Whitespace as Provenance argues that a machine representation can preserve every proposition in a poem and still delete the poem.
Before text reaches a model it passes through artifact capture, digitization, normalization, serialization, tokenization, model processing, and rendering. Each stage can remove spatial, typographic, prosodic, stanzaic, or manuscript structure.
The paper’s decisive correction is that tokenization alone is not the problem. A byte-perfect tokenizer can round-trip a linear character sequence while the work’s two-dimensional composition was already lost at serialization. Character preservation is not compositional preservation.
An empirical demonstration compares a spatial fragment with its linearized phrase list under GPT-2 byte-level BPE. Both reconstruct their character strings, but neither machine representation contains the calligram as visual argument. Whitespace survives as tokens without being interpreted as compositional relation.
The paper sketches an extinction gradient: some Whitman lineation is reduced, Dickinson’s manuscript variation substantially lost, Hopkins’s prosodic notation catastrophically weakened, and concrete or calligrammatic poetry extinguished in text-only pipelines.
Pearl becomes the exemplary case of double invisibility. The archive’s copy is a non-OCR PDF, preventing machine-text access; an ordinary OCR stream would still erase the concatenation and group structure through which the poem performs its theology.
The argument extends provenance below words and ideas. Composition is not decoration surrounding content. It is the operation by which the content exists as this work.