Chapter 29 — Open Problems

RUNNING EXAMPLE — Priya’s Ledger

Priya is closing the laboratory notebook. On the left margin she has listed what the geometric framework has done: a seventeen-sigma effect in books; a cross-lingual invariance of ρ≈0.70 across ten language families; a twenty-eight-sigma signal in music after genre is residualized out; a pair of sign flips, robust at p < 10⁻²⁸, that separate literary coherence from musical contrast. On the right margin she has started a second column. This one is longer. It is titled “What we have not done.” Some of the entries are empirical gaps we could plausibly close in a semester of compute. Some are theoretical openings that will outlive this volume. One — the Sinitic corpus gap — is a reminder that the world’s aesthetic cultures are not uniformly legible to the instruments we have built. Priya closes the notebook with the right-hand column twice as long as the left. This is not a confession of failure. It is what the next decade of work looks like written in outline.

29.1 What We Do Not Know

The preceding twenty-eight chapters have developed a geometric framework for aesthetic judgment and tested it against two empirical corpora. The framework makes structural claims: that aesthetic judgment is multi-dimensional; that works occupy locations in a manifold rather than points on a line; that the tensor hierarchy exposes directional information a scalar discards; that aesthetic evaluation admits stratification, parallel transport, and cross-lingual invariance; and that modality-specific sign flips reveal a directionality that no universal aesthetic metric could contain.

Some of these claims now have empirical support. The internal-coherence channel pair_sim_mean registers at 8.4σ on books, distinct from divergence-family features (Chapter 17). Genre-residualized aesthetic signal survives at R=0.093, z=6.5σ, p=5.7×10⁻¹¹ (Chapter 8, Chapter 17) — small, but unambiguously present. Cross-lingual invariance holds across six language families with headline ρ=+0.77 (EN↔FI Hellinger, n=288, p=8×10⁻⁵⁷), witnessing Chapter 12’s Noether-style symmetry claim in data rather than analogy. In music, genre-residualized spectrum-Lasso hits z=28.3σ while hand features collapse by 91% — the starker cousin of books’ 85% collapse (Chapter 21). And the sign flip between books and music on pair_sim_mean and step_mean, each robust at p < 10⁻²⁸, is a directional finding no scalar aesthetics could have made.

But much remains unknown. The framework’s deepest structural predictions — a full aesthetic field equation, the metric signature, genuine holonomy measurement on aesthetic obligations — have not been attempted. Two of its most interesting empirical claims rest on n=2 modalities. Its cross-lingual result, while robust across six families, contains a corpus-shaped hole where the world’s largest written tradition should sit. And its low residualized effect sizes point to a genuine question about how much of aesthetic judgment any embedding-geometry framework can recover in principle.

This chapter surveys nine open problems. It is organized by proximity: near-term empirical gaps (§29.2–§29.4), mid-term theoretical and methodological questions (§29.5–§29.7), and far-horizon integrative and real-time problems (§29.8–§29.9). Each section states the problem, explains why it matters, and sketches what progress would look like. The problems are not defects. They are the research agenda the framework both makes possible and demands.

29.2 The Sinitic Corpus Gap

Chapter 17 reported a cross-lingual invariance result spanning 4,683 non-English books across nineteen languages and ten language families. The top features — pair_sim_mean (ρ=+0.712), mahal_mean (+0.710), Hellinger (+0.675), Bhattacharyya (+0.675), JS (+0.674) — transferred across Germanic, Uralic, Romance, Hellenic, Italic-ancient, and Constructed (Esperanto) corpora with headline values above ρ=+0.77 in the largest bundles. The framework’s symmetry claim (Chapter 12) passed the hardest test its author could construct from available data.

It did not pass one test. The Sinitic branch of the experiment did not happen.

The constraint was not methodological but archival. Project Gutenberg’s Chinese corpus is overwhelmingly composed of classical originals — Confucius’s Analects, Sunzi’s Art of War, the Shijing — rather than Chinese translations of the same Western works whose English, French, Finnish, German, and Dutch translations we could align and compare. Only five Chinese-language bundles reached the n≥20 threshold; at n=5 they sit below statistical power. The geometric signature of Shakespeare-in-Chinese, or Dickens-in-Chinese, or Austen-in-Chinese, cannot be read from a corpus that does not contain them.

This is a corpus-design gap, not a framework gap. But until it is closed, the invariance claim for written Chinese is an absence of evidence rather than evidence of absence. And Chinese is not ornamental to the claim: it is a non-Indo-European, logographic, tonally-inflected written tradition that touches the largest living readership of any human language. An invariance result that holds across Europe but fails to transfer to Chinese would be a different kind of finding than the one we now possess. An invariance result that holds would strengthen the Noether analogy considerably.

What progress would look like. A purpose-built Sino-Western translation corpus: perhaps 300–500 Chinese translations of English-corpus works, drawn from academic and trade editions where machine-readable text exists, quality-controlled for translation-vs-adaptation, then LaBSE-encoded and projected into the English PCA-128 basis exactly as the other nineteen languages were. The labor is archival and legal, not algorithmic. The predicted outcome, if the framework is correct, is that pair_sim_mean and the divergence family should register ρ values in the 0.60–0.75 band seen in Germanic and Uralic — modestly lower, perhaps, given translator-mediated distance from source, but clearly above chance. A null result here would be one of the more interesting failures the framework could suffer.

29.3 The Origin of the Cross-Modal Sign Flip

The most surprising finding in this volume is not the cross-lingual invariance but its modality-specific shadow: the robust sign flip between books and music on two features we had no reason to expect would flip.

pair_sim_mean — internal coherence, measured as the mean cosine similarity between consecutive embedded windows — correlates positively with rating in books (ρ=+0.126, 8.4σ) and negatively with listens in music (ρ=−0.076, p=5×10⁻³³). step_mean — the mean embedding displacement between adjacent windows — correlates negatively with book rating (ρ=−0.096, 6.4σ) and positively with music listens (ρ=+0.071, p=4×10⁻²⁹). Both flips are more than robust: they are of a statistical magnitude that all but rules out sampling artifact. Books reward continuity. Music rewards contrast. The aesthetic geometry has directionality, and that directionality is not universal across modalities.

The question we cannot yet answer is whether this flip reflects a real asymmetry in how audiences weigh coherence across sensory modes, or an artifact of the two encoders we used. LaBSE was trained on sentence-level translation alignment across 109 languages; MERT-v1-330M was trained on 160,000 hours of self-supervised audio with music-specific heads. These encoders learn different notions of “adjacency.” For LaBSE, adjacent paragraphs in a novel are nearly paraphrastic when the author is building an argument; low pair_sim may indicate narrative rupture, which long-form prose tends to penalize. For MERT, adjacent 30-second windows in a track that repeat verbatim may register as loop-padded filler; high pair_sim may indicate aesthetic-cheapness rather than coherence. Under this reading, the sign flip is partly a feature of what each encoder has been trained to consider similar.

A full disambiguation requires a third encoder, a third modality, and ideally both. Would CLIP-encoded image sequences in film — shot-to-shot embeddings across a cinematic trajectory — show the book-like pattern (rewards coherence) or the music-like pattern (rewards contrast)? Would replacing LaBSE with XLM-R or multilingual-E5 move the book effect? Would replacing MERT with CLAP or wav2vec2 move the music effect?

What progress would look like. A full 2×3 design: two content modalities (novels, music tracks) crossed with three encoders per modality (for text: LaBSE, XLM-R, multilingual-E5; for music: MERT, CLAP, M2UGen). If the sign flip survives across all six cells, we have strong evidence it reflects a cross-modal aesthetic asymmetry. If it collapses in some cells, we have evidence of encoder-dependence that pushes the question toward representation learning and away from aesthetics proper. Either outcome is scientifically informative. The current result — n=2 modalities, n=1 encoder each — cannot adjudicate.

29.4 Encoder Dependence

The sign-flip question is a special case of a more general problem. Every numerical result in this book is an image of its corpus as seen through a particular learned encoder. Change the encoder, and the image changes — by how much, we do not know.

LaBSE is a reasonable choice for multilingual sentence embedding: it was built to make translations sit close in embedding space, which is exactly what our cross-lingual invariance test requires. MERT is a reasonable choice for music: its self-supervised pretraining on 160k hours of audio covers a musical ontology far wider than MIR hand features. But reasonableness is not uniqueness. XLM-R and mBERT emphasize different dimensions of linguistic variation. Multilingual-E5 weights retrieval-style similarity over translation alignment. CLAP builds shared text-audio embeddings that would let us bridge modalities in ways MERT alone cannot. wav2vec2 strips to acoustic phonetic content; M2UGen injects generative priors.

Each of these alternatives would yield its own PCA-128 basis, its own channel magnitudes, its own Noether-like invariance test. Would pair_sim_mean still dominate? Would the Hellinger family still lead the cross-lingual ranking? Would the book/music sign flip survive?

We do not know. The conservative expectation is that the qualitative findings — the existence of internal-coherence channels distinct from divergence channels; the existence of genre-residualized signal beyond headline R²; the existence of cross-lingual invariance — will transfer across encoders, while the precise ranking of features and the exact ρ values will shift. The less conservative expectation is that the sign flip may be an encoder signature as much as an aesthetic fact.

What progress would look like. A systematic encoder-ablation study: hold the corpus fixed, swap encoders, report a table of ρ per feature per encoder. Chapter 17’s rankings become a column rather than the whole story. If the rankings are stable across LaBSE / XLM-R / multilingual-E5, we have earned the right to talk about an encoder-invariant aesthetic signal. If they are not, we have earned a narrower, more honest finding: that LaBSE-space aesthetic geometry is what our pipeline measures.

29.5 The Hermeneutic Residue

Genre-residualized R² on books is 0.009. On the nominal scale of predictive accuracy this is a small number. It is also, at z=6.5σ, unambiguously real. And it raises a question the framework must take seriously: how much of aesthetic judgment is in-principle unrecoverable from any embedding-geometry feature set, no matter how sophisticated?

Call the unrecoverable portion the hermeneutic residue: the component of aesthetic valence that depends on reading — on context the text does not contain, on cultural moments embeddings do not see, on individual history and association, on the fact that the reader closing a novel at 2 a.m. is not the same person who opened it at breakfast. A theory of aesthetic geometry can make progress on the recoverable component and still leave the residue untouched.

We suspect the residue is substantial. We cannot presently measure how large. A first step is to construct an empirical ceiling: given the best conceivable embedding and the best conceivable feature set, what fraction of human-rating variance can in principle be predicted? The present result (R²=0.009 post-genre for books, R²≈0.031 post-genre for music’s spectrum-Lasso) is a lower bound. The ceiling is unknown.

What progress would look like. Human-judgment-limited benchmarks: collect pairwise preference data with sufficient rater overlap to estimate annotator agreement, treat inter-annotator agreement as the upper bound on any predictor’s accuracy, and report all model R² as fractions of this ceiling rather than absolute values. A framework that recovers 60% of the inter-annotator-agreement ceiling is doing substantial work even if its absolute R² is 0.05. A framework that recovers 5% of the ceiling is mostly reporting residue. The Chapter 28 program on personal-aesthetic manifolds is the natural home for this measurement.

29.6 Canon Formation as Dynamical System

Chapter 14 sketched a speculative account of canon formation: aesthetic canons are cluster centers in the aesthetic manifold, stabilized by the aggregation of individual judgments and the feedback loop between critical attention and reader exposure. The account is consistent with our Lasso-discovered genre axes (71 non-zero interpretable directions, Chapter 17 §17.4), each of which is plausibly the geometric shadow of a historical canonization process.

We have not tested it. Our data are synchronic: Goodreads ratings are a present-day snapshot, and our Gutenberg corpus is already filtered by the prior century’s canonization. To study canon formation as dynamics we need multi-decade rating trajectories — how does a work move through the manifold over time as it is read, re-read, forgotten, revived?

What progress would look like. A longitudinal corpus. Library-of-Congress circulation records, New York Times bestseller lists going back to 1931, JSTOR citation counts stratified by decade, Goodreads shelving histories where date-of-shelf is available. For each work in an English-language subset of Chapter 17’s corpus, reconstruct the time series of its aggregated reception. Ask whether canon-cluster centers are stable, drifting, or bifurcating. Ask whether newly canonized works arrive at the cluster center before or after acquiring scale. Ask whether the geometric distance from the canonical centroid predicts the probability of canonization — a falsifiable version of the Chapter 14 speculation. The methodological machinery (cluster tracking in Riemannian spaces, non-stationary drift estimation) exists in other literatures; what is missing is the cross-disciplinary work to port it into geometric aesthetics.

29.7 Personal Manifolds and Subject-Disjoint Cross-Validation

Chapter 28 raised the possibility of personal-aesthetic manifolds: a geometry whose metric is fitted per-subject rather than per-population, where “beauty” is replaced by “beauty-to-this-reader” and aesthetic distance becomes a relation indexed by the observer. The chapter argued that such manifolds are computationally tractable — the per-subject fit is just a low-rank adaptation of the population metric — and that they would resolve one of the oldest puzzles of aesthetics: how universal structure coexists with radically individual taste.

We have no empirical test of this claim. A proper test requires subject-disjoint cross-validation on preference data: hold out entire individuals, fit the personal metric on held-in subjects, test whether the fitted metrics generalize to the held-out subjects in any structured way. The question is not “does everyone have a different metric” (they do) but “are the differences drawn from a low-dimensional space of taste variation, or are they idiosyncratic in ways the framework cannot capture?”

What progress would look like. A preference-pair dataset with dense per-subject annotation — at least 500 items per subject, at least 500 subjects. Fit a population metric. Fit per-subject low-rank adaptations. Ask whether the rank of the adaptation subspace is small (taste lives in a low-dimensional variation of a shared geometry) or large (personal taste escapes the population framework). Both outcomes would inform the Chapter 28 program. Neither has been attempted in our lab.

29.8 The Third Modality

The sign-flip finding of §29.3 is an n=2 observation. Books and music disagree about whether coherence is a virtue and contrast is a vice. The disagreement is statistically robust but structurally underdetermined. With two modalities, any pattern of agreement or disagreement has a trivial binary resolution.

The load-bearing next experiment is a third modality. Film is the natural candidate: CLIP-encoded shot-sequence trajectories against IMDb or Letterboxd ratings, with genre and director-disjoint cross-validation. Visual art is a second: CLIP-encoded painting traversals (scan-path simulations or gallery-walk sequences) against expert ratings or auction prices. Architecture is a third: rendered-facade or walkthrough embeddings against architectural-review ratings. Games, where a playthrough is intrinsically temporal and embedding-ready, is a fourth.

If the book/music sign flip extends: pair_sim_mean should be positively correlated with aesthetic valence in some modalities and negatively in others, and the pattern should not be random. The framework predicts the split tracks something like “extended-attention modalities reward coherence” (books, architecture) versus “dynamic-arousal modalities reward contrast” (music, games) — but this prediction is a conjecture awaiting test. A third modality that lines up cleanly with either books or music would promote the prediction from conjecture to hypothesis. A third modality that splits differently would refine the taxonomy.

What progress would look like. Choose one modality. Run the full pipeline: encoder, feature extraction, genre-analogue residualization, correlation with aesthetic-valence proxy. Report with the same discipline as Chapters 17 and 21. One replication would double the evidentiary base of the cross-modal program.

29.9 Real-Time Aesthetic Judgment

Our pipeline is batch-offline. A book is encoded in its entirety, its trajectory computed, its features extracted, its verdict contracted once. Music is analyzed in 30-second clips, aggregated, scored. This matches the data we have. It does not match the temporal structure of aesthetic life.

Live performance, improvised music, gameplay, conversation, interactive art: these modalities resolve over time, in the presence of the audience, and often with the audience’s participation as a variable in the work’s own trajectory. The framework’s geometric machinery does not obviously extend. Parallel transport along a fixed trajectory is mathematically clean; parallel transport along a trajectory whose future depends on the observer’s evaluation of its past is an online control problem, not an offline geometry.

What progress would look like. Online variants of the core pipeline. Streaming embedding with bounded memory. Running estimates of trajectory features (step_mean, pair_sim_mean, recur_rate) updatable in O(1) per window. A real-time aesthetic-valence predictor whose output is itself part of the dynamical system generating the work. The theoretical analogue is non-equilibrium statistical mechanics; the engineering analogue is streaming signal processing; the aesthetic analogue is the moment a jazz musician decides whether to resolve a phrase. None of this has been done in our framework. It is one of the places where aesthetic geometry would need to meet control theory on honest terms.

29.10 Integration with Geometric Ethics

The sibling volume to this book is Geometric Ethics. Its central claim is that moral evaluation is a tensor, not a scalar, and that the moral manifold has geometric structure measurable in cross-lingual and contextual experiments. The parallel with aesthetics is deliberate. The two volumes were written as companions.

But the relationship between the two manifolds — moral and aesthetic — is not worked out. Do they coincide in some substructure? Do they intersect transversally, such that the aesthetic valence of a work is independent of its moral valence at first order but coupled at higher order? Is the alignment between the manifolds itself a cultural variable, tighter in some traditions and looser in others?

The empirical leverage here is considerable. Works exist with high aesthetic consensus and morally divisive content: Céline’s Journey to the End of the Night; Leni Riefenstahl’s Triumph of the Will; Pound’s Cantos; large portions of classical tragedy. Works exist with high moral consensus and low aesthetic consensus: most didactic literature of the Victorian improvement movement. Works exist with both-high and both-low corners of the joint distribution. If the two manifolds are deeply intertwined, the both-high and both-low corners should dominate the joint density. If they are largely orthogonal, the off-diagonal corners should be populated at rates predicted by the marginals. The test is tractable.

What progress would look like. A joint-geometry experiment: encode works with LaBSE (or equivalent), project into both the aesthetic PCA basis (this volume’s Chapter 17) and the moral-structure basis from the Ethics volume (Chapter 17 of that book), measure the angular relationship between the two projections, and correlate the joint location with independent aesthetic and moral-judgment ratings. A positive finding would support a unified geometric theory of judgment, with aesthetics and ethics as two axes of one structure. A null finding would affirm the Kantian intuition that the two realms have independent legislations. A mixed finding — coupling in some genres, decoupling in others — would be the most interesting outcome and the one a framework built on stratification (Chapter 8) should expect.

29.11 A Taxonomy of the Gaps

The nine problems surveyed here do not have the same shape, and it is useful to close with the rough taxonomy the framework imposes on them.

Corpus gaps (§29.2): problems where the framework is in principle ready but the data to test it does not yet exist. The Sinitic corpus is the paradigm case. The remedy is archival labor.

Encoder-dependence gaps (§29.3, §29.4): problems where the finding is empirically real in our setup but its transferability to alternative representations is untested. The remedy is systematic ablation.

Effect-size ceiling gaps (§29.5): problems where small absolute numbers demand a principled account of how small is small. The remedy is calibrating against inter-annotator-agreement bounds.

Speculative-framework gaps (§29.6, §29.7): problems where the book has advanced a theoretical structure that its data cannot yet test. Canon dynamics and personal manifolds fall here. The remedy is longitudinal and subject-disjoint data, respectively.

Generalization gaps (§29.8, §29.9): problems where our n=2 modality result and our batch-offline pipeline leave entire domains of aesthetic life outside the framework’s demonstrated reach. The remedy is a third modality and an online variant of the core methods.

Integration gaps (§29.10): problems where the relationship between this framework and its sibling volumes is asserted in parallel construction but not yet witnessed in joint data. The remedy is a cross-volume experiment.

No one team can close all of these in a decade. Several may not close at all — the hermeneutic residue, in particular, may be the name of a limit rather than a gap. But the framework earns the right to pose the questions it cannot yet answer, and it earns this right by having made empirically falsifiable claims that survived the tests we did run.

29.12 What We Have and What We Lack

The book began with a scalar failure and built, through twenty-seven chapters, a geometric alternative. The alternative has produced: a 17σ discovery-phase result, a 6.5σ residualized-phase result, a ρ=0.70 cross-lingual invariance across six language families, a 28σ music result with a 91% genre confound correctly diagnosed, a pair of cross-modal sign flips at p < 10⁻²⁸, a framework for literary and musical application, and a speculative extension into seven further modalities that explicitly marks its claims as speculative.

It has not produced: a Chinese replication of the invariance result; a test of encoder-independence for any of its core findings; a measurement of the hermeneutic ceiling; a longitudinal test of canon-formation dynamics; a subject-disjoint test of personal-aesthetic manifolds; a third-modality check on the book/music sign-flip taxonomy; an online variant of its offline methods; or a joint experiment tying its manifold to the moral manifold of the sibling volume.

This is the honest inventory. The framework is neither a theorem we have proved nor a mystery we have cleared. It is a subject we have begun to measure, and the measurement has produced both findings and gaps in a proportion a decade of work could not have predicted at the outset.

The next volume, or the next generation of work, will close some of these gaps. Others will turn out to be the framework’s permanent boundary — the place where geometric aesthetics hands the rest of the problem back to hermeneutics, to history, to the reader’s own 2 a.m. encounter with the page. A theory that knows where its boundary lies is the only kind that can be trusted about what lies inside.

We close the ledger with the right-hand column longer than the left. That is where honest research should be.