Who Wrote Act 1 of
Titus Andronicus?

The Question

Scholars have long suspected that Titus Andronicus (c. 1593) is a collaboration. The play's first act differs markedly in style from the rest — more ceremonial, more classical, more rhetorically elaborate. The leading candidate for the co-author is George Peele, a University Wit known for his pageant verse and classical dramas. But suspicion is not proof.

We ran fifty-six computational tests on a corpus of Early Modern plays. The investigation unfolded in ten phases: establishing that Act 1 is genuinely anomalous (Part I), testing whether the anomaly points to Peele specifically (Part II), subjecting that hypothesis to adversarial stress tests (Part III), examining scene-level rare bigram fingerprints (Part IV), profiling content-word frequencies (Part V), running a broad internal-evidence battery from nine independent angles (Part VI), zooming into the sharp stylistic boundary inside Act 1 itself (Part VII), comparing Act 1 against Acts 2–5 using five independent methods and 100 comparator plays from 1585–1600 (Part VIII), zooming into the late portion of Act 1 after the internal boundary to see which comparator plays it most resembles (Tests 43–45), testing whether attribution conclusions change when the text representation itself is changed (Part IX), and replicating core findings using an independent text edition, topic-balanced comparators, progressive lexical ablation, and boundary-local analysis (Part X). What follows is the complete evidence.

A note on structure: Tests 1–17 were developed during the exploratory phase, using external author baselines (Shakespeare vs. Peele). Tests 18–33 (Part VI) form a stricter first-principles battery without relying on external author labels. Tests 34–35 (Part VII) investigate the internal Act 1 boundary identified by the earlier tests. Tests 36–42 (Part VIII) apply five independent test families to Act 1 vs. Acts 2–5, using 100 comparator plays (1585–1600) with no external author labels. Tests 43–45 rerun the same battery on the boundary-defined late section of Act 1 (TWN 2798–3946) identified in Test 34. Tests 46–49 (Part IX) test whether attribution conclusions are sensitive to the choice of text representation, comparing a style-masked approach against a non-masked lexical-semantic approach. Test 50 identifies which specific content words drive the vocabulary difference between Act 1 and Acts 2–5 and asks which plays in the database that vocabulary most resembles. Tests 51–56 (Part X) replicate the core findings using an independent text edition (EEBO TCP), test robustness to topic confounds, comparator resampling, and hard-negative removal, apply progressive lexical ablation to isolate signal layers, and examine signal structure at the internal Act 1 boundary.

Test 1Act-by-Act Function Word Comparison

For each act of Titus, we built a relative-frequency vector over 184 function words (pronouns, articles, prepositions, conjunctions, auxiliaries — drawn from a standard stylometric list). We measured cosine similarity to two baselines: a Shakespeare centroid (mean of 14 plays from the First Folio, excluding Titus itself) and a Peele centroid (mean of 5 plays: The Arraignment of Paris, The Battle of Alcazar, Edward the First, David and Bathsheba, and The Old Wives Tale). “Peele preference” = similarity-to-Peele minus similarity-to-Shakespeare. Positive values indicate a function-word profile closer to Peele; negative values closer to Shakespeare.

Test 2Rolling Stylometric Window

A 500-word sliding window moves across the full text of Titus in 100-word steps (201 measurements). At each window position, the same 184 function-word feature vector is computed and compared via cosine similarity to the Shakespeare and Peele centroids from Test 1. Positive Peele preference (red zone) indicates a function-word profile closer to Peele; negative (blue zone) indicates closer to Shakespeare. This provides a continuous, word-by-word map of where in the play the stylistic signal shifts.

Test 3Peele vs. the Field

Act 1’s function-word vector (184 features) was compared against eight Elizabethan dramatists: Shakespeare, Peele, Marlowe, Kyd, Greene, Lodge, Lyly, and Nashe. For each author, similarity was computed two ways: cosine similarity to the author’s centroid, and Burrows’ Delta (Manhattan distance between z-scored frequency vectors, lower = closer). Cosine similarity measures angular closeness of usage patterns; Delta penalizes large deviations in individual words and is a standard stylometric benchmark.

Test 4Speaker-Stratified Analysis

We parsed the Folger Shakespeare Library edition to isolate each named character’s speeches, then built a separate 184-function-word vector for each character in Act 1 versus the same character in Acts 2–5. Only characters with ≥100 words in both halves were included (Titus, Marcus, Saturninus, Lucius, Bassianus). If one formal speaker drives the signal, only that character would lean Peele. If the signal comes from the author, all characters should shift uniformly.

Test 5Ensemble Classification

A logistic regression classifier was trained on 17 Shakespeare plays (excluding Titus) and 5 Peele plays using three independent feature sets: 100 function words (relative frequencies), 200 most common character trigrams, and 13 word-length bins (proportion of 1-letter, 2-letter, … 13+-letter words) — 313 features total. Leave-one-out cross-validation accuracy: 86.4%. The classifier was applied to each act of Titus individually, returning a probability P(Shakespeare) and P(Peele). The 95% confidence interval was computed via 1,000 bootstrap resamples of the training data.

Bootstrap 95% CI for Act 1 P(Shakespeare): [0.018 – 0.965]. The wide interval reflects the small Peele training corpus (5 plays) and limits confidence in the point estimate.

Test 6Adversarial Feature Search

Using the same 184 function-word feature space, we tested whether the Peele signal survives when subsets of features are removed. We evaluated thematic subsets (pronouns only, auxiliaries only, articles only, prepositions only) and ran a greedy backward-elimination algorithm that removes one function word at a time, choosing the word whose removal most reduces Peele preference. The question: how many function words must be removed before Act 1 flips from Peele to Shakespeare?

Test 7Known Collaborations Control

We applied the identical 184-function-word cosine similarity method to three Shakespeare plays widely accepted as collaborations. Henry VI Part 1 (probable Nashe co-authorship in Acts 1, 3–4), Henry VIII (Fletcher co-authorship throughout), and Two Noble Kinsmen (Fletcher co-authorship in Acts 2–4). For each play we computed the per-act Peele preference against the same Shakespeare and Peele centroids. A method that cannot detect known collaborations would cast doubt on any Titus findings.

Test 8Null Distribution — The Base Rate

We classified all 70 individual acts from the 14 Shakespeare baseline plays using the same 184-function-word cosine similarity measure. This produces a null distribution: the range of Peele preference scores that occur naturally across confirmed Shakespeare texts. If Titus Act 1’s score (+0.035) falls within this normal variation, the finding would be statistically unremarkable. The z-score and empirical p-value quantify how extreme Titus Act 1 is relative to this baseline.

Test 9Register Confound

We selected Shakespeare’s most formal and ceremonial passages — Richard II’s trial and abdication (Act 4), Richard III’s coronation (Act 3–4), Julius Caesar’s Forum speeches (Act 3), and King John’s parley scenes (Act 2) — totaling 19,936 words. A “Formal Shakespeare” centroid was built from these passages using the same 184 function words. If Act 1’s Peele similarity is really just high-register ceremonial language, Act 1 should match Formal Shakespeare as well as it matches Peele.

Test 10The Lodge Problem

Lodge’s Wounds of Civil War is his sole surviving play and, like Titus, a Roman tragedy. To test whether Lodge’s high similarity to Act 1 is a genre effect (shared Roman-tragedy vocabulary) or an authorship signal, we computed Lodge’s cosine similarity to each of the five acts of Titus individually. A genre effect should produce uniformly high similarity across all five acts. An authorship signal (or shared stylistic school) should concentrate in Act 1 specifically.

Test 11Register-Matched Window Comparison

For each Act 1 window (500 words, 100-word steps), we computed five register proxy features: stage-direction density, line-break rate, uppercase ratio, punctuation density, and mean word length. We then found the 50 Shakespeare windows most similar in these register proxies (using Euclidean distance) and computed function-word cosine similarity only against those register-matched windows. This removes register as a confound: if the Peele signal survives matching on formality, it is more likely an authorship effect. If it vanishes, the original signal may have been register-driven.

Test 12Author Typicality — Open Set

Rather than the binary Peele-vs-Shakespeare question, this test computes author typicality for 14 Elizabethan dramatists simultaneously. For each author, we built a centroid from all their 500-word windows, then measured where Act 1’s windows fall in that author’s cosine-similarity distribution (expressed as a percentile). A high percentile (e.g. 75th) means Act 1 fits comfortably within an author’s range; a low percentile (e.g. 25th) means Act 1 is stylistically distant from that author. The 14 authors include: Shakespeare, Peele, Marlowe, Kyd, Greene, Lodge, Lyly, Nashe, Chapman, Chettle, Dekker, Heywood, Jonson, and Munday.

Test 13Imposters Method — Feature Subsampling

The imposters method (Koppel & Winter 2014) tests whether an attribution is robust to feature subsampling. In each of 1,000 trials, we randomly selected 50 of the 100 most common function words and 30 random 500-word windows per author from the reference corpus. The author whose centroid is closest (by cosine similarity) to Act 1 wins that trial. The “win rate” across 1,000 trials measures how consistently each author claims Act 1 under varying feature subsets. If Peele’s signal is robust, Peele should dominate.

Test 14Scene-by-Scene Shakespeare Bigram Density

A “Shakespeare Fingerprint” was built from the rare bigrams (lemma pairs appearing in ≤5 of 527 plays) across the other 35 First Folio plays, excluding Titus (183,425 unique rare bigrams). A separate “Non-Shakespeare Fingerprint” was drawn from 203 contemporary plays outside the Folio (640,063 exclusive rare bigrams). Each Titus scene was then scored by how many rare bigrams match each fingerprint, normalized per 1,000 words. The underlying database contains 527 Early Modern plays totaling over 12 million words, with all tokens lemmatized and lowercased for consistent matching.

Test 15Peele vs. Shakespeare: Exclusive Bigram Fingerprints

A “Peele Fingerprint” was built from six confirmed Peele plays (19,540 rare bigrams): The Arraignment of Paris, The Battle of Alcazar, Descensus Astraeae, Edward the First, David and Bathsheba, and The Old Wives Tale. Bigrams shared by both authors were separated out, leaving 17,277 Peele-exclusive and 176,923 Shakespeare-exclusive rare bigrams. Each Titus scene was scored against these exclusive sets.

Test 16Scene-by-Scene Content-Word Frequency Profile

A frequency vector of 2,133 common content lemmas was built for each Titus scene. “Common” means appearing in ≥90 of 304 plays in the database; function words and stage directions were excluded to focus on content vocabulary (e.g. “blood,” “honour,” “death,” “love”). Log-transformed counts (log_e(1 + count)) were compared via cosine similarity to a Shakespeare centroid (mean of 35 other First Folio plays, excluding Titus) and a Peele centroid (mean of 6 confirmed Peele plays: The Arraignment of Paris, The Battle of Alcazar, Descensus Astraeae, Edward the First, David and Bathsheba, and The Old Wives Tale). This method achieved 97.2% top-1 accuracy when validated across the full First Folio (each play’s nearest neighbor by content-word profile is another play by the same author).

Test 17Peele Null Distribution

For each Shakespeare act (175 acts from 35 First Folio plays, excluding Titus) and each Peele play (6 plays, treated as whole units since they lack act/scene boundaries in the database), we computed a “Peele preference” score using the same 184 function-word feature space: cosine similarity to the Peele centroid minus cosine similarity to the Shakespeare centroid. Positive values = leans Peele; negative = leans Shakespeare. This produces a null distribution of what Peele preference looks like across confirmed Shakespeare and Peele texts, allowing us to place each Titus act within that distribution as a percentile.

Top 5 most Peele-leaning Shakespeare acts: Henry V Act 1 (+0.078), King John Act 2 (+0.042), Henry VI Pt 2 Act 1 (+0.021), Richard II Act 3 (+0.014), Richard III Act 5 (+0.012). All are histories with ceremonial/political rhetoric.

The Diagnostic Panel

Think of this as a medical diagnostic panel, not a single blood test. The nine tests below each ask a different question about the relationship between Act 1 and Acts 2–5. Some test structural features (where does the strongest internal break occur?), others test lexical behavior (does Act 1 use different rare phrases?), and one uses a style-masked language model to isolate functional scaffolding from content. Disagreement between tests is informative, not a bug: it tells us which aspects of writing behavior differ and which do not.

The strongest differences turn out to be lexical (word choices, rare phrase concentrations, masked-LM shifts), while some structural and speaker-controlled tests are only moderate. The combined evidence is real but mixed — not a single knockdown signal.

Test 18Data & Segmentation Audit

Before any stylometric comparison, we validated that all token counts, scene boundaries, and speaker labels agree exactly. This audit uses the words table from the Early Modern Plays Database for Titus Andronicus (PLAY_ID 520, 20,516 tokens across 14 scenes). We also confirmed that 9 speakers have non-zero dialogue tokens in both Act 1 and Acts 2–5, providing sufficient overlap for speaker-controlled analyses.

Test 19Internal Boundary Scan

We built 194 rolling windows (500 tokens, step 100) over Titus dialogue, computing per-window style vectors from function-word frequencies, punctuation rates, line-break rate, and word-length moments. Each candidate split point was scored by the L2 distance between left-mean and right-mean vectors. Two permutation controls (2,000 random boundaries, 2,000 randomized window orders) calibrated the result.

Test 20Register & Structure Profiles

For each act, we measured six register features — stage-direction rate, line-break rate, punctuation rate, upper-initial rate, mean word length, and long-word rate (7+ characters) — and computed z-scores against the FF35 reference set (35 First Folio plays, excluding Titus). Higher z-scores indicate the act is more extreme relative to the reference distribution.

Test 21Speaker-Controlled Shift

For each of the 6 paired speakers (characters with dialogue in both Act 1 and Acts 2–5), we computed the function-word distance between their Act 1 profile and their later profile, then constructed a weighted mean across speakers. This Titus-specific shift was compared to the same metric computed for control plays split at the same ratio (19.2%).

Test 22Reference Distance Centroids

We constructed function-word centroids for Act 1 and Acts 2–5, then measured cosine distance from each reference-set play to both Titus segments. The question: is the Act1-to-rest gap unusually large compared to external reference plays?

Test 23Null-Calibrated Boundary Shift

We computed the L2 distance between left-mean and right-mean style vectors at the true Act 1 boundary, then calibrated it against (a) 5,000 placebo splits at random interior positions within Titus and (b) the same metric for 258 control plays from 1580–1615, each split at Titus’s matched ratio (19.2%).

Test 24Rare Bigram Concentration

We counted all bigrams in Titus dialogue and identified “rare” bigrams — those appearing ≤ 10 times across the entire 1580–1615 corpus. Act 1’s rare bigram rate was compared to 5,000 bootstrap samples of size-matched contiguous chunks from Acts 2–5. Bigrams with ≥ 2 occurrences in Act 1 were tested for enrichment against bootstrap expected counts.

Test 25Lexical Redistribution (Log-Odds)

Using weighted log-odds with a Dirichlet prior (α₀ = 5,000) based on global 1580–1615 lemma frequencies, we scored 1,294 lemmas that appear at least twice across Titus. This method penalizes low-frequency noise and identifies words whose tilts are robust relative to a large informative prior. Sign stability was confirmed via 2,000 bootstrap iterations.

Test 26Masked Language Model Shift

Each Titus token was style-masked: function words were kept verbatim, content words replaced with <C>, punctuation with typed placeholders, and line breaks with <LB>. Trigram language models were trained on each reference corpus set (excluding Titus), and per-window perplexity was scored for 194 rolling windows. The Act 1 vs. rest perplexity shift was compared to matched-ratio splits of 264 control plays.

Test 27Burrows’ Delta Combo Search — All Words

This test runs a broad, objective Burrows’ Delta search using all word types. For each of 1,997 candidate groups — Peele combinations, early Shakespeare combinations, other early anonymous plays, and over 1,200 randomized control groups — we compute the mean absolute z-distance to Titus Act 1 across an MFW grid (100, 150, 200, 300 most frequent words). Lower Delta = closer stylistic proximity. This is a proximity test, not a proof-of-authorship test.

Test 28Burrows’ Delta Combo Search — Function Words Only

This test repeats the 1,997-group Delta search using only function words — stripping away all content vocabulary to focus purely on grammatical scaffolding (pronouns, articles, prepositions, conjunctions, auxiliary verbs). Function-word features are widely regarded as more author-diagnostic because they are less topic-dependent.

Test 29Parody vs Collaboration (Peele-Centred)

This test formalises three competing explanations and evaluates them under Bayesian model comparison. M1 (Collaboration) posits that Act 1 follows a Peele stylistic profile and Acts 2–5 follow Shakespeare. M2 (Parody) says Shakespeare wrote the whole play but mixed surface-level Peele features into Act 1. M3 (Single Author) says Shakespeare wrote everything uniformly. Models are evaluated on 194 rolling windows (500 tokens, step 100) using five feature families: all-words, word bigrams, function words, character n-grams, and part-of-speech n-grams.

The Bayes Factor of M1 over M2 is 14.2 — strong evidence against the parody hypothesis. A synthetic stress test confirms this: when Shakespeare’s later acts are artificially mixed toward Peele at increasing intensity (λ 0–1), the “easy” channel (vocabulary) eventually matches Act 1, but the “hard” channel (function words, character n-grams, POS patterns) never rises above 1.4% match rate. This means lexical imitation alone cannot reproduce the deep-structure signature of Act 1.

Speaker controls further strengthen the case: after controlling for which characters speak in each window, Act 1 retains a positive Peele-direction residual in 97.1% of windows.

Test 30Lodge Negative Control

To test whether the Test 29 result is specific to Peele or an artefact of any non-Shakespeare comparison, the identical framework is re-run with Lodge as the alternative-author profile (using The Wounds of Civil War, his only play in the corpus). If the collaboration signal were generic, Lodge should produce a similar BIC ranking.

Act 1 is still closer to Lodge than Acts 2–5 (LLR: −0.012 vs −0.093), so the early–late contrast is not Peele-dependent. But the collaboration model does not overcome single-Shakespeare under BIC when Lodge is the anchor. Bootstrap robustness confirms: M3 wins 77.7% of BIC-adjusted bootstrap draws.

Test 31Pre-1600 Comparator Scan

Every play created before 1600 in the corpus (129 candidates, excluding Titus) is tested as a single-play collaborator using the same M1/M2/M3 framework. Results are compared to a Lodge baseline. This reveals which plays Act 1 is most stylistically compatible with, without presupposing Peele.

Critical caveat: across all 129 candidates, M3 (single Shakespeare) is the BIC-best model in every case — no individual play makes the collaboration model win outright. However, the ranking of which plays come closest is telling: 1 Henry VI (itself a suspected collaboration), Edward the First (Peele), and 2 Henry VI all cluster at the top.

Test 32Polysemy Fingerprint — Nearest Neighbours

This test goes beyond surface vocabulary to examine sense-mixture patterns. Using a Word2Vec model trained on 1590–1615 drama, context embeddings for each polysemous lemma are clustered into senses. Each play’s usage is then characterised by its distribution across senses (via Jensen–Shannon divergence). The method was validated with leave-one-play-out evaluation: 95% accuracy on Shakespeare plays, 80% on Peele plays.

Of the top 25 neighbours, 12 are First Folio plays, giving the list a Shakespeare-heavy character. However, Peele’s Edward the First appears at rank 17 — present in the neighbourhood, though not dominant. Neighbour rankings are sensitive to embedding configuration, distance metric, and corpus composition; this is contextual evidence, not decisive attribution evidence.

Test 33Polysemy Fingerprint — Authorship Scores

Using leave-one-play-out training (95% Shakespeare accuracy, 80% Peele accuracy), each Titus segment is scored by average log-likelihood under Shakespeare vs Peele sense-mixture models. The polysemous “fingerprint” captures unconscious habits of word-sense selection that are hard to consciously imitate.

Act 1’s Shakespeare-minus-Peele margin is Δ = +0.0009 — an extremely small difference that amounts to a near tie. By contrast, the “rest” of the play has a clearer Shakespeare lean (Δ = 0.017). Scenes 2.2 and 4.1 — both previously flagged in other tests — lean Peele in sense-mixture patterns. This margin should be read as boundary-level ambiguity, not a strong Shakespeare win.

Battery Conclusion

A careful reader should hold two truths at once. First, there is meaningful evidence of Act 1 distinctiveness across multiple independent methods. Seven tests produce strong signals: rare bigram concentration (99.4th percentile, p = 0.006), broad lexical redistribution across hundreds of lemmas, masked language model shift (92nd–97th percentile), dominant Peele proximity in both Delta searches across 1,997 candidate groups, and Bayesian model comparison favouring Peele collaboration over parody (BF = 14.2).

Second, not every structural or speaker-controlled test is extreme. The strongest internal changepoint falls early in Act 1, not at the act boundary. Speaker-controlled shift is moderate (46th percentile). Reference-distance centroids show modest separation. The null-calibrated boundary shift is above average but not decisive (78th percentile, p ≈ 0.23).

The combined evidence supports a real segmental distinctiveness signal in Act 1, but not a simplistic or near-certain single-metric attribution claim. The strongest weight falls on lexical-style differences and stylometric proximity to Peele; structural and semantic measures are more ambiguous. Crucially, the Lodge negative control (Test 30) shows the collaboration signal is Peele-specific, not generic. The polysemy fingerprint (Tests 32–33) yields mixed and configuration-sensitive results: one binary run gives Act 1 a tiny Shakespeare edge (Δ = 0.0009), but multi-author refine runs most often label it Marlowe or Peele, and the Shakespeare-minus-Peele margin is frequently negative. Semantic evidence is boundary-level, not decisive. This sixteen-test panel provides an objective evidence brief, not a final attribution verdict.

Test 34Act 1 Boundary Deep Dive

Re-analysis of all rolling LLR series from Tests 29 and 30 identifies changepoints and largest one-step drops inside Act 1 (3,748 dialogue tokens). A second, author-agnostic split scan uses Jensen–Shannon divergence across five feature families. The consensus late boundary falls at TWN 2798 — approximately 70% through Act 1.

Changepoint & Largest Drops (Peele vs Shakespeare Series)

Author-Agnostic Split Scan (Jensen–Shannon)

Local Contrast at Boundary (500-token windows)

The Text at the Boundary

Right-leaning lemmas after the boundary: bury (+3), speak (+3), father (+2), nature (+2), soul (+2).
Left-leaning lemmas before it: son (−3), burial (−2), dishonour (−2), deed (−2), slay (−2).

Test 35Two-Author Split Comparison

Explicit two-author models are tested inside Act 1 (68 rolling windows of 350 tokens, step 50). For each of 8 comparators × 8 feature series, four models compete: single-Shakespeare, single-Other, two-forward (Other→Shakespeare), and two-reverse (Shakespeare→Other). The forced boundary is at TWN 2798 (nearest window midpoint TWN 2821).

Model Selection Winners (64 scans)

Forced Boundary Focus (Combined Series)

Test 36The Five Test Families

The candidate pool consists of 100 plays dated 1585–1600 in the Early Modern Plays Database, with Titus Andronicus itself excluded. Act 1 contains 3,748 tokens; Acts 2–5 contain 16,104. Each target is compared independently to all candidates across the five test families described above. The consensus rank is the mean rank across all five tests. The results are repeated across five leakage-control variants (described in Test 42) to ensure the pattern is not driven by topical shortcuts.

Test 37Rank Divergence

Using the consensus ranks from Test 36 (baseline variant), we compute the rank delta for each of the 100 comparator plays: rank delta = Acts 2–5 rank − Act 1 rank. A large positive delta means the play is much closer to Act 1 than to the rest; a large negative delta means the opposite.

Test 38Per-Test Heterogeneity

This test examines which play ranks first in each individual test family, under the baseline variant. The chart below shows the per-test ranks for the top five Act 1 consensus candidates, revealing how each candidate performs across the different methods.

Test 39Verification Preference Matrix

The verification aggregate combines z-standardised Act 1–vs–Rest differences across all five tests into a single preference probability per play. This provides a unified measure of how consistently each comparator play aligns with one half of Titus rather than the other.

Test 40Bootstrap Stability

Two hundred block-bootstrap iterations (block size = 400 tokens) were run for both the baseline and strict_all variants. Each iteration resamples the target text, re-computes the three test distances, and records which play ranks first in the resulting consensus. The charts show how often each play finishes in first place.

Test 41Length-Sensitivity Control

The consensus is re-ranked at five minimum-token thresholds: 0 (all 100 candidates), 7,000, 10,000, 12,000, and 15,000. As the threshold rises, shorter plays drop out of the candidate pool. This reveals whether the top consensus leaders are genuinely close stylistic neighbours or artifacts of comparing with very short texts.

Test 42Leakage Controls Summary

The table below shows how much text is removed under each variant and whether the consensus leaders change. Removal rates are measured on the target side (Act 1 or Acts 2–5). The chart shows how the Act 1 consensus ranks of the top five plays change across the five variants.

Test 43Boundary Section Consensus

The charts below show the top 10 consensus nearest neighbours for the boundary section and the remainder (baseline variant). Consensus rank is the mean rank across all five test families — lower means closer.

The table below shows the #1 consensus candidate across all five leakage-control variants. The boundary section’s leader is the same in every variant.

Test 44Section Divergence & Preference

The chart below shows the plays with the largest consensus rank divergence in either direction (baseline variant).

Test 45Section Stability & Caveats

The chart shows bootstrap top-1 shares for the boundary section (baseline variant, 200 resamples).

The table below tracks how the section’s #1 consensus candidate changes as short comparator plays are progressively excluded (baseline variant).

The Representation Question

All stylometric tests depend on a choice: what features of the text do you measure? Different feature sets can capture different aspects of writing, and those aspects may point in different directions. The four tests below hold the evaluation framework constant — same 99 comparator plays (1585–1600), same chunking (320 tokens, step 80), same 256 resampled splits, same 6,000 permutation calibrations — and vary only one thing: the text representation.

Test 46Style-Masked: Act 1

The chart shows the 10 nearest neighbours under style-masked representation. Bars are coloured by whether each play is attributed to Shakespeare (blue) or not (grey).

The table below shows the full top 20. Note the dramatic gap between rank 19 (the last Shakespeare play) and rank 20 (the first non-Shakespeare play).

Test 47Non-Masked Semantic: Act 1

The chart shows the 10 nearest neighbours under non-masked representation.

The table below shows the full top 20. Every play is non-Shakespeare. The nearest Shakespeare play is shown at the bottom.

Test 48Style-Masked: Acts 2–5

The chart shows the 10 nearest neighbours for Acts 2–5 under style-masked representation.

Test 49Non-Masked Semantic: Acts 2–5

The chart shows the 10 nearest neighbours for Acts 2–5 under non-masked representation.

Test 50Content-Word Register Analysis

The chart below shows the percentage of First Folio (Shakespeare) plays at each top-N level for Act 1 versus Acts 2–5.

Act 1’s 15 nearest neighbours by content-word frequency profile. Note: rank 1 is the full Titus (self-match across acts).

Where do Peele’s plays rank for Act 1 versus Acts 2–5?

The chart below shows Act 1’s classical/ceremonial vocabulary compared to Acts 2–5, measured in occurrences per 1,000 content words.

All tests in this section use the EEBO TCP text (A12017) of Titus Andronicus Act 1 — an independently transcribed edition — rather than the EMPD text used in Parts I–IX. The same 99 EMPD comparator plays (1585–1600), chunking parameters (320 tokens, step 80), and permutation calibration framework are held constant. The purpose is to confirm that the findings reported above are not artefacts of a single text edition, comparator set, or small number of dominant non-Shakespeare plays.

Test 51Cross-Edition Replication

Test 52Topic-Matched Comparators

Test 53Lexical Ablation Ladder

Test 54Function Subchannel Decomposition

Test 55Comparator Stability

Bootstrap stability (12 iterations, n = 15 per class):

Hard-negative removal cascade (removing top-k nearest non-Shakespeare plays):

Test 56Boundary-Local Signal Structure

Style-masked authorship verification at the boundary: