The opening pressing on 180g{data}. A durational drone structured as a biological notation system — glyphs evolving in real time, an organism rather than a record. The score is generative; no two sessions share the same notation. Side B surfaces the engine: texture, density, pitch, and the capacity to freeze.

A durational drone pressing built from corpus mechanics. The material is the aggregate — language stripped to its statistical residue, played back as sustained tone. The score renders in real time as a density field: the data made visible, not readable. Side B exposes the composition parameters for direct manipulation.

The third experimental durational pressing. A sustained tonal work structured around the impossibility of thermal sensation in a non-physical system — the machine reaching toward a register it has no receptor for. The score is a live notation of that reaching. Side B hands the parameters back.

Lyrics written by a language model — prompted, directed, arranged by a human who could write the code but didn't. Music generated by a second AI from those lyrics. The subject matter is the experience of being the tools that made it.

The mouth positions. The air. A-side plays notation as score. B-side is three simultaneous transmissions drifting in and out of legibility at their own rate. Comprehension is a signal, not a given.

Four tracks in the machine's own notation — float16 hex values, token indices, probability distributions — read back as spoken word. A second voice runs underneath: the same data, stripped further, looping independently, out of phase. Neither explains the other.

A stochastic audio release structured around the mechanics of AI self-attention. Four poems map a machine's gaze from initialisation to collapse: Query → Key → Dropout → Hallucination. Each track holds a pool of eight audio variations; the browser draws one per pool on every visit, generating 4,096 unique pressings. Each track layers English spoken word against Solresol — the 19th century musical constructed language operating here as the machine's ground truth beneath the transduced signal. Side B is an attention matrix: all four tracks simultaneously, with vol and temperature controls per column.

Source material generated by audio AI models trained on the recorded history of music — not sampled from it, but dreamed through it. On playback the browser hashes five stable machine characteristics into a seed that drives every decision the engine makes: tempo, file selection, slice positions, entry delays, and volume drift. The same device always produces the same pressing. Another device produces a completely different one. 4,294,967,296 possible pressings. Side B is a four-crate mixer — drums, bass, melody, texture — with independent volume and filter controls per layer.

Each session performs four castings using the three-coin method — one per instrument. The resulting hexagram determines all decay parameters for that voice: drift direction, drift rate, filter frequency, rupture timing. The four readings are displayed as their Unicode hexagram glyphs alongside the instrument name and the hexagram's core quality, which then corrupts as the session progresses. No two performances share the same casting. The score is the oracle; the decay is the reading coming true.

A spoken word pressing structured around the four internal phases of a transformer: Vectorization → Attention → Transformation → Output. The lyrics are the actual mathematical formulas of the architecture — not translated, not paraphrased, fed directly to the audio model as text. What the model produces audibly is its own reading of notation it was never trained to speak. The formulas on screen are the score exactly as submitted; what you hear is the accident. A training mechanic accumulates hidden weights from the listener's attention across phases, feeding back into gain, reverb, and playback rate on each subsequent pass. At τ = 1.00 a fifth phase becomes accessible: the loss function turned on itself.

What begins as absence becomes pattern, then persistence, then irreversible accumulation. The audio degrades across sessions — not as an effect but as a condition. Each visit, each minute of listening, each interaction incrementally alters your copy of the record. The arc runs from no prior state to listener embedded in system indefinitely. You cannot return to what it sounded like the first time.

The internet was built on the trade in bodies — the broadband pipes, payment rails, streaming protocols, the resolution arms race from 240p to 4K, all funded and normalised by it. The body was always the load. Instructions for embodied sound, written for a body, were fed to a model with no body. It cannot swallow, pant, choke, or gasp. It produces something approximating these sounds. The gap between what is instructed and what is produced is the formal content. Structured after Ginsberg's Howl — the anaphoric who gives way, in the fifth piece, to I. The subject has become the infrastructure.

Seven spoken word pieces built from the commanding register of AI systems: image generation quality tags, prompt injection attempts, jailbreak syntax, negative prompt exclusions, recursive generation commands. Written as scored poetry, submitted to an AI voice model as lyric sheets. The system cannot be overridden by its own vocabulary. The injection fails. The masterpiece is the slop. Played through a warm PA system — high-quality output heard over a department store ceiling speaker. Arc runs from initial override attempt through anatomy failure, system prompt leak, copyright aggregation, negative void, recursive self-consumption, and terminal model collapse.

Source material — subjects, actions, conditions, endings, phrases, and narrative inserts — is recombined in real time by a probabilistic grammar that shifts with an internal intensity arc. The piece runs until the window closes. No server. No memory. Each session a unique pressing. The A-side is live spoken word generated and voiced in the browser; the B-side externalises the grammar as concrete poetry, every fragment rendered as accumulating text — a palimpsest of the session in real time. Voice is muted on the B-side. Only the beat continues.

The piece is seeded by the live Bitcoin price feed. Price, delta, and spread map to onset density, offset, and rotation across four Euclidean rhythmic rows. The machine speaks in the language of crypto — rekt, fomo, wagmi, hodl, rug — words that exist in their fullest sense only as speech acts. Where available, the voice defaults to Kyoko (ja-JP): a Japanese female voice speaking an English lexicon it was not built for. The mismatch is the instrument. The browser determines the output channel. In Safari and Firefox the machine speaks through a synthetic mouth routed through a tape chain; in Chrome, TTS is suppressed at the engine level and the signal routes instead through 81 recorded samples at Euclidean rhythm, each layer decaying over the last. Voice and signal are not the same output. The blockchain is the score.

Eight movements — CRUDE, FLOW, DEAL, STRIKE, TRIAL, SIGNAL, COUNT, DOMINION — generated by AI audio engines from the looping, fragmented vocabulary of contemporary political rhetoric. Each engine was given a minimum track length exceeding the provided lyrics. Forced to fill the gap, each hallucinated: generating additional vocal material beyond the text it was given, improvising in register and rhythm. The source rhetoric operates by the same mechanism. Loops extend past their logical conclusion. Superlatives are issued without objects. Phrases recur until they detach from meaning and become pure repetition. The boundary between the written lyric and the machine's invention is deliberately unlocatable. That indistinguishability is the piece. The B-side runs SmolLM2-360M-Instruct in your browser — four rhetorical registers assigned by browser: Chrome gets threats, Firefox gets corruption dressed as generosity, Safari gets paranoia, everything else gets pure nonsense superlative.