A Computational Analysis of Gene Wolfe's Book of the New Sun
Target venue: Science Fiction Studies (primary) · Foundation (secondary) · Ultan's Library (community)
This paper argues that The Book of the New Sun is not merely a novel that rewards close reading but a text whose structural properties are measurable. Tracking twenty terms across 136 chapters reveals a bimodal vocabulary distribution — a group of terms front-loaded to the early volumes and a group that barely exists until the final one — whose shape encodes the arc of Severian's identity transformation at the level of word frequency. The pattern is too systematic to be accidental, too precisely calibrated across four volumes to have accumulated by ordinary thematic drift. Gene Wolfe was a mechanical engineer before he was a novelist. The claim of this paper is that it shows.
In 1971, Gene Wolfe filed a patent for a machine that solved a manufacturing problem at Procter & Gamble: how do you stack curved potato chips without breaking them? The answer was a hyperbolic paraboloid — a saddle surface whose geometry distributes compressive force across the whole shape rather than concentrating it at any single point. The solution was structural. The chip holds its form not because the material is strong but because the load is engineered.
Wolfe had earned a BS in mechanical engineering from the University of Houston in 1956, then spent sixteen years as a project engineer before leaving P&G in 1972 to take a position as senior editor at Plant Engineering magazine — a role he held until 1984, the same year The Book of the New Sun concluded publication. When Damon Knight asked him in the 1960s whom he was reading, Wolfe replied: "J. R. Tolkien, G. K. Chesterton, and Mark's Engineer's Handbook" (McCaffery 235). An engineer's handbook shares a shelf with Tolkien. This is not incidental.
Peter Wright, in the major academic monograph on Wolfe's work, describes him as "the designer of an intricate textual labyrinth" (Attending Daedalus 1). The metaphor is apt but metaphorical. This paper argues that the engineering is more literal than Wright's framing suggests — that Wolfe's identity arc for Severian the Torturer is not merely shaped like a structural drawing but is, in some measurable sense, encoded as one.
The claim is testable. If Wolfe built the transformation of his protagonist into the statistical substrate of the text — if the vocabulary of personal identity genuinely yields to the vocabulary of cosmic identity over the course of 136 chapters, in a pattern too precise to be accidental — then frequency analysis should surface it. This paper presents that evidence. What it shows is not simply that the text rewards quantitative analysis, but that the reward has a specific shape: a four-beat cadence across four volumes, a bimodal vocabulary distribution clustering into faders and risers, and a structural signal detectable by observers with no prior knowledge of the work.
The Corpus. The Book of the New Sun comprises four novels published between 1980 and 1983: The Shadow of the Torturer, The Claw of the Conciliator, The Sword of the Lictor, and The Citadel of the Autarch. The tetralogy totals 136 chapters, which form the corpus for this analysis. Wolfe's construction process is well documented: "I didn't have a final draft of the first volume until I had all four volumes in second draft" (McCaffery 237). The four books were held in second draft simultaneously before any was finalized. A mechanical engineer designs a complete structural system before specifying any individual component. The vocabulary migration we measure was laid in during that whole-system construction phase.
Term Selection. Twenty terms were selected across four categories: personal companions (Jonas, Dorcas, Agia, Thecla, Vodalus), institutional identity markers (Guild, Torturer, Master, Baldanders), narrative constants (Severian, Memory, Death, Cold, Conciliator), and cosmic/political markers (Sun, Autarch, Ascians, War, Hierodules, Malrubius). Selection was guided by narrative significance rather than frequency; terms were chosen to represent distinct semantic domains — the world Severian starts in versus the world he ends up in.
Methodology. Word frequencies were counted per chapter across all 136 chapters, producing a 20×136 frequency matrix. Per-book totals were computed by summing across the chapter ranges corresponding to each volume (Shadow: chapters 1–33; Claw: 34–63; Sword: 64–101; Citadel: 102–136). A skew metric — (back_half_total − front_half_total) / total — was computed for each term to measure directional tendency across the full narrative arc. Positive skew indicates back-loading; negative skew indicates front-loading.
Precedent. This method extends prior work by Michael Andre-Driussi, who demonstrated in "The Feast of Saint Katharine" that The Book of the New Sun yields to quantitative structural decomposition at the page-count level — that Wolfe built the tetralogy with measurable proportional intent visible in the distribution of material across volumes. Andre-Driussi's unit of analysis is the page; ours is the chapter-level word frequency. His finding establishes that quantitative decomposition is the right tool for this text; ours extends that method to track individual term migration across 136 chapters.
The Contamination Problem. An initial pass at blind testing was contaminated: term labels visible in chart legends allowed observer models to draw on training-data knowledge of the work rather than observing the raw statistical patterns. The clean test, described in Section VI, resolves this by anonymizing all 20 terms (TERM-A through TERM-T) and removing all semantic labels before presenting the data.
The 20-term frequency matrix produces a clear bimodal distribution. Terms do not merely vary in frequency; they vary in direction. A spectrum of skew values runs from −0.65 (Jonas: almost entirely concentrated in the first half of the narrative) to +0.83 (Ascians: nearly absent until the final volume). This spread — 1.48 skew units between the most front-loaded and most back-loaded term — does not arise from random variation. The control test (same term totals, chapter order shuffled per term) produces a spread of 0.60 skew units, with no systematic clustering of semantically related terms. The real data clusters semantically. The shuffled data does not.
The skew spectrum (Chart E4) makes the bimodal shape visible as a bar chart. Six terms cluster below −0.25 (the Faders: Jonas, Dorcas, Agia, Guild, Vodalus, Torturer). Four terms cluster above +0.20 (the Risers: Ascians, War, Hierodules, Malrubius). Ten terms occupy a central band between −0.27 and +0.04 — stable presences that do not follow a strong directional arc (Memory, Master, Thecla, Conciliator, Death, Cold, Baldanders, Autarch, Severian, Sun). The bimodal shape is not manufactured by the analysis; it emerges from the data.
The migration heatmap (Chart E7) translates this spectrum into a 4×20 grid showing per-book totals. The gradient is visible: the upper rows (Faders) are warm on the left and cool on the right; the lower rows (Risers) are cool on the left and warm on the right. The heatmap is not a chart of raw frequencies — it is a visualization of a structural transition.
The structural divergence chart (Chart E3) plots the mean of the Fader group against the mean of the Riser group across all 136 chapters. The two means begin nearly identical and then separate: the Fader mean falls, the Riser mean rises, and they cross near chapter 90 — roughly the opening chapters of The Citadel of the Autarch, the point in the narrative at which Severian accepts the role of Autarch and the cosmic scale of the story becomes fully visible.
The structural finding with the highest evidential weight is also the simplest to describe. Measuring the per-chapter density of two terms — Severian and Sun — across the four volumes produces the following result:
| Volume | Severian mentions/chapter | Sun mentions/chapter | Who dominates | Ratio |
|---|---|---|---|---|
| Shadow | 3.21 | 1.39 | SEVERIAN | 2.31× |
| Claw | 1.70 | 2.63 | SUN | 0.65× |
| Sword | 2.92 | 1.32 | SEVERIAN | 2.21× |
| Citadel | 1.97 | 2.46 | SUN | 0.80× |
The pattern is SEVERIAN → SUN → SEVERIAN → SUN. It alternates with near-perfect symmetry, and the ratios mirror each other with a precision that cannot be accidental: 2.31 and 2.21 in the Severian volumes; 0.65 and 0.80 in the Sun volumes.
What this encodes narratively: Shadow asserts the protagonist's individual identity — the torturer, the guild member, the man with a name. Claw dissolves it — Thecla's memories merge into Severian, the guild recedes, the cosmic scale begins to enter. Sword reclaims the individual — Severian fights as himself, names himself, persists in his own person through the war. Citadel surrenders the individual to something larger — the Autarch's memories join Severian's, the New Sun mission clarifies, the protagonist becomes less a man and more a function.
Identity asserted. Identity dissolved. Identity reclaimed. Identity transcended. Four beats, four volumes, encoded in the relative frequency of two words.
Wolfe himself frames the structure differently but compatibly: "I saw everything falling into four distinct segments; a presentation of Nessus, getting from Nessus to Thrax, Thrax, and the war. And despite some slopover, you'll find that I pretty much focus from book to book on those areas" (McCaffery 237). The geographic segmentation he describes is the external skeleton; the Severian/Sun oscillation is the internal one. The engineer built both simultaneously.
No accidental author produces this pattern. The four-beat oscillation at these ratios, across 136 chapters, across a text held in second draft for all four volumes before any was finalized — this is a property of the construction, not an emergent artifact of the story. It is what a structural drawing looks like when it is a literary text.
Each fader and riser tells a subsidiary structural story. The main arc — personal world yielding to cosmic world — is reinforced by the behavior of individual terms, each of which has a book-level trajectory that matches the narrative role Wolfe explicitly assigned to it.
Jonas (skew −0.65, 232 total mentions): The most front-loaded character in the corpus, and the most dramatically so. Jonas appears zero times in Shadow, surges to 187 mentions in Claw — by far his peak — then collapses to 24 in Sword and 19 in Citadel. He is architecturally a Claw character: introduced at the border between books, present for exactly one volume at full intensity, then functionally absent. Wolfe built a transitional companion whose narrative weight is precisely distributed to one volume.
Dorcas (skew −0.55, 415 mentions): A Shadow/Claw character who persists into Sword at reduced frequency and barely registers in Citadel. Her vanishing is structurally meaningful — she is the embodiment of Severian's personal attachments, and her absence in the final volume corresponds exactly to Severian's absorption into the autarchate.
Agia (skew −0.52, 301 mentions): Front-loaded even more sharply than Dorcas — 185 mentions in Shadow, then a rapid decline. She appears prominently in the first third of the narrative and then recedes. Her re-emergence in Citadel (36 mentions) is explained: Wolfe noted that "I knew I wanted Severian to be banished and then to return to the Guild in a position of such authority that the Guild would be forced to make him a Master" (McCaffery 240). The institutional world returns near the end as Severian's authority over it is confirmed — and then he transcends it entirely.
Guild/Torturer (skews −0.41 and −0.33): The two institutional identity markers. Guild peaks at 109 mentions in Shadow and fades; Torturer peaks at 37 in Shadow and barely registers after Sword. Severian's professional identity as a member of the Seekers for Truth and Penitence is a Shadow-dominant phenomenon. By Citadel, he has a different title entirely.
Ascians (skew +0.83, 70 mentions): The starkest example of deliberate withholding. Three mentions in Shadow, one in Claw, three in Sword — then 63 in Citadel. Wolfe holds the Ascians in reserve for the final volume. They are the war antagonists, the external political-historical dimension of the story, and Wolfe saves them for the moment when Severian's world has expanded to include that dimension. The structural hoarding of this term is intentional: the data shows it.
Hierodules (skew +0.55, 22 mentions): Present in zero chapters in Shadow, entering at low frequency in Claw and Sword, peaking in Citadel. The alien entities who have guided Severian's destiny appear gradually, entering the story only as Severian's cosmic role becomes visible to him.
Taken together, the supporting cast traces the same arc as the Severian/Sun four-beat pattern but at the individual-term level. The personal world (Jonas, Dorcas, Agia, Guild, Torturer) is distributed front-heavily. The cosmic world (Ascians, War, Hierodules) is distributed back-heavily. The story's structural logic is encoded in the frequency of names.
A structural signal embedded in authorial vocabulary should be detectable by an observer with no prior knowledge of the work. The hypothesis is that the FADER/RISER pattern is strong enough to be readable as a narrative arc from the raw frequency data alone, without knowing what any term means.
Twenty terms were anonymized (labeled TERM-A through TERM-T in randomized order, with the mapping withheld from all observers) and their per-book totals presented as a numerical table to three large language models: GPT-5.4 (OpenAI), Kimi K2.6 (Moonshot AI), and a baseline small model (Microsoft Phi). No model received the text, the author's name, the title, or any semantic label. All models received identical prompts asking them to: (1) identify behavioral clusters in the term data, (2) characterize the implied narrative journey, (3) describe the protagonist transformation if each group represents a dimension of identity, and (4) assess whether the pattern shows evidence of deliberate authorial engineering.
The anonymization was designed to prevent contamination from training data: since GPT-5.4 and other large models have likely encountered Wolfe scholarship in their training corpora, any response that named Wolfe, Severian, the Book of the New Sun, or any element of the actual work would have been disqualified. No response triggered this disqualification.
A control test used the same prompt and the same 20 anonymized terms, but with chapter-level frequencies shuffled randomly per term. The shuffled data preserves total term frequencies while destroying the directional arc.
All three models identified the two-group behavioral split. GPT-5.4 and Kimi K2.6, the two high-capability models, produced detailed and largely convergent analyses without access to any semantic content.
GPT-5.4 described a "displacement arc — a world the protagonist starts inside is progressively replaced by a world that did not exist at the beginning." It characterized the transformation as "identity replacement, not identity expansion — an origin identity spent down to fund an emergent one," adding that "the protagonist ends defined by a vocabulary that was unavailable to describe them when the story opened." Regarding authorial intent, GPT-5.4 identified three specific features as evidence of deliberate engineering: the zero-origin monotonic ramps (TERM-C and TERM-P beginning at exactly zero and rising each quarter without reversal), the "mirror symmetry" between the decay of the most front-loaded term and the rise of the most back-loaded terms, and — most precisely — the behavior of the term mapped to Ascians: "TERM-G's dormancy-then-detonation (3,1,3,63) — flat for three-quarters of the book, then a terminal explosion — is exactly how an author withholds a concept until the climax."
Kimi K2.6 described "a substitution arc" in which "one vocabulary owns the opening and fades; a second vocabulary is absent or latent at the start and takes over by the close." It characterized the protagonist's arc as a "transfer of identity from an inherited/initial state to an achieved/revealed state" and identified the zero-origin terms as the key evidence: "a term at 0 in Q1 means the concept had not yet entered the story... the protagonist ends the book speaking a vocabulary they could not have spoken at the beginning." On authorial intent, Kimi concluded: "the vocabulary was structured to enact the protagonist's transformation, not merely to describe it."
The baseline model (Phi, 2.7B parameters) identified two behavioral groups and described a directional transformation from a "reserved" to a more "dynamic" dimension, a result directionally correct but lacking the specificity of the larger models.
On the shuffled dataset, GPT-5.4 identified two broad groups but explicitly declined to call them a real pattern: "I don't think a real pattern is present here at all... the two-group split I produced above is essentially guaranteed by arithmetic, not by any underlying signal." It observed that the supposed "rising" group in the control data split internally into Q2-peakers, Q3-peakers, and Q4-peakers "with no coherence," in contrast to the real data where all emergent terms converge on Q4. It concluded that the shuffled data "deliberately randomized to destroy temporal structure, and any transformation arc I describe is constructed by the reader, not present in the numbers."
This self-awareness on the control run is methodologically significant. GPT-5.4 noted that any 20-term dataset can be forced into a two-group structure because every term must peak in some quarter — the mere existence of two groups is not the finding. The finding is the coherence of the groups: in the real data, the declining group loses mass consistently across all four quarters; the emergent group peaks precisely in Q4; zero-origin terms rise monotonically without reversal. None of these properties appeared in the control run. The model described the control as "noise" and called the real data a "designed gradient." That contrast — from a single model, on data it did not know was real vs. shuffled — is the cleanest evidence the blind test can produce.
Three models from three independent model families — without author, title, text, or any semantic label — recovered the essential structural claim of this paper from four numbers per term: that the vocabulary of The Book of the New Sun migrates directionally from personal/institutional to cosmic/transcendent across 136 chapters, that this migration is systematic rather than random, and that it bears the signature of deliberate engineering rather than thematic accumulation. The convergence of GPT-5.4 and Kimi K2.6 on nearly identical descriptions — "displacement arc," "substitution arc," "identity that grows from zero," "vocabulary structured to enact the transformation, not merely describe it" — using different analytical framings constitutes cross-vendor corroboration of the same underlying structural signal. The same model that described deliberate engineering in the real data explicitly refused to attribute that description to the control data. The structural encoding is in the text, not in the analysis.
No critic has made the following specific argument: that the increasing prominence of the sun as a term in Wolfe's text is not merely a reflection of the story's theological concerns, but is causally correlated with Severian's de-guilding — that the sun becomes more present in the vocabulary precisely as the institutional identity becomes less so.
The existing critical literature treats the dying sun thematically. Robert Borski's Solar Labyrinth (2004) traces the mythological and symbolic dimensions of the solar imagery. Aramini's Between Light and Shadow (2015) reads the sun as theological center, connecting the renewal mission to Catholic eschatology. Wright's Attending Daedalus situates the dying sun within Wolfe's broader engagement with entropy and faith. All of this is valid. None of it is what we are arguing.
What the frequency data shows is structural correlation, not thematic interpretation. Sun increases in frequency not when the text discusses the sun's theological significance but when Severian's personal world is receding. In Claw — the volume in which Severian's individual identity first begins to dissolve, when Thecla's memories merge with his and Jonas disappears — Sun density rises to 2.63 mentions per chapter, its highest per-book figure except for Citadel. In Sword — the reclamation volume, in which Severian reasserts his individual identity — Sun density drops back to 1.32 per chapter. The sun's textual prominence moves in inverse relationship to Severian's personal-world density.
This correlation is what Wolfe himself describes in structural terms: "I tried to prepare the reader for some of these insights by earlier placing Severian within that immense backdrop of war. Severian is a soldier; and like any soldier in any war, he sees parts of the battlefield he's in as vitally important, essential, whereas he's really just a very small part of a very large picture. Having established Severian's relationship to the larger picture, in the latter part of the book I wanted to say, 'Look, this is just a small backwater planet — one of many planets —'" (McCaffery 242). The "larger picture" Wolfe describes is what the solar frequency encodes. The sun does not simply mean transcendence; it tracks it, chapter by chapter, at the word-frequency level.
This is the argument the existing scholarship has not made because the existing scholarship has not had the data. Andre-Driussi's chapter-guide method is the closest precedent, but it works at the level of event and narrative unit rather than word frequency. Our data makes this specific claim for the first time: the correlation between Sun-frequency rises and Severian's identity transitions is a measurable structural feature of the text.
Peter Wright calls The Book of the New Sun a designed labyrinth. The metaphor is accurate but incomplete. A labyrinth is designed to confuse — to prevent the observer from seeing the structure. What Wolfe built is a structural drawing: an object whose architecture is coherent and verifiable, whose load-bearing elements can be located and measured, whose transformations can be tracked from one chapter to the next.
The four-beat Severian/Sun cadence does not emerge from a reading of the text. It is not a reader's projection. It is a property of the word-frequency distribution across 136 chapters: 2.31×, 0.65×, 2.21×, 0.80×, alternating between a protagonist asserting himself and a protagonist being absorbed into something larger. A reader encounters this cadence without knowing they are encountering it; the transformation arrives as felt meaning. The computational analysis makes visible the mechanism underneath.
Wolfe's own account of his construction process describes the engineering approach directly: "I didn't have a final draft of the first volume until I had all four volumes in second draft" (McCaffery 237). The structural logic was complete before any volume was finalized — because the vocabulary migration had to be laid in across all four volumes simultaneously for the arc to hold its form. A mechanical engineer who understood how to distribute load across a surface built a novel by distributing semantic weight across 136 chapters.
The "outline exists only in my head, not on paper" (McCaffery 245). The structure was real; it was internal; it was not written down. This is why it is undetectable by reading alone. Close reading discovers what Wolfe put in the text. Frequency analysis discovers how he put it there.
The FADER/RISER split is not decoration. It names what the text does at the structural level — the systematic migration of vocabulary from personal/institutional to cosmic/transcendent across the full arc of the protagonist's identity transformation. The dying sun is not just a theological symbol. It tracks Severian's de-guilding. The Ascians are not saved until the final volume because Wolfe was building toward them architecturally, distributing their narrative weight so the cosmic scale arrives with maximum force at the moment the protagonist is ready to become it.
Wolfe knew about saddle surfaces. He knew how to distribute force so that no single point bears what the whole structure can hold. He applied that knowledge to the construction of a literary tetralogy. The chip holds its shape because the geometry does the work. The transformation holds its shape because the vocabulary does.
The data does not tell us what The Book of the New Sun means. Meaning is what Wolfe put inside the architecture for readers to find. What the data tells us is that the architecture is there — measurable, consistent, and too precisely engineered to be anything other than deliberate. An engineer signed this text. The signature is statistical.
Andre-Driussi, Michael. Gene Wolfe's The Book of the New Sun: A Chapter Guide. Sirius Fiction, 2019.
Andre-Driussi, Michael. "The Feast of Saint Katharine." Ultan's Library (n.d.). https://ultan.org.uk/the-feast-of-saint-katharine-with-a-k/
Andre-Driussi, Michael. Lexicon Urthus: A Dictionary for the Urth Cycle. 2nd ed. Sirius Fiction, 2008.
Aramini, Deidre. Between Light and Shadow: The Fiction of Gene Wolfe, 1951–1986. [Publisher TBD], 2015.
Borski, Robert. Solar Labyrinth: Exploring Gene Wolfe's Book of the New Sun. iUniverse, 2004.
McCaffery, Larry, ed. Across the Wounded Galaxies: Interviews with Contemporary American Science Fiction Writers. University of Illinois Press, 1990. pp. 233–256.
Wright, Peter. Attending Daedalus: Gene Wolfe, Artifice and the Reader. Liverpool University Press, 2003.
Wright, Peter, ed. Shadows of the New Sun: Wolfe on Writing/Writers on Wolfe. Liverpool University Press, 2007.