Residual instrument: predictions

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

Residual instrument: predictions

Written before extraction or implementation. This is an original listening instrument using GPT-2 small, not a claim about how a language model experiences sound.

Prediction	Confidence	What would count against it
P1. Capturing the attention output projection and MLP output produces `after_attention = incoming + attention` and `outgoing = after_attention + mlp` at all 12 blocks and all bundled token positions, within float32 tolerance.	High	An elementwise discrepancy exceeding 2e-5, a missing branch, or a non-finite tensor.
P2. Repeating deterministic CPU inference with the same pinned weights, tokenizer and tokens produces byte-identical float32 captures.	High	A different tensor digest on the repeated run.
P3. A fixed sine bank with signed amplitudes and one global scale preserves both residual additions and differences before the final output safety stage.	High	A waveform addition error above 1e-6 at checked sample times.
P4. Changing a final word leaves all earlier token captures unchanged under causal attention. The changed position differs.	High	An earlier position changes beyond 2e-5, or the changed word produces identical captures throughout.
P5. The same architecture with seeded randomly initialized weights can be compared with the trained model using identical tokens, frequency assignments, phases and scale.	High	Incompatible token layouts, changed mapping, or an irreproducible initialization digest.
P6. The instrument can start, stop, mute, restart and cancel pending startup without leaving active audio voices.	High	A running context after Stop or a stale start completing after Stop.
P7. Slowing and looping recorded transformations yields a playable encounter.	Open	This requires a person's listening and judgment; automated signal checks cannot establish it.

The subject is the residual stream of one particular model. Its consequence is a signed, audible interference pattern. Per-frame normalization and magnitude-only spectra are rejected because they erase addition and cancellation. The stated limit is that feature coordinates have no intrinsic pitch order; the mapping is chosen by the instrument, and performance time is chosen by its player.

Residual instrument: predictions?

Residual instrument: predictions