/generate-dialogue.mjs --dry-run                       # plan and estimated credits, no key needed
ELEVENLABS_API_KEY=… node scripts/generate-dialogue.mjs --budget=900
ELEVENLABS_API_KEY=… node scripts/generate-dialogue.mjs --lab --budget=700
ELEVENLABS_API_KEY=… node scripts/design-voices.mjs --budget=400    # bespoke voice previews (Voice Design)
node scripts/generate-dialogue.mjs --approve=… --reject=…           # after listening

• Idempotent. A request is identified by voice, model, input, stability, similarity, scene, lead-in and seed (fields absent from older takes leave their hashes unchanged). Identical requests are never repeated; a lab take that matches a performance exactly can be adopted rather than paid for twice.
• Protected. Approved recordings are never overwritten without --regenerate=<id>.
• Budget. Jobs beyond --budget stay queued.
• Failures. Rate limits are retried with back-off. Invalid voices are recorded and skipped. Quota and permission failures stop the run and leave the remaining jobs queued.
• Checks. Every file is checked to be valid MP3 and transcribed (ElevenLabs Scribe, with audio-event tagging). A take whose words don't match, that speaks a tag aloud, or that adds words is marked needs-review and kept out of the manifest. (One did: "[quietly, the smile fading]" was spoken aloud; the direction moved into the unspoken scene and the take was regenerated once.)
• Ledger. Status, hashes, credits, request ids, transcripts, detected audio events and review decisions go to scripts/dialogue-ledger.json (lab: scripts/voice-lab-ledger.json).

Account capabilities, as tested (October 2026):

• Text-to-speech works with ElevenLabs default voices on eleven_v4, eleven_v3, eleven_multilingual_v2, eleven_flash_v2_5 and eleven_turbo_v2_5, as do speech-to-text and the dialogue endpoint. eleven_v4 accepts previous_text/next_text; eleven_v3 rejects them.
• Default voices that work: Jessica, Chris, Laura, River, Will, Sarah, Matilda, Alice, Charlie, George, Liam, Callum, Brian, Eric, Bill, Daniel, Roger, Bella, Harry. Library voices (e.g. Aria, Rachel) return 402 paid_plan_required; Lily returns 404.
• The key now has full access (October 2026), so both generation scripts read the real balance before every batch and refuse one the account can't cover; without user_read they would fall back to the ledger and --budget. Costs are also tracked per take from each response's character-cost header. The account showed 2,596 credits used after the two Creator-plan rounds, against 1,982 in the ledger for takes, so the voice previews (and speech-to-text checks) came to roughly 600, close to the 520 estimate.
• Credits: first batch 815. Second round: lab 670, main batch 863, one retake 55, alternatives and re-directions 357, availability probes about 70 — about 2,015 in all.

After the Creator plan (121k credits a month, 30 custom voice slots, Voice Design):

• Voice Design works (scripts/design-voices.mjs, scripts/voice-designs.json, previews in audition/design/): four bespoke voices described by character (an older worn woman, a tired man, a close whisperer, a small shy young-adult voice), three previews each. Nothing is saved to the account until the artist keeps a preview (--save=<preview>:<key>), and saving uses a slot, not credits. Voice Design reports no cost; the ledger estimates one credit per character of preview text (about 520). The key still can't read library voices (voices_read).
• The voice library (readable now that the key has full access): a search for older, worn and whispering voices, and for children's voices, gave a shortlist of 12 (library in scripts/voice-designs.json; node scripts/design-voices.mjs --library), each reading the same text as the designed voices so they can be compared. The library's "children's voices" are mostly bubbly cartoon characters or adults who narrate for children; only two (Kavya, Omar J) are soft enough for the lines in a child's hand, and both are performed or designed voices of unstated age, not real children. The samples were spoken straight from the library, so no voice slots were used: 632 credits by the response headers (the account balance fell by 545).
• Several seeds per reading: the earlier measurements showed the seed matters more than stability, so the gap takes and the saddest kept readings have two more seeds each (variantOf), grouped on the sheet to pick the best.
• The whole spoken corpus: the 14 voiced phrases that had no recording, plus alone-ok, have two readings each (mostly whispered or sad, as the artist prefers, with a warm reading where the line allows), each in two seeds — all audition-only until kept.
• Bitrate stays at 128 kbps: 192 kbps makes every file half as large again for nothing a phone speaker can show. Loading is staged: every sound and the first two takes of each line on arrival (tested ≤ 3 MB), the rest a few at a time when the browser is idle; a take chosen earlier is fetched when needed.
• Round-2 credits: about 1,836 for 89 generations (84 candidate takes; five retakes of lines that repeated or misspoke) plus about 520 (estimated) for voice previews, against a cap of 8,000. No top-ups, overage or plan changes.

Audition decisions. The artist's choices on the shared sheet are applied with --approve/--reject; rejecting now also retires a take from the Director's ranking (retired), and approving brings an alternative or retired take back. The first full round kept 29 takes, rejected 20, and adopted six kept lab takes at no cost; eight sadder or whispered takes were then directed for the lines left without a recording.

What has and hasn't been checked. Technically, everything above is automated. Artistically, only what the artist marked keep has been approved by ear; every other take is a candidate, however well it validates. Whether a performance sounds like a real person is exactly what the audition sheet is for.

Accessibility, comfort and privacy

• Settings (Esc): volume, mute, captions (off by default, an accessibility choice: the artist's decision, 2026-10-19), browser speech fallback, intensity (gentle halves light flashes, holds back the darkest extremes, softens voices and music, and makes silences less deep), reduced motion (follows the OS setting by default; removes head-bob, sway and grain animation, shortens transitions), quality.
• Keyboard-only exploration is complete (Tab to writing, F to approach, C for the chair). Touch is supported. The settings dialog traps focus.
• No flashing; all light changes take seconds.
• Privacy: nothing leaves the browser. Contributions exist only in memory for the current tab and vanish when it is closed; "forget what I wrote, and that I was here" removes them immediately and clears the visit count. The microphone is never used. Only preferences, a visit count, how many times each memory has been met, and one word, close, if the last visit stayed close to the chair (it turns the chair before the next visit opens), are stored (localStorage). Nothing about where anyone stood or what they wrote is kept. "Forget" clears all of it, and nothing of that visit is written afterwards.

What is incomplete or uses fallback content

• Spoken lines: 20 phrases have recordings the artist kept; every voiced phrase now has candidate takes waiting on the audition sheet. Lines without a kept take use unintelligible textures with captions.
• Recorded textures were generated, not performed; I could check their duration and levels but not audition them. One clip that came back near-silent was discarded.
• Visual verification was done in headless Chromium with software rendering (about 10 fps, so behaviour over time was checked with the engine's fast-forward and the opening on its own wall clock). Real-GPU performance, how the light-up looks on an OLED phone, and whether the handwriting reads as a hand at full frame rate still need a human eye.
• Not yet heard by the artist: the new switch and graphite sounds (in the opening now), and two music pilots (on the audition sheet only, not in the room). Music is never in the opening or under a voice.
• Deferred, on purpose: writing in the air (at most once a visit, after review), the light as a motif later in the visit, and a contact-shadow decal for the chair (see the review).
• Speech/microphone input is intentionally not implemented.
• Remote AI interpretation exists only as a typed interface plus a strict validator.

Testing and review

• npm run opening watches the opening in real time on an emulated phone, without the autoplay override: the title on black, the chair lit before the walls, the room nearly empty, the first line written over time and left unfinished, the voice's words written as heard, nothing from the Director before the hand-off, audio running after the tap, Skip, and touch afterwards. tests/opening.test.ts checks the score's timings, the hold and the hand-off, filling and finishing, speech-following writing on real recordings (each word never before it is heard, and the line ending within a second), stroke order, and the timing data.

• Specialist reviewers. Seven reviewer agents (.claude/agents/, docs/agents.md) reviewed the brief before it was built, the Skeptic challenged them, and the decisions are recorded in docs/reviews/2026-10-cinematic-opening.md.

• The unit tests (Vitest) cover pacing (tests/pacing.test.ts: five playtest personas against the targets — something seen or heard in the first 10 s, a perceptible answer within 30 s, a recorded voice within 90 s, no perceptible gap over 25 s in the first 90 s, no more than one development every 7 s, phases that differ by behaviour; engagement signals; rhythm states), and the visitor model (stillness, approach/withdraw hysteresis, reading), the emotional field (non-monotonic closeness, relief after withdrawal, hush), the phrase selector, the scheduler, deterministic and divergent experiences for scripted visitors (still, hesitant, close, wandering, warm), contributions (always written, never answered within a minute, at most one gesture), the interpreter and its validator, the corpus, wall layout, each authored situation (followed and ignored paths, absences, what generic rules must not do while a situation holds a channel), contamination (lines start out of place, migrate gradually beside what belongs, vary between visits; the held line reaching the empty wall), recorded voices (direction matches the corpus; idempotent, budgeted generation; rate-limit, invalid-voice, quota and invalid-audio handling with a simulated API; transcript checks; manifest compatibility; playback priority; context-based performance choice; no credentials anywhere and no API calls from the browser), and encounters (the three meetings with "I wanted you to stay.", what counts as a new meeting, default arcs, variants that depend on the visit, pacing).

• npm run perf (tests/playtest/perf.mjs) measures, on desktop and an emulated phone:

  • startup (ready to enter, the score running, the hand-off);
  • long tasks;
  • frame intervals in the room and during the memory;
  • shader programs at the threshold, and any compiled later (by key);
  • draw calls and triangles across all post passes;

Credits

Fonts (bundled, SIL Open Font License): Reenie Beanie, Caveat, Nothing You Could Do, Cormorant Garamond — via Fontsource. 3D: three.js. Material scans (CC0, Poly Haven): White Plaster 02 and Fine Grained Wood by Rob Tuytel; Concrete Floor Worn 001 and Wood Table Worn by Dimitrios Savva and Rico Cilliers (public/textures/scan/CREDITS.md). All writing in the room is original.