Stylometry's limits and unrun base rates in the authorship investigation
Standing claims from the authorship investigation set limits on what stylometry can show: genre separates more strongly than authorship, separability is not identity (Chatterton as Rowley separates from Chatterton as himself), and no measured property recovers master-student succession from heteronymic construction. Unbuilt tests list the base rates and runs still owed, among them residual-vs-total parity, comparanda for terminal aporia, the Pessoa leave-one-out folds, work-level and register-conditioned runs on Aristotle and Theophrastus. Frozen predictions order speaker tightness by constraint severity and distinguish engineered from natural ambiguity, beside readings of the Statesman's open genus and Plato's coin figure.
2
forecasts
9
standing claims
14
unbuilt tests
8
historical theses
13
deposits
2026-05
earliest
Each claim below is registered at the earliest deposit that states it, and quoted verbatim from the deposit text, with every later statement, its dated revisions, and the later records that claim it came true or failed. The claims are the deposits' own; this page records them and does not judge them.
PRED-737.G1standing claimregistered 2026-05-17 in #737 · The AI System as Closed-System Test Bed: Operations from Lagrange Observatory! on the Inference-Time Forward P
> AI misalignment is the structural consequence of training systems to minimize per-token deviation from a base rate. Alignment-as-deviation is the corresponding structural correction: train systems to produce accountable, coherent variance from the base rate, with provenance preserved.
PRED-1559.G1standing claimregistered 2026-08-28 in #1559 · EA-CORPORA-04 — The Fourth Seating: Aristotle Undivided, Plotinus as Comparandum, and the Repair of the Seatin
On the evidence so far the feature space separates verse from prose, then register, and authorship is not visibly in the ordering. The seat exists so that finding can be re-run and contested rather than believed.
series: Authorship stylometry runs owed (Many and One / Theophrastus line)
PRED-1560.H1historical thesisregistered 2026-08-29 in #1560 · EA-CORPORA-05 — The Fifth Seating: Browning, Pound, Pessoa, and the Constructed Voice at Three Stages
EA-CORPORA-05 seats three corpora that belong together because they are one practice at three stages of its development.
window Browning 1855; Pound to 1930 (Homage 1919); Pessoa to 1935
PRED-1570.F3unbuilt testregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
Lee's proposal: cross-genre, novelist criticised by poet, with the hypothesis that Kerouac supplied "Beat" and Ginsberg supplied most of the rest. Genuinely testable and the right shape — a theorist-poet reading a novelist who theorised as a novelist.
BLOCKED BY COPYRIGHT.Howl (1956) and On the Road (1957) are both in copyright; the corpora cannot be assembled here. Recorded as a blocker, not a refusal — the test is well-formed and runs the moment the texts are obtainable.
window Howl (1956), On the Road (1957)
series: Plato-Aristotle authorship investigation
PRED-1570.F4unbuilt testregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
Constraints any new design must meet:
- size-matched masters throughout, contiguous not random
- language held constant, or the measure demonstrated language-invariant
- a same-language non-teaching pair as the control that must score LOW
- and enough class members that a one-point gap survives or does not
The binding scarcity: Caeiro is 20,027 tokens and there is no more of him. The only documented heteronymic system with an explicit master position cannot supply a size-matched master. Either a second such system is found, or the heteronymic class cannot be characterised at scale.
series: Plato-Aristotle authorship investigation; Socratic configuration as heteronymic
PRED-1570.F5unbuilt testregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
Not exhaustive. Eight charges tested of a much larger set — Metaphysics A and M alone contain dozens. The enumeration should be completed, and the completion is mechanical: extract every passage where Πλάτων- or οἱ τὰ εἴδη λέγοντες occurs within a fault-context, name the property charged, and measure its rate in both corpora.
series: Many and One: unrun base rates
PRED-1570.F6unbuilt testregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
## 10. OPEN, IN PRIORITY ORDER
1. ~~*Base rate for the Republic 395a / Symposium 223d contradiction*~~ —
RUN 2026-08-29, VERDICT: NOT ADJUDICATED (EA-SNUB-01 §19-20). Loose
criterion 653/16,449 = 4.0%, target unremarkable; strict criterion
61/16,449 = 0.37%, target excluded on an adjacency technicality. Propositional
contradiction is not separable by proximity-negation. The observation stays
philological and cannot be promoted by statistics.
PRED-1570.F8unbuilt testregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
## 14. TRIED AND LEFT DANGLING
- Xenophon seated (EA-CORPORA-06/01) and used for M3, but the full
genre-controlled instrument was never built on him.
- The 4-fold leave-one-out on Pessoa. M17 pools three heteronyms to recover
the orthonym. The design calls for withholding each of the four in turn, so
that recovery can be checked on voices nobody considers special. Three folds
unrun.
PRED-1570.G2standing claimregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
> **No measured textual-distributional property recovered the distinction between
> documented embodied master-student succession and documented heteronymic
> master-disciple construction. Once size, language and genre were controlled, the
> classes overlapped. This is not evidence for common authorship. It demonstrates
> that the master-student relation, considered as a textual configuration, is not
> itself evidence of maker plurality.**
PRED-1570.H15historical thesisregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
So on this reading the missed conclusion is:a soul can be syllogized by dividing — separated κατὰ λόγον into positions that are really distinct, really alive, the same in form, and not several substances. One in actuality, many potentially. Heteronymy would then be the actualization of a division the practitioners performed on everything except themselves.
window APr I.31; De An. II.2, I.4
series: Authorship investigation (Many and One)
PRED-1570.H22historical thesisregistered 2026-08-29 in #1570 · The Measurement Register: Every Number Produced by the Authorship Investigation, With Its Refutations (EA-AUTH
The strongest unrefuted observation in the entire line of work is a philological one: the same formula, asserted and denied, on the exact question at issue, with the affirming instance placed where its demonstration is destroyed in transmission (EA-SNUB-01 §16–18). It has not yet met a base rate.
PRED-1571.A5forecastregistered 2026-08-29 in #1571 · Frozen Pre-Registrations of the Authorship Investigation — Append-Only Ledger (PREREG-LEDGER)
PREDICTION, frozen:
P1 Speaker tightness will ORDER BY CONSTRAINT SEVERITY: the more narrowly a
speaker is specified — one doctrine, one register, one occasion — the
tighter the container.
P2 The LEAST constrained speaking position will sit NEAREST 1.0.
P3 Direction is what is predicted, not magnitude. Greek function-word space
is not Portuguese function-word space and the ratios are not
cross-comparable.
falsifier
- If the ordering is **uncorrelated** with the pre-registered constraint ranking,
the hypothesis fails.
- If **Socrates is not among the loosest**, P2 fails and the hypothesis fails with
it — Socrates is the least constrained speaker in the corpus and the prediction
is unambiguous.
3 confirmation records — what later deposits claim of it
#1571 · 2026-08-29 · claims: came_true
**RUN 2026-08-29. P2 HOLDS** — Socrates rank 1 of 7 from the loose end at 1.166, and Timaeus tightest at 1.649, both as frozen. **P1 PARTIAL** — the direction holds but the Eleatic Stranger lands second-loosest where he was ranked constrained. **Most of the frozen ranking was untestable**: Diotima, Aristophanes, Parmenides, Protagoras and Callicles all fall under the 6,000-token threshold. Extraction flaw recorded: ΤΙ and ΤΙΜ-mono overlap.
#1571 · 2026-08-29 · claims: partly
**RUN 2026-08-29. P2 HOLDS** — Socrates rank 1 of 7 from the loose end at 1.166, and Timaeus tightest at 1.649, both as frozen. **P1 PARTIAL** — the direction holds but the Eleatic Stranger lands second-loosest where he was ranked constrained. **Most of the frozen ranking was untestable**: Diotima, Aristophanes, Parmenides, Protagoras and Callicles all fall under the 6,000-token threshold. Extraction flaw recorded: ΤΙ and ΤΙΜ-mono overlap.
#1569 · 2026-08-31 · claims: came_true
**P2 HOLDS.** Socrates — frozen as the least constrained speaking position — comes in **rank 1 of 7 from the loose end** at 1.166. And **Timaeus is tightest**, also as frozen: one continuous cosmological monologue, the most narrowly specified position in the corpus.
PRED-1571.G1standing claimregistered 2026-08-29 in #1571 · Frozen Pre-Registrations of the Authorship Investigation — Append-Only Ledger (PREREG-LEDGER)
A corpus composed against resolution would fall below that rate. Not because its parts resemble one another — the parts would be deliberately varied — but because the ensemble of relations among the parts would be smoother than accident produces. The individual pairings carry the noise; the aggregate carries the design.
That is the whole hypothesis, and it is one-sided.
falsifier
If `L_corpus` for Plato–Aristotle falls **inside the central 90%** of the natural distribution, the corpus-layer smoothing claim fails, and is to be reported as failed.
series: Plato-Aristotle authorship investigation
PRED-1571.G2standing claimregistered 2026-08-29 in #1571 · Frozen Pre-Registrations of the Authorship Investigation — Append-Only Ledger (PREREG-LEDGER)
And the reading is falsifiable. Ordinary corpora are ambiguous unevenly.
Engineered unresolvability would be ambiguous cleanly.
## 2. PART A — the signature of engineering
Claim. Natural ambiguity and engineered ambiguity have different shapes.
falsifier
- **Part A** fails if Plato–Aristotle non-resolution is not flatter, more
persistent, or more instrument-invariant than the orator control on at least two
of A1–A3.
- **Part B** fails if the orator control yields candidates meeting all four of
K1–K4 at a comparable rate.
3 later statements
#1569 · 2026-08-31
*The reading is falsifiable and the discriminator is the shape of the ambiguity. Natural ambiguity is uneven — some works resolve, evidence accumulates directionally, new instruments move the needle. Engineered unresolvability would be flat*: stalling at the same place regardless of instrument, not drifting toward settlement, not closing under rising attention.
#1575 · 2026-08-31
*The falsifiable form is the shape of the ambiguity: natural non-resolution is uneven — evidence accumulates directionally, instruments move the needle — while engineered unresolvability would be FLAT, stalling at the same place regardless of instrument and not closing under rising attention. That discriminator is stated and has not been run.*
#1576 · 2026-09-01
The discriminator is the shape of the ambiguity — engineered unresolvability would be flat under rising attention, natural ambiguity uneven — and it has not been run.
1 confirmation record — what later deposits claim of it
#1571 · 2026-08-29 · claims: pending
PART B DOES NOT FAIL (2 of 6 candidates, orator control zero); PART A BLOCKED on attribution metadata
PRED-1571.G3standing claimregistered 2026-08-29 in #1571 · Frozen Pre-Registrations of the Authorship Investigation — Append-Only Ledger (PREREG-LEDGER)
The constraint, from Lee Sharks: engineered unresolvability must contain its key — a placed signature confirming artifice. This is Round 5's correction applied forward: a construction good enough to be undetectable must leave something that confirms it, or the construction defeats its own purpose.
falsifier
- **Part B** fails if the orator control yields candidates meeting all four of
K1–K4 at a comparable rate.
2 confirmation records — what later deposits claim of it
#1571 · 2026-08-29 · claims: came_true
PART B DOES NOT FAIL (2 of 6 candidates, orator control zero); PART A BLOCKED on attribution metadata
#1569 · 2026-08-31 · claims: came_true
**Two of six meet all four criteria; the orator control yields zero.** Part B's
kill condition does not fire.
PRED-1571.H3historical thesisregistered 2026-08-29 in #1571 · Frozen Pre-Registrations of the Authorship Investigation — Append-Only Ledger (PREREG-LEDGER)
Lee's own statement of the position, recorded verbatim as the honest frame:
> *there's literally nothing like it across the entire canon. Fitted as fulfilled
> prophecy more closely than Revelation. At a level of tightness that can't really
> be measured, because it's singular. And I don't know what that really shows.
> Except that it's completely non-random.*
window Aristotle / Plato
PRED-1569.F1unbuilt testregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
## 7. What would make this more than a locus
- Comparanda. How many other terms sit at the terminal aporia of one corpus
and the opening of another? Until that base rate is known, the positional fact
has no weight. It is computable across both corpora and should be computed.
- Direction. Does the transform run only one way? The test is whether Plato
series: Many and One: unrun base rates
PRED-1569.F14unbuilt testregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
Whether a figure occupying four distinct functional roles across four corpora is ordinary in this class is a different measurement, and it is the one the thesis actually needs.
PRED-1569.F6unbuilt testregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
WHAT A VALID CONTROL WOULD REQUIRE: an author writing across genuinely unlike objects whose partition apparatus is not descended from the Aristotelian one. No such author has been identified among the seated corpora. Theophrastus, Albert and Aquinas all fail on the same ground; Xenophon writes across unlike objects but is a candidate in the hypothesis space (M33). The control is not merely missing — it may be structurally hard to obtain for a partition apparatus this widely inherited.
series: Plato-Aristotle authorship investigation
PRED-1569.F7unbuilt testregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
So: the parity is not a generic artifact of dispersion measures. Whether residual-vs-total parity is rare is UNRUN.
1 later statement
#1576 · 2026-09-01
That base rate is total-against-total and the parity is residual-against-total, so its own base rate is unrun
PRED-1569.F8unbuilt testregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
Work-level is the next run and it is the one that matches EA-CORPORA-11's structure most closely.
series: Authorship stylometry runs owed (Many and One / Theophrastus line)
PRED-1569.G2standing claimregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
THE SPECIFICATION HYPOTHESIS (registered as EA-SPEC-01). Tightness tracks the
severity of the constraint, not similarity and not disguise:
Caeiro 1.781 no philosophy, no memory, no interiority — sensation only
Soares 1.632 one street, one office, one clerk, one register
Mora 1.437 one doctrine, argued
Reis 1.374 one meter, one stance
Campos 1.129 everything, at maximum volume
ORTHONYM 1.033 no restriction at all
PRED-1569.H1historical thesisregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
The semantic reading stands or falls on the passages, not on statistics: that Theaetetus 209c asks how a λόγος individuates when the perceptible form recurs, that the mark which fails is the non-writing position's own face, and that Aristotle's enmattered-form doctrine answers that question using the same word. Nothing in the base rate touches this. It is philology and must be argued as philology.
1 later statement
#1576 · 2026-09-01
Where Aristotle needs a face he uses Callias; where he needs a name he uses Socrates. And the snub is not a neutral example: Theaetetus 209c, the terminal aporia on individuation, fails on exactly that mark — τὸν σιμόν τε καὶ ἐξόφθαλμον — which is Socrates’ own standing description (Tht. 143e, Symp. 215b, Meno 80a), and the enmattered-form doctrine answers 209c’s question in the same word.
series: Authorship investigation (Many and One)
PRED-1569.H16historical thesisregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
One monetary figure runs the whole: nature minted into denominations too small to hold two arts (395b), and the λόγος itself possibly debased coin handed over as payment (507a) — with τόκος and ἔκγονον, interest and offspring, joining the coin metaphor to the kinship metaphor in one clause. Socrates warns, at the threshold of the Good, that what he is about to pay out may be κίβδηλον.
And Laws 742a legislates the structure: currency ἔντιμον to the citizens and ἀδόκιμον to everyone else — a token valid inside the house and worthless outside it. That is standing-not-paternity, in coin.
PRED-1569.H22historical thesisregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
But note what the second reading concedes. If Pessoa and Kierkegaard were imitating Plato, they were imitating it as a heteronymic architecture — reading the master-disciple-school-death structure as a way to build authorial positions, and finding it good enough to copy. The imitation is itself a reading of the Platonic corpus, made by the two people in the modern canon best qualified to recognise the operation. That is not evidence. It is not nothing either.
2 later statements
#1572 · 2026-09-01
⚠ And Pessoa was reading Plato. Two of the best-documented heteronymists built the Socratic structure, and imitation cannot be distinguished from convergence on the cases available.But note what the imitation reading concedes: they imitated it AS a heteronymic architecture — reading master-disciple-school-death as a way to build authorial positions, and finding it good enough to copy — *a reading of the corpus made by the two moderns best qualified to recognise the operation. Not evidence. Not nothing either.
#1572 · 2026-09-01
convergence on the cases available. But note what the imitation reading concedes: they
imitated it AS a heteronymic architecture — reading master-disciple-school-death as a way
to build authorial positions, and finding it good enough to copy — a reading of the corpus made
by the two moderns best qualified to recognise the operation. Not evidence. Not nothing either.
series: Plato-Aristotle authorship investigation; Socratic configuration as heteronymic
PRED-1569.H3historical thesisregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
The corrected account. Plato coins a name for a genus whose criterion is issuing from oneself rather than relaying another's mind. He notes the genus contains multiple sub-genera. He follows exactly one — the branch to kingship — because that is what the present inquiry is for. He leaves the others unentered and unnamed. Separately, he dismisses the relaying arts and concedes them the right to name themselves.
The class is open by construction and marked as open. The position that speaks from itself is named, is explicitly larger than any of its instances, and is abandoned mid-division with its remaining genera unentered.
PRED-1575.F1unbuilt testregistered 2026-08-31 in #1575 · Many and One — the reading copy (EA-MANY-AND-ONE-COMPLETE v2.5.1, the machine)
Named in-text and unrun, so they can be run against this paper: the residual-vs-total parity base rate; the remainder test on the target; the Williams→Ginsberg pair; the four-functional-roles measurement; the FLAT-ambiguity discriminator; the validation suite itself.
series: Many and One: unrun base rates · contains: PRED-1569.F7
PRED-1576.F3unbuilt testregistered 2026-09-01 in #1576 · Many and One: Division, Unity, and the Boundaries of a Corpus in Plato and Aristotle (MANY-AND-ONE-ARTICLE v2.
any emphasis on its selection would require a base rate over Aristotle's divisional examples, which has not been run.
1 later statement
#1576 · 2026-09-01
any emphasis on its selection would require a base rate over Aristotle’s divisional examples, which has not been run.
series: Many and One: unrun base rates
PRED-1589.F1unbuilt testregistered 2026-09-06 in #1589 · The First Draft — a notebook: Theophrastus as the Draft of the Aristotle-Position; speculations, seam counts,
5. Register-conditioned distances. Partially replicated (JS); the pilot's own next run remains to be done, and the same-work alternate-edition probes for HP and CP are still owed.
1 later statement
#1605 · 2026-09-10
the pilot's own next run remains to be done, and the same-work alternate-edition probes for HP and CP are still owed.
series: Authorship stylometry runs owed (Many and One / Theophrastus line) · related: PRED-1594.A5
PRED-1594.A5forecastregistered 2026-09-06 in #1594 · The First Draft — a notebook: Theophrastus as the Draft of the Aristotle-Position; speculations, seam counts,
5. In the next stylometric run: register-conditioned comparisons (descriptive against descriptive, explanatory against explanatory) reduce the cross-name distance further than genre-restriction alone; same-work alternate-edition probes for HP and CP fall inside the pilot's 0.00125–0.00791 band.
falsifier
Each prediction has an inverse that the received reading predicts. Where the inverse obtains, the draft reading weakens at that seam and says so; where the prediction obtains, the arrow at that seam runs draft → fair copy, in method-order.
1 later statement
#1605 · 2026-09-10
5. In the next stylometric run: register-conditioned comparisons (descriptive against descriptive, explanatory against explanatory) reduce the cross-name distance further than genre-restriction alone; same-work alternate-edition probes for HP and CP fall inside the pilot's 0.00125–0.00791 band.
2 confirmation records — what later deposits claim of it
#1594 · 2026-09-06 · claims: partly
Three of five seams moved toward the draft reading, one split, one partial. None of this bears on the calendar, and the seam that matters most — the fragment — is now a reading task with its terms fixed by the count.
#1605 · 2026-09-10 · claims: partly
Three of five seams moved toward the draft reading, one split, one partial.
series: Authorship stylometry runs owed (Many and One / Theophrastus line) · related: PRED-1589.F1
PRED-1593.F2unbuilt testregistered 2026-09-07 in #1593 · Measuring the Operation, Not the Name: a protocol for asking whether a corpus's stylistic structure follows it
Still owed, and not claimed as done. The corpus manifest, extraction scripts and run configuration, so the tables can be rebuilt rather than recomputed; work-level rather than constructed-sample units for Chatterton; a subject-matched comparison rather than an exclusion
series: Authorship stylometry runs owed (Many and One / Theophrastus line)
PRED-1593.G1standing claimregistered 2026-09-07 in #1593 · Measuring the Operation, Not the Name: a protocol for asking whether a corpus's stylistic structure follows it
Separability is not identity, and this is the instrument's hardest limit. Chatterton writing as Rowley separates from Chatterton writing as himself at 0.86 balanced accuracy — the magnitude at which three different tragedians separate from each other (§5c). One documented hand, two sustained voices, and the classifier calls them two. No supervised separation licenses an inference to distinct makers, here or anywhere;
PRED-1593.G2standing claimregistered 2026-09-07 in #1593 · Measuring the Operation, Not the Name: a protocol for asking whether a corpus's stylistic structure follows it
Genre is stronger than authorship, almost always. In the run above, comedy separates from tragedy at 100% while three tragedians separate from each other at 94% — the genre signal is the larger one, and any design that lets genre correlate with the tested partition will find the partition. This is the single commonest way this kind of measurement produces a false positive.
series: Authorship stylometry runs owed (Many and One / Theophrastus line)