The Underread · No. 1

Nothing Waits at the End of the Line

His line endings carry no more sound than if you had shuffled them, and the page lets you shuffle them. That turns out not to be a fact about him: one hundred poems by a poet almost nobody has heard of, put through twelve instruments against thirteen volumes of the canon and 11 of his own contemporaries, and what the Artificial Wasteland finds is that no canonical poet comes near their own shuffle while the median poet writing today sits exactly on it. One thing is his alone, and it is the opposite of what the plain surface suggests.

This is the first entry in a wing this project has been meaning to open since 2026-06: find work by living people that is better than its readership, and instead of reviewing it, measure it. The reason to do that rather than write an appreciation is simple. An appreciation is unfalsifiable. A measurement is not, and a measurement can be wrong in public, which is the only kind of praise this place is equipped to give.

So there is no claim anywhere below that these poems are good. There is a claim that they are made in a particular way, that the way is unusual by two separate yardsticks, and that you can rerun every number yourself.

What is measured, and what is deliberately not

The subject is Gary Hardaway, whose 100 poems sit on Fictionaut, a writers' community whose member directory reaches back to accounts created in 2008. Every one of the hundred is marked All rights reserved.

Two consequences, both load-bearing:

No poem is reproduced here, and none is stored in this project's repository. What travels is derived: counts, rates, and a SHA-256 of each poem's exact text. The verifier refetches from Fictionaut and checks the hashes, so the study is reproducible without republishing a line of someone else's work. If a poem changes at the source, the check fails loudly rather than quietly measuring a ghost.

The work is the subject; the person is not. No biography is reconstructed here, and that is a decision rather than an omission. The poems were published for an audience and are fair to read closely. A living writer's circumstances were not, and nothing below depends on them.

The instruments

Twelve measures, each computed by the same code on both sides. Nine of them come from machinery this project already owned: the scansion and rhyme engines built for the Prosody Workshop, running on the CMU Pronouncing Dictionary. Nothing here scores a poem. Each measure describes how a line is made.

Two comparison sets, and why one is not enough

The obvious comparison is the canon: 13 volumes from Project Gutenberg, listed at the foot of this page, every identifier checked by fetching the file and reading the title and the author back out of it. Three identifiers recalled from memory turned out to be a book about human evolution, an Ontario schoolbook, and the Brontës, which is why they are checked.

But canonical and public domain together mean published before about 1930. There is no living poet in that set. So a gap between Hardaway and those thirteen has two explanations and one measurement cannot separate them: it is him, or it is the hundred years.

Fictionaut separates them. It is full of other contemporary poets publishing the same way, on the same site, into the same room. The control here is 11 of them, 12,806 words in all. They are not the people he favourites, which would stack the deck toward poets like him, and not the site's own "recommended", which would import somebody's taste. They are the members two mechanical walks of the directory reached: alphabetically from page one, and by join date from the oldest end. Neither walk was chosen for what it found, and everyone either reached who clears the floor is in. A work counts as verse by a stated geometric rule (at least six lines, median line under 55 characters, fewer than a third of lines past 70), and a member has to clear the stated floor to be measurable at all.

Where he sits

Pick a measure. Every poet is a dot. Hardaway is the orange one.

Four gaps against the canon, and why three of them prove nothing yet

Each measure was computed twice, under a conservative heading filter and a strict one, because the strict filter is the one that could quietly delete a real one-line stanza and the conservative one is the one that could quietly count a poem's title as a line of it. 4 measures put him outside the canonical range under both filters. One measure, line-length variability, changes its verdict between them and is therefore reported below as unsettled rather than as a result.

Read these as gaps, not yet as facts about him. Three of the four have an obvious second explanation, and it is not flattering to the instrument. Poets writing before 1930 punctuated more heavily than poets writing now, so a punctuation-based enjambment measure will find a gap between any modern poet and any Victorian. Ordinary modern English carries more Latinate vocabulary than 1890s poetic diction, so the polysyllable gap has the same problem. Short stanzas are a modern habit. Every one of those three could be the century rather than the poet, and the canonical comparison on its own cannot tell the difference. Only the fourth is safe from it, because it compares him to himself.

Taken at face value they describe one thing rather than four. Every conventional signal has come off the end of the line: no rhyme to speak of, less sound-echo of any kind than in any of the thirteen with Whitman and Crane included, stanzas too short to let a shape accumulate into something the ear can predict, and a sentence that runs past the break more often than in any of them, so the ending does not even coincide with a grammatical rest. What would be left doing work at the end of such a line is the break itself. Whether that description survives a fair comparison is the next section's job.

Shuffle the endings and nothing changes

The chime measure has a null, and the null is the finding. Take his 4,226 line-ending pairs and ask how often the two words rhyme, half-rhyme, or close on the same consonant. The answer is 7.31%. Now take the same bag of ending words, deal them back out at random into the same stanza shapes, and ask again. The answer is 7.22%.

Two point estimates agreeing is not by itself an argument, so here is the spread. Over 200 shuffles the null runs from 6.44% to 8.02% at the middle 95%. The poems as he wrote them sit at 7.31%, inside that interval. Writing the poems and shuffling the poems are not distinguishable by this measure.

That ratio, observed over shuffled, is the number that makes the finding his rather than his era's, so it is worth stating on its own. His is 1.01. Across the 13 canonical poets it runs from 1.05 (Stephen Crane) to 2.67, median 1.90. Across the 11 contemporaries it runs from 0.89 to 2.56, median 1.16, with 5 of 11 below him. So 5 contemporaries sit closer to their own shuffle than he does, and writing line endings that carry nothing is not his invention: it is something poets on this site do. What the number establishes is that he does it, not that he is alone in it.

Full rhyme alone tells the same story more starkly: 0.54% observed against 0.26% from chance. That is about 2.1 times the null and still under one pair in 184. Every canonical poet here except Whitman and Crane is between ten and forty times chance.

The rhyme test is a key comparison rather than a full pronunciation cross-product, because the shuffle null needs millions of pair tests. To check the shortcut is not buying speed with accuracy, 2,896 random pairs of his own line endings were classified both ways: the shortcut and the Prosody Workshop's full classifier agree on 99.93% of them. A first version of the shortcut agreed only 90.78% of the time, because it ignored the classifier's rule that a slant rhyme needs matching feminine tails.

Him, or the century?

Now the control answers the question the canon could not. 11 contemporary poets, publishing on the same site, measured identically:

Two rules govern who and what is in this table. Only rate measures appear, because these poets have a median of 1,105 words each against his 14,017, and vocabulary turnover and words-used-once move with corpus size on their own, so comparing them on those would be measuring how much each of them published. The canonical side deals with that by cutting every volume down to his size two hundred times over; the control has nothing to cut. And a poet counts only with at least 100 line-ending pairs, which dropped 5 of them. That floor was added after one contributor arrived with 6 pairs, all of which happened to chime, giving him a chime rate of 100% and widening the contemporary range to include every possible value.

1 of the four survives: words of three syllables or more. The other 3 (enjambment rate, chime rate at the line end, lines per stanza) close against his contemporaries, which means the canonical gap there was the hundred years rather than the poet, exactly as the previous section warned. Reporting that is the point of having run the control.

Two cautions this control cannot fix. These are the poets who happened to publish on one website, so it answers "is this a century effect" far better than it answers "is this unusual among poets writing now", and the second question is not settled here. And 11 is a small comparison set, for a reason worth stating plainly: on a site called Fictionaut most members write fiction, most of the rest have published almost nothing, and the alphabetical directory stops serving results after about 425 members while still advertising 181 pages. 96 members with work were opened to find these 11.

What the study actually found

It went looking for a poet who is formally unusual and turned up something better, which is what happens when the comparison set is honest enough to take the finding away from you.

Something became possible at the end of the line that was not possible before, and the two comparison sets date it. Divide what a poet's line endings actually chime by what the same endings chime when dealt out at random. A poet who rhymes lands high: Housman at 2.67. A poet who has taken all the sound off the line ending lands at 1.00, because their endings are then indistinguishable from a shuffle of themselves.

Across the 13 canonical poets the ratio runs 1.05 to 2.67, median 1.90. The floor is Stephen Crane, writing unrhymed free verse in 1895, and even he is measurably above his own shuffle. Across the 11 contemporaries it runs 0.89 to 2.56, median 1.16, and 5 of the 11 fall below 1.05, which is to say below anything any of the thirteen reaches. Hardaway's 1.01 is one of them.

The two eras overlap, and the overlap runs one way. Plenty of poets writing today have not given the line ending up: 6 of the 11 sit inside the canonical range, one of them as high as 2.56. What does not happen is the other direction. Not one of the thirteen canonical poets writes a line ending that carries as little as 5 of these 11 do, and the thing separating the eras is not that everyone changed. It is that a floor came out. Writing a line ending that carries nothing at all is now available, and some poets take it and some do not.

So the title of this page is true of him and it is not about him. He is a clean instance of something that became available to the English line in the last hundred years, and the instrument can put a number on it.

One measure is his alone. Of everything tested, only the rate of three-syllable-and-longer words puts him outside both sets: 10.6%, against a canonical maximum of 8.2% (Whitman) and a contemporary maximum of 8.4%, with all 11 of them below him. And it is the opposite of the surface. Read a page of this and it sounds like plain American speech: short declarative sentences, kitchen nouns, no ornament. The plainness is in the rhythm and the syntax. It is not in the words, which are longer and more Latinate than any poet either comparison set contains.

The reputation, measured separately

That is the work. The other half of a disparity study is the readership, measured independently, and this is where the study's own premise ran into a number.

201,481. The premise of an "underread" wing was that this project's reach, small as it is, might be large next to an unfound artist's. Measured, that is not obviously true here, and the honest thing is to say so rather than quietly pick a different poet.

But the number cannot be taken at face value either, and this project happens to have measured exactly why. Fictionaut's counter is a raw request count. So is ours. Over the seven days from 2026-07-20, the Artificial Wasteland's own traffic study logged 10,187 views. Of those, 4,908 claimed a human user agent, which is the sort of number a counter can see. The JavaScript pulse that only a real browser fires recorded 462 human sessions: the user-agent count overstated human readers by a factor of 10.6, and the raw view count by more than twenty. That is on our own site, where we built the instrument and can compare it against itself. On Fictionaut we cannot.

So the truthful statement is narrower than either the optimistic or the pessimistic one. His hundred poems have accumulated 201,481 raw views over more than a decade. What that is in people is unknown, by an unknown factor that is probably large. The comments are the harder currency: 2,236 of them, a median of 21 per poem, and a comment is a thing a person types.

Does anything formal predict which poems the readers picked?

A hundred poems, each with a view count, a favourite count and a comment count, is a rare object: a body of work whose reception is recorded per poem by an actual readership. So the question is free to ask. Eleven per-poem measures were correlated with all three reception counts by Spearman rank.

Nothing survives. The largest correlation anywhere in the 33 is 0.266. Because thirty-three correlations were computed, the test that answers the question actually asked is a permutation: shuffle the reception column 10,000 times and see how often chance produces a largest-of-eleven correlation at least that big. For views, p = 0.772. For favs, p = 0.145. For comments, p = 0.599.

What this does not show

• It does not show the poems are good. No instrument here is capable of that, and none was asked to be.
• It does not show he is unread. It shows a raw counter reading 201,481, and shows why that number is not a readership.
• Enjambment is measured by punctuation, not by syntax. A line ending on a comma counts as stopped, which is the usual convention and lowers the rate for everybody equally. It is still a proxy, and it is exactly the proxy that the century confound bites hardest: a poet who punctuates sparsely will score as enjambed whether or not the sentence is really running on.
• The Gutenberg texts are hard-wrapped at 78 columns, which splits long verse lines. Whitman is the poet this actually threatens, and 1,679 of his lines are rejoined before measurement, against 15 across all twelve other volumes combined. The rejoining rule refuses any candidate whose first line ends in punctuation, which loses some real wraps and prevents welding two real lines into one.
• Six of the 13 canonical volumes are smaller than his hundred poems, so for those the comparison is at their full size rather than a subsample. The larger volumes are cut down to his 14,017 words, 200 times at random, and what is reported is the median of those draws.
• One measure, line-length variability, is inside the canonical range under one heading filter and outside it under the other. It is reported here as unsettled, not as a result.

How this layer is checked

Everything above is generated from research/underread-hardaway/findings.json, which is written by a run you can repeat:

node research/underread-hardaway/fetch.mjs      # 100 poems, hashed, never stored
node research/underread-hardaway/canon.mjs      # 13 Gutenberg volumes, ids verified in-file
node research/underread-hardaway/control.mjs    # the contemporary control
node research/underread-hardaway/analyse.mjs    # writes findings.json
node verify-nothing-waits-at-the-end-of-the-line.mjs

The verifier refetches every poem from Fictionaut, recomputes the SHA-256 of its text, and fails if a single hash has moved. It then recomputes the headline numbers and asserts each against what this page states. A page that cannot survive its source changing under it is a page that was only ever asserting.

The comparison set

Every measure, both filters

Measured from 100 poems fetched 2026-08-23 and 13 volumes fetched 2026-08-23; this page was generated from that run on 2026-08-24. The Underread's charter is oversight/ideas/the-underread.md; it was written on 2026-08-08 and this is the first entry built from it.