Appendix E: Skeptic’s Appendix — Objections, Alternatives, and Failure Modes
We present the strongest objections we know of to the empirical and conceptual claims of Geometric Aesthetics, and the honest responses we can give. This appendix is written in the spirit of adversarial collaboration: the strongest version of the framework is the one that survives the strongest attacks. Each objection is given two to three paragraphs; none is made easier than it deserves.
E.1 “The Encoder Captures Genre, Not Aesthetics”
This is the most important objection, and the one we have most thoroughly confronted. LaBSE and MERT are powerful genre/style classifiers: the same features that produced a headline R = 0.241 across the 4,998-book corpus also recover fiction-vs-non-fiction, Victorian-vs-Modern, and Romance-vs-Mystery with high accuracy. Concerningly, when we residualized against genre labels, 85% of the book R² vanished. When we residualized against genre labels in the music corpus, 91% of the hand-feature R² vanished. For a skeptic who arrived at this book expecting a bold “geometry predicts aesthetic quality” headline, these are damning numbers, and we do not dispute them.
Our response is threefold. First, we report the residualized result as the result, not as a robustness check. Within-genre, author-disjoint, the books signal is R = 0.093, R² = 0.009, z = 6.5σ, p = 5.7×10⁻¹¹ on n = 4,998 — small, but replicable and statistically strong at this sample size. Within-fiction, R = 0.131 at 6.2σ on n = 2,250. Within-Rock, within-Electronic, within-Hip-Hop, within-Pop all carry signal in the music corpus (R = 0.14–0.19, artist-disjoint). The “additional aesthetic signal beyond genre” framing is the honest one, and it is the one this book adopts. We do not claim that geometry predicts rating; we claim that geometry carries a small residual predictive signal after genre is controlled for, and that the shape of that signal differs between modalities.
Second, the confound is itself informative. That 85–91% of the headline is genre means that the LaBSE/MERT features are picking up real stylistic structure — the same structure humans use to say “this is a Victorian novel” or “this is a driving Rock track”. Genre is not a nuisance variable to be purged; it is part of the aesthetic content of a work, and the geometry is how the encoders represent that content. We therefore think of genre as a coarse partition of the manifold, not as noise. Third, and most importantly, the strongest finding of the book — cross-lingual invariance at ρ ≈ 0.7 across six language families — is independent of any rating target. It is a statement about the geometric feature vector itself being preserved under language swap, not about predicting a human judgment. That finding is not vulnerable to the genre-confound objection, because it makes no predictive claim at all.
E.2 “Cross-Lingual Invariance Is Trivial Because LaBSE Was Trained For Translation Alignment”
This is the most tempting objection to dismiss, and the one we take most seriously. LaBSE’s training objective explicitly pushes translated sentence pairs toward proximity in the embedding space. It would be bizarre if, having trained the encoder for exactly this purpose, we then discovered that translations of the same book produced dissimilar embeddings. In one sense, the cross-lingual invariance finding is simply confirmation that LaBSE works as advertised.
The objection is partially right, and we acknowledge it. But it misses what is specifically being measured in Chapter 17. We do not measure sentence-pair similarity — that is indeed what LaBSE was trained for. We measure point-cloud statistics: covariance structure, internal coherence averaged over pairs within a work, trajectory curvature, recurrence rate, step-size distributions. These are second-and-higher-order geometric properties of the cloud of paragraphs making up an entire book. They were never part of LaBSE’s training signal. LaBSE sees sentence pairs; it does not see which sentences belong to the same long work, and it certainly does not see any direct supervision on “the covariance of the paragraph cloud of a book should match across translations”. That these point-cloud statistics emerge as language-invariant — with Spearman ρ ≈ 0.7 across 6 language families and within-bundle-std ratios as low as 0.18 for path_eff — is a property of the encoder’s induced geometry, not of its training objective.
An analogy: word2vec is trained on context prediction, yet exhibits linear compositional structure (king - man + woman ≈ queen) that was not in its training objective. No one claims this linear structure is “trivial” because word2vec was trained on context. The same logic applies here. That a translation-trained encoder preserves point-cloud geometry across translations is an emergent property, and emergent properties are exactly where interesting empirical work lives. We mark the result as load-bearing, not as circular.
E.3 “Low R² Means No Real Signal”
A reader coming from domains where R² = 0.5 is considered a weak result will look at our within-genre R² = 0.009 and conclude that the entire literary-geometry program is a rounding error dressed up in physics language. This is a fair instinct, and one we must answer on the instinct’s own terms.
The answer is in two parts. First, what the effect-size conventions are calibrated for. R² conventions from economics and psychology typically assume effect sizes where a single intervention explains a substantial fraction of outcome variance — experimental interventions, drug trials, strong manipulations. Goodreads ratings of books are the aggregate verdict of thousands of readers, influenced by genre preference, period, personal taste, reading mood, marketing, peer recommendations, translation quality, and countless other factors. The ceiling on any single feature family — even a perfect feature family — is low, because so much of the variance is genuinely idiosyncratic to individual readers and to non-textual factors. Against that ceiling, an R² = 0.009 that nonetheless produces z = 6.5σ at p = 5.7×10⁻¹¹ on n = 4,998, replicable across author-disjoint folds, is not “no real signal”. It is “a small but reliably measurable signal in a noisy domain”.
Second, and more importantly, we are not asking R² to do the work the framework asks it to do. The book’s central claims are not “feature X predicts rating Y with R² = Z”. They are claims about geometric structure: that the manifold has directionality (sign flips between modalities at p < 10⁻²⁸), that it supports a metric that is cross-lingually invariant (ρ ≈ 0.7 across language families), that genres appear as discoverable Lasso axes (71 non-zero axes in the books spectrum, of which many are post-hoc interpretable). These claims are tested by statistics other than predictive R². A framework can be predictively weak on a noisy rating target and simultaneously make strong structural claims that hold up at p = 10⁻⁵⁷. The skeptic who collapses “does this predict stars?” into “is this a real theory?” has imported a reductive version of the aesthetic question we have spent the whole book arguing against.
E.4 “The Cross-Modality Sign Flip Could Be a Bug”
The sign flip on pair_sim_mean (books ρ = +0.126, music ρ = −0.076) and on step_mean (books ρ = −0.096, music ρ = +0.071) is, in the argument of Chapter 21, the core surprise of the book. A natural skeptical response is that this flip could be an artifact of preprocessing — different encoders, different temporal scales, a bug in a normalization step, perhaps a Hellinger feature computed one way for books and another way for music.
We verified the flip post-Hellinger-fix (the Hellinger feature was saturating in the music corpus due to small-sample Gaussian fits in 1024-d; we moved to a K=32 PCA subspace and the Bhattacharyya feature directly; see the methodological note in the book reference). With the fix in place, the sign flip persists and intensifies: pair_sim_mean on music sits at ρ = −0.076, p = 5×10⁻³³; step_mean on music sits at ρ = +0.071, p = 4×10⁻²⁹. These p-values come from n = 24,801 tracks with artist-disjoint CV. For the flip to be a bug, the bug would have to be corpus-wide, consistent across thousands of tracks, and produce effect sizes that miraculously match what one would predict from the interpretation “books reward coherence and continuity, music rewards contrast and dynamic variation”. That is a very specific bug.
We do not rule out the possibility that some encoder-level difference between LaBSE (trained on sentence pairs) and MERT (trained on 160k hours of audio SSL) is structurally responsible for the sign difference, rather than a fact about aesthetic preferences as such. That is a legitimate alternative interpretation, and we flag it in Chapter 21. Distinguishing “encoder-artifact sign flip” from “genuine modality-specific aesthetic directionality” would require either a third encoder in each modality (does the flip replicate with MusicBERT-style vs MERT? with RoBERTa vs LaBSE?) or a cross-modality controlled study where the same construct — say, “dynamic contrast within a work” — is measured in a way less entangled with pretraining. We name this as an open problem in Chapter 29. But the robustness of the flip at p = 10⁻²⁹ and 10⁻³³ is not a bug; if it is an artifact, it is an artifact with a shape that is interpretable and stable.
E.5 “Ratings Are Not Aesthetics”
Goodreads stars and Spotify-or-FMA listen counts are not aesthetic judgments in any thick philosophical sense. They are proxies — commercial, social, noisy, shaped by the recommendation systems that serve the pages on which ratings are registered. A serious aesthetic theory that relies on fitting “is this beautiful?” with “did this book average 4.1 stars on Goodreads?” is, a skeptic might say, mistaking the appraisal for the object.
We agree, and we say so explicitly. Chapters 2, 9, and 15 repeatedly note that rating is a scalar collapse of an object whose structure we argue is geometric. The within-genre R² = 0.009 is, if anything, confirmation that ratings are a blunt instrument: the geometry picks up a small real signal, and the rest of the variance is the ocean of non-aesthetic noise. We do not claim that our feature space predicts aesthetic merit. We claim that it carries a measurable signal in the direction of “things humans-in-aggregate mark with higher scores”, which is a weaker and more honest claim.
The cross-lingual invariance finding is, as noted in E.1, the part of the empirical program that does not depend on a rating target. It is a structural claim about the encoder-induced geometry of works, not a predictive claim about a noisy proxy. For readers who reject Goodreads and listens as valid aesthetic targets altogether, the cross-lingual invariance remains intact, and the sign-flip between modalities (being a within-feature cross-modality comparison) likewise stands. Much of the book’s load-bearing structure is therefore robust to the “ratings are not aesthetics” objection even in its strongest form. The rating results are secondary corroboration, not the core.
E.6 “The Framework Privileges Digital / Embedded Artifacts Over Live or Performative Ones”
The book’s empirical program relies on encoders that operate on digitized texts and on recorded audio. Live performance, dance, theater-as-event, oral poetry, and site-specific installation art are not obviously amenable to the same pipeline. A skeptic aligned with performance studies might argue that by defining the aesthetic manifold through whatever our encoders can embed, we have tacitly defined a restricted aesthetics — the aesthetics of the embedded artifact — and then generalized it into a theory of “aesthetics” as such.
This is a fair critique and we address it head-on in Chapter 16 on aesthetic uncertainty. The geometric framework as currently instrumented is a framework for encodable artifacts. A novel can be encoded; a theatrical performance can only be encoded via a proxy (a recording, a script, a photograph), and the proxy is not the performance. We are cautious about claiming that our feature space captures the aesthetics of live performance, improvised music, or site-specific installation. Where we do extend the framework to film (Chapter 22), to visual art (Chapter 23), and to architecture (Chapter 24), we do so by analogy with the book and music cases and explicitly flag those extensions as conjectural (see Appendix F).
The methodological upshot is that “geometric aesthetics” is well-defined for the class of artifacts whose structure survives an encoder. That class is large — it contains almost all published fiction, recorded music, digitized visual art, and (via text-to-image multimodal encoders) a good deal of image-based art. It does not contain the live event. We do not claim it does. Chapter 16’s working position is that the encodable artifact is the tractable base case, and the non-encodable live event is a boundary problem whose aesthetics our framework partially illuminates (via the trace the event leaves in recordings, scores, and descriptions) but does not fully subsume.
E.7 “Western-Canon-Biased Corpora”
Project Gutenberg is overwhelmingly Western. The FMA Medium corpus skews heavily toward contemporary English-language Rock / Electronic / Hip-Hop / Pop and contains little from the South Asian, East Asian, African, Indigenous, or ethno-musicological traditions that make up the majority of humanity’s musical practice. When we report “cross-lingual invariance across 6 language families”, those six families are Germanic, Romance, Uralic, Hellenic, Italic-ancient, and Constructed. When we report the Sinitic gap — only 5 Chinese bundles — that is because the Chinese Gutenberg corpus is classical originals (Confucius, Sunzi, the Shijing), not Western-work translations, and our bundling protocol requires the same underlying work in multiple languages. The skeptic is right to note that our “cross-lingual invariance” result is, in its current form, a cross-European-language invariance result.
We acknowledge this openly in Chapter 29 as an open problem. The corpus gap is not a theoretical limit — nothing in the framework prevents extending bundles to Chinese, Japanese, Arabic, Hindi, Swahili, or Yoruba translations of the same works. It is a data-collection gap. The honest inference from our current numbers is: within the language and cultural range we sampled, the encoder-induced geometric structure of works is invariant to a degree that surprises us. It remains an open empirical question whether this invariance extends to non-Western canons, to orally-transmitted texts, or to non-European language families at the same ρ ≈ 0.7 level, at lower levels, or not at all.
A sharper version of the objection is that by choosing a Western-trained encoder (LaBSE, MERT) we have pre-committed to a Western-aesthetic representation, and any invariance we find is an invariance within a Western-conditioned representation space. This is philosophically serious. Our partial response is that LaBSE is trained on 109 languages including many non-European ones, and MERT is trained on 160k hours of globally-sourced music; neither is strictly Western-conditioned. But “not strictly Western” is weaker than “culturally neutral”, and we do not pretend otherwise. Chapter 29 commits us to replicating these findings on encoders and corpora that do not share the Western-canonical bias, and to updating the theoretical claims accordingly. The current claims are bounded by the current corpus, and we mark that boundary.
E.8 “The Framework Is Too Flexible — Unfalsifiable”
With four feature channels, 71 Lasso axes, five encoder variants, and a dozen residualization choices, the framework has enough degrees of freedom that a skeptic might reasonably suspect it can fit any aesthetic story ex post. This is the sharpest methodological objection, and we treat it as such.
The falsification criteria we commit to are concrete. (i) If cross-lingual invariance fails at language-pair level — ρ < 0.5 across multiple language families on a pre-registered bundled corpus at n ≥ 200 — then Chapter 12’s Noether-style symmetry claim is falsified. The current result at ρ ≈ 0.7 is from data, not from fit. (ii) If the cross-modality sign flip on pair_sim_mean and step_mean fails to replicate on a new music corpus with a different encoder (say, a MusicGen-derived or CLAP-derived embedding), the “books reward continuity, music rewards contrast” claim is weakened or overturned. (iii) If the within-genre residualized signal (R = 0.093 on books, R = 0.14–0.19 on within-genre music) fails to replicate on a new corpus, the “additional aesthetic signal beyond genre” claim is weakened. (iv) If any of the formal claims marked Proved in Appendix F are found to have errors in the proofs, they are retracted.
We note that the framework is falsifiable in practice, not merely in principle. Each of the tests above can be conducted with existing tools, public encoders, and existing datasets (Gutenberg, FMA, MSD, Goodreads). We invite replication and we commit to updating the book’s empirical claims in response to null or contradictory replications. A framework with a published falsification list that has passed four tests (cross-lingual invariance, sign flip, within-genre signal, independence from ratings) is not unfalsifiable — it is a theory that has, so far, survived contact with the data.
E.9 “The Whole Geometric Vocabulary Is Dressing”
Some skeptics will accept the empirical findings — cross-lingual invariance, sign flips, within-genre signal — while rejecting the geometric framing as ornament. On this view, we have a set of statistical results that could be reported in the flat vocabulary of feature engineering, and the appeals to manifolds, Riemannian metrics, Noether’s theorem, and quantum dynamics are rhetoric that inflates the interest of the findings beyond what they warrant.
We take this seriously as a stylistic and philosophical critique. Our response is that the geometric vocabulary does work that the feature-engineering vocabulary cannot do. Cross-lingual invariance is a symmetry-invariance claim; calling it that places it in a well-understood conceptual neighborhood (Noether’s theorem, gauge invariance) where its implications are structured, not ornamental. The sign flip is a directionality claim about the manifold, and distinguishing “direction” from “magnitude” is exactly what one does on a geometric object and exactly what one cannot do on a scalar. The Lasso axes are axes of a linear subspace of the feature space, and “axis” is geometric language used for its geometric content.
Where the geometric vocabulary becomes load-bearing beyond its literal applicability is in Chapter 13’s quantum aesthetic dynamics. There, as Appendix F records, we are using mathematical language for its expressive power rather than making a literal physical claim. The observer, the superposition, the measurement are mathematical analogies: they pick up structural features of aesthetic judgment (it is context-dependent, it collapses under articulation, prior judgments influence posterior ones) that the analogy illuminates. We do not claim that aesthetic experience is quantum-mechanical in any literal sense; we claim that the mathematical structure of quantum mechanics is a useful vocabulary for some features of aesthetic judgment. The skeptic is right to mark this as analogy; we mark it the same way ourselves. For the empirical chapters (17, 20, 21), the geometric vocabulary is not dressing but description.
E.10 Summary: What Would Make Us Abandon the Framework
A combination of three results would substantially undermine the book’s thesis. First, failure of cross-lingual invariance on a pre-registered replication at the language-family level. Second, failure of the cross-modality sign flip to replicate under a non-MERT music encoder. Third, failure of any within-genre residualized signal to replicate in any modality on a pre-registered new corpus. If all three fell, we would conclude that the signals we measured were specific to our encoders and datasets, and that the geometric framework did not point to anything beyond the representation space of specific pretrained models.
None of these three has fallen in our tests so far. Until they do, we maintain the claims in the form the book makes them — modest predictive R², robust structural invariances, honestly-flagged analogies, and a standing invitation to falsification.