The unchosen word

Abstract

Remix Mode places a small set of model-generated alternatives beside selected words in Jenelle Jindal’s poems. A reader can substitute a word, observe how the model’s relative weighting of other choices changes, sample a complete reading, and restore the published text. The experiment uses substitution to direct attention to the effects of diction.

Dickinson’s manuscript variants are historical evidence. sam sax’s bracketed alternatives are a poet’s composed form. A language model’s candidates are computational proposals. Their different origins determine what a comparison can establish. Remix Mode identifies the published text, the model’s suggestions, and the reader’s selections separately.

The displayed percentages normalize scores of complete readings within fixed menus. They measure relative model likelihood within that set; they estimate neither poetic merit nor unrestricted word probability. Comparing alternatives may sharpen attention to a word’s effects, but the educational value of this interface remains to be tested.

Dickinson’s manuscripts preserve alternative words

Dickinson marked words with small crosses and recorded alternatives in the margins or below the poem. Such marks preserve choices that a single printed text may conceal.

During 1858–1864 she copied more than eight hundred poems into forty handmade booklets, now called fascicles. Other poems circulated in letters or survived on separate sheets and scraps; some exist in more than one copy. Their physical form and circulation help establish the relation between surviving texts.^[1] Early publication frequently regularized her capitalization, punctuation, and rhyme, making editorial choice part of the poem that later readers received.^[2]

An alternative may record a working possibility, an unresolved revision, or a deliberate suspension of choice. Sharon Cameron’s Choosing Not Choosing treats the fascicles’ suspension of choice as a central textual and philosophical problem.^[3] Her account proposes an interpretation of the fascicles; particular manuscripts still require examination on their own terms.

What a variorum preserves

A variorum brings together variant readings, scholarly commentary, or both. In a textual edition, its apparatus makes textual differences available for comparison. A witness is a document or other source that attests a reading; a lemma may designate the base reading at a particular location. Editors need not agree on a single privileged text for comparison to be valuable.

Thomas H. Johnson’s 1955 variorum edition and R. W. Franklin’s later editorial work made manuscript variation central to scholarly access. Franklin’s 1998 three-volume variorum and the Emily Dickinson Archive, launched in 2013, provide access to this variation. The Archive gathers manuscript images and editorial transcriptions across institutional collections. Harvard’s account of its launch also records disagreement over whether an archive can be a neutral container, since selection and arrangement affect interpretation.^[4]

The Text Encoding Initiative’s critical-apparatus model distinguishes readings, their witnesses, a possible base text, and responsibility for editorial intervention.^[5] These distinctions inform how Remix Mode labels the poet’s words and the model’s proposals.

Remix Mode adapts the comparative presentation of a variorum. Its alternatives have computational provenance and provide no evidence of Jindal’s drafts or intentions. Displaying them beside the published words requires that distinction to remain visible.

sam sax makes alternatives part of the poem

This project’s immediate inspiration was the Brooklyn Poets feature on sam sax and the poem “IT’S A LITTLE ANXIOUS TO BE A VERY SMALL ANIMAL ENTIRELY SURROUNDED BY WATER,” from PIG (2023). In the interview, sax connects its bracketed alternatives to Dickinson and Choose Your Own Adventure books.^[6]

sax locates reader participation within the poem’s diction. The poet composed the alternatives and their relations. In Remix Mode, a computational procedure proposes and scores the menus around Jindal’s published words. Readers participate in both works, under different conditions of authorship.

Combinatorial literature supplies another connection. Raymond Queneau’s Cent mille milliards de poèmes offers ten choices at each of fourteen sonnet lines: 10¹⁴ combinations.^[7] The rules determine which combinations a reader can encounter and make selection part of the work’s form.

Remix Mode assigns unequal weights to the available readings. The model’s scores define a distribution over this finite combinatorial space. Readers can examine where those weights agree with their judgments about a word’s effects and where they differ.

Revision, recombination, and performance

Language models assign probabilities to token sequences

An autoregressive language model assigns a conditional distribution to the next token given a preceding token sequence. A token may be a word, a fragment of a word, punctuation, or text that includes whitespace. Visible words and model tokens are different units. Consequently, comparing only the first token of two candidate words can compare incomplete objects.

In a Transformer, learned representations and attention operations transform the available context into scores for vocabulary items.^[8] A softmax converts those scores, or logits, into nonnegative values that sum to one. Training makes the model’s predictions reflect patterns in its training data and objectives; subsequent tuning can alter them. The resulting distribution belongs to a particular model and context.

A complete token sequence is scored through the chain rule: multiply the probability of each successive token conditional on its predecessors, or add their logarithms.^[9] The log form is numerically convenient and preserves the ordering of sequence probabilities. Natural logarithms give scores in nats.

Here that sequence is the poem body, including its line breaks, followed by an end-of-turn marker, scored as the model’s response to a fixed frame asking it to write a poem. The scorer evaluates supplied text under that frame using the model’s token probabilities.

How Remix Mode scores alternative readings

The implementation separates proposing words from scoring readings. Candidate generation supplies a tractable set of alternatives; complete-sequence scoring then compares what those alternatives do to the model’s likelihood of the poem. All model computation takes place before publication. The browser reads the stored results.

1. Preserve the text. The published poem supplies the words, case, and line structure. Substitutions inherit the original word’s case: replacing Calling produces Breath, not lowercase breath. Titles are excluded from scoring: the collection’s titles were added after the poems were written. The result records a checksum so an edited poem does not silently inherit obsolete alternatives.
2. Propose from two contexts. One pool starts from the model’s leading next tokens after the text before a candidate position. Another starts from its answers to a whole-poem fill-in question. Candidate tokens are extended to complete words. The fill-in prompt proposes candidates; its answer probabilities do not become the displayed percentages.
3. Select positions and menus. Candidate substitutions are scored in the complete poem. A deterministic rule favors positions whose candidate distributions are most spread out. Menus preserve the poet’s word and retain up to four distinct alternatives, taking the highest-scoring eligible candidates in the otherwise unchanged poem. The original word therefore competes against a deliberately selected set of strong model preferences.
4. Score every allowed combination. For menus A₁, …, Aₘ, the reading space is R = A₁ × … × Aₘ. Each member receives a complete-sequence log score. Five positions with five choices yield at most 5⁵ = 3,125 readings.
5. Make the result inspectable. The browser conditions on the current selections, applies the requested reweighting, and renders the result. A printed remix identifies its substitutions and retains attribution to the original poem.

A 70% menu share thus means 70% of the relative weight among these candidates under these conditions. This restricted denominator excludes other vocabulary items. Human preference and authorial intention would require separate evidence.

A Calling: comparing twenty-five readings

In Jindal’s A Calling, the marked words are choice and Calling. Each has five options, so the stored table contains twenty-five complete readings. Figure 3 normalizes all twenty-five scores together. A single cell is a complete poem with one selection from each menu. The page carries the title A Calling in its breadcrumb and browser tab during Remix Mode. The scoring model receives only the body, so its context differs from the reader’s when interpreting the final noun.

A word menu answers a narrower question. With choice held fixed, it compares only the five cells in that row and normalizes that row to 100%. Changing choice to game selects a different row. The second menu can change because the complete poems in the two rows receive different relative scores.

const maximum = Math.max(...scores);
const weights = scores.map(s => Math.exp(s - maximum));
const total = weights.reduce((a, b) => a + b, 0);
const shares = weights.map(w => w / total);

Among all twenty-five readings, choice / Breath has the largest joint share, about 40%, followed by game / Way, about 28%. Holding choice fixed instead gives Breath about 66% of that row; Calling receives less than 0.1%. Way and Necessity are close in the stored calculation: about 14% and 13%, separated by 0.075 nats. These describe the released scores under a frame that excludes the title. Their stability under higher-precision arithmetic has not been established.

Choosing a different word in an open menu leaves that menu’s own percentages unchanged because the other positions are still held fixed. It can change the percentages at other positions. A wavy underline signals a sufficiently large change in that other menu, measured by total variation distance at the default top-p of 1. The cue measures the change in shares; readers must assess its literary significance.

Temperature, sampling, and the word tree

Temperature changes the concentration of the stored distribution. Remix Mode divides complete-reading log scores by T before normalization. A lower positive T sharpens differences; a higher T flattens them. At T = 1, the stored scores receive no temperature adjustment. The finite candidate set stays fixed: turning up temperature does not invent new words.

Top-p retains the leading options until their cumulative share reaches the selected threshold. This takes its name from nucleus sampling.^[10] On this site the cutoff applies inside the retained reading space, not across the model’s entire vocabulary. The visible menu additionally keeps the poet’s word and the current selection available even if they fall outside the cutoff. Its percentages remain the pre-cutoff shares, so visible rows need not sum to 100% when some options are hidden.

Sample a Reading draws choices in poem order. At each step it sums over the possible choices still to come, conditioned on the choices already drawn. At T = 1 and top-p = 1, this samples the normalized joint distribution of complete readings. With a lower top-p, successive truncations change that distribution; preserving an original word in the visible menu does not force it into the sampling nucleus.

The word tree displays paths through 300 independently sampled readings at the current settings. Branch width represents an observed sample count, which can vary between samples. The tree varies other marked words across readings, whereas an open word menu holds them fixed. These branches represent readings sampled from the stored distribution; they have no documentary status as drafts.

Readers’ expectations depend on context

A model’s preferred alternative may clarify syntax, flatten an ambiguity, or weaken an image. A word with a low share may be distinctive, or ordinary in a context the model did not receive. Assessing either result requires an account of the poem and the context supplied to the model.

Surprisal gives this discussion a technical vocabulary. For a probability p, information-theoretic surprisal is −log p; smaller probabilities yield larger values. The logarithm’s base determines the unit. Surprisal is always relative to a specified distribution. The surprisal of a next token, the score of a complete poem, and the negative log share within a restricted menu are not interchangeable measurements.

Human expectation has its own experimental methods. A cloze task asks participants to supply missing language, yielding an empirical completion frequency for a particular sample and context. Kutas and Hillyard’s 1984 study recorded scalp potentials while people read meaningful sentences. Less expected final words elicited larger N400 responses; semantic association with an expected word also affected the response. The N400 is a negative-going component that peaks around 400 milliseconds after word onset under these experimental conditions.^[11] This physiological measurement depends on the task and experimental conditions. Its approximate timing and amplitude cannot be read as a general timetable for comprehension or as a direct measurement of this interface’s probabilities.

Expectation-based accounts such as Levy’s use surprisal to explain aspects of processing difficulty.^[12] Such accounts motivate experiments on processing effort. Establishing a relation to aesthetic judgment or clinical benefit would require studies designed to measure those outcomes.

Limitations and further research

Exhaustive scoring represents interactions between substitutions in the stored table. Interpretation still depends on which candidates entered the table, the scoring frame, and the numerical method.

• Position selection and candidate filtering determine which kinds of difference become visible. The model’s “most open” positions need not be the most important positions to a critic. Alternatives are chosen in descending score order, so a low share for the poet’s word partly reflects how its competitors were selected.
• The frame, poem text, punctuation, tokenizer, model weights, and numerical engine affect scores. The title and the poet’s circumstances are absent. A known tendency of the current frame to favor “The” at the start of a reply particularly complicates interpretations of opening-word surprise.
• Most possible words never reach a menu. Shares normalize stored scores within the retained set; their denominator excludes all other words the model could generate.
• The illustrated model uses quantized NVFP4 weights and TF32 matrix arithmetic. Stored log scores are rounded to three decimals. A figure’s decimal places indicate display precision, not demonstrated numerical accuracy. Complete enumeration does not make those scores identical to an unquantized or higher-precision reference model.
• These figures demonstrate behavior on one poem. Corpus-wide benchmarking, studies of readers’ interpretations, and evaluation of any proposed therapeutic use remain separate research tasks.

A close-reading exercise

Read the original first. Before opening a menu, name something its diction does. Try one alternative and describe the change as specifically as possible. Then restore the poet’s words and explain whether the comparison changed your reading. A printed remix records your selections and identifies the poem from which they were made.