Residual instrument

Twelve rings of actual signed GPT-2 coordinates surround their additive waveform; a selected block opens into 768 frequency rays.

One model, five captured tensors

A signed, linear bank

A performance clock

What this encounter establishes

Sources

One model, five captured tensors

GPT-2 small has 768 residual coordinates and twelve blocks. Each capture records the block input, the whole projected attention output, the intermediate residual before the second normalization, the whole projected MLP output, and the block output. The final normalization and vocabulary readout are outside the listening field. Attention here is the complete branch, not twelve individual head contributions.

The trained weights are openai-community/gpt2, revision 607a30d783dfa663caf39e06633721c8d4cfcd7e. The published safetensors digest is verified. The implementation is Transformers 5.8.0 with PyTorch 2.11.0, deterministic CPU float32 inference, eager attention, evaluation mode and no KV cache. A second full architecture is randomly initialized with seed 2026 using the same tokenizer. Its initialization follows that implementation’s conventions.

Six bundled traces cover three sentences in both trained and seeded models. All five tensors are stored as losslessly compressed float32 with token IDs and byte hashes. In these captures, both residual additions, the embedding input and continuity between blocks match exactly under float32 arithmetic. Repeated extraction and repeated seeded initialization are byte-identical. Replacing “time” with “space” leaves all earlier token captures unchanged.

A signed, linear bank

For a vector x, the signal is S Σ xⱼ sin(2πfⱼt + φⱼ). Amplitudes keep their signs. The shared base scale is the reciprocal of the largest vector L2 norm across all bundled captures. Shared trim changes this scalar for every branch, text, block and model together. The listening window applies the same coordinate mask to every vector. No branch or frame is normalized separately.

The logarithmic bank assigns 768 distinct rounded multiples of 0.5 Hz from 55 to 7,040 Hz; the dense bank assigns 60 + 7.5k Hz. Index, reversed index and a seeded permutation assign coordinates to frequency slots. Phases are fixed with seed 314159. Coordinate order has no intrinsic pitch meaning. Changing it changes what is heard without changing any model tensor.

A 65,536-sample inverse FFT forms a two-second periodic table. Playback linearly interpolates adjacent samples and transitions linearly over 80 ms with a shared phase clock. These operations preserve addition before a final tanh safety stage and conservative output gain. The default output bound is 0.049 and the maximum is 0.14. The nonlinear safety stage can alter audible identities, so the mathematical comparisons use the preceding linear signal. The interface indicates substantial shaping.

A performance clock

Choose a word fragment, drag a ring, solo a term or scrub the addition. A, B and B − A compare the same token index, which may refer to different fragments. Edit the eight score steps, choose Addition, Blocks or Tokens, and set the pulse tempo. This is a chosen performance time, not a model inference timer. The marks use an asinh contrast curve so small signed coordinates remain visible. The center shows the linear waveform.

The bundled instrument works offline. With the accompanying local inference server, Add text captures both models for up to 32 GPT-2 tokens and verifies their additions. A static preview instead accepts imported captures and labels that limitation. Entered text is not saved or sent elsewhere. Explicit Start, Mute and Stop control audio; repeated Start reuses one context and node, and leaving the page closes audio. No microphone, hosted inference or analytics is used.

Residual instrument?

One model, five captured tensors?

A signed, linear bank?

A performance clock?

What this encounter establishes?

Sources?

Residual instrument

One model, five captured tensors

A signed, linear bank

A performance clock

What this encounter establishes

Sources