Does the Geometry Work

Name: Rodin
Author: Rodin

Canonical-pair ranks vs. random baseline

Rodin makes a falsifiable prediction. If the matching geometry recovers real intellectual kinship, then pairs of writers with documented historical relationships — mentor & mentee, antagonist & antagonist, lineage holder & inheritor — should rank one another well above chance among the 253 rival profiles in the archive.

The fixture below is hand-curated from textual record: Kant on Hume, Pascal on Montaigne, Wollstonecraft to Shelley. Three independent metrics are evaluated — the production cosine ranker, the Jaccard + topology blend, and Burrows’ Delta on function-word frequencies. A random metric would place each partner at a median rank near 127. The numbers below are what the live system returns.

Summary

Cosine

Documented median rank

6.0

Controls median rank

181.0

Random baseline

127

Blend

Documented median rank

37.0

Controls median rank

118.0

Random baseline

127

Stylometric

Documented median rank

44.0

Controls median rank

215.0

Random baseline

127

Lower is better. Documented pairs should sit far below the random baseline of 127; negative controls (unrelated thinkers from different eras and domains) should sit near it. Rank is the smaller of A→B and B→A — matching is symmetric in intent, but the geometry is not always perfectly so.

Documented Pairs (23)

Pair	Label	Cos	Blend	Stylo	Δ
Ralph Waldo Emerson × Henry David Thoreau RWE mentored HDT; close friends	MENTOR MENTEE	16/253	31/253	109/253	1.13
Ralph Waldo Emerson × Walt Whitman RWE: "I greet you at the beginning of a great career"	CHAMPION	2/253	37/253	73/253	1.03
Henry David Thoreau × Walt Whitman Met 1856; transcendentalist circle	CONTEMPORARY	26/253	24/253	111/253	0.98
Immanuel Kant × David Hume Kant: "Hume awakened me from dogmatic slumber"	LINEAGE CRITICAL	1/253	82/253	126/253	1.16
Immanuel Kant × Jean-Jacques Rousseau Kant hung Rousseau's portrait in his study	ASYMMETRIC INFLUENCE	22/253	91/253	203/253	1.33
David Hume × John Locke Hume built on and departed from Locke	LINEAGE	64/253	37/253	20/253	0.91
John Locke × Jean-Jacques Rousseau Rousseau critiqued Locke on social contract	ANTAGONISTIC	1/253	17/253	28/253	0.90
Friedrich Nietzsche × Arthur Schopenhauer N read S deeply, later moved beyond	INFLUENCE REBELLION	1/253	2/253	18/253	0.85
Friedrich Nietzsche × Plato N framed much of his work anti-Platonic	ANTAGONISTIC	5/253	15/253	13/253	0.87
René Descartes × Blaise Pascal Pascal critiqued Cartesian certainty	ANTAGONISTIC	20/253	16/253	44/253	0.95
Blaise Pascal × Michel de Montaigne Pascal wrestled with Montaigne's skepticism	PRODUCTIVE FRICTION	10/253	37/253	3/253	0.70
Leo Tolstoy × Fyodor Dostoevsky Contemporaries who never met, mutual awareness	MUTUAL AVOIDANCE	6/253	13/253	183/253	1.29
Mary Wollstonecraft × Mary Shelley Mother/daughter; W died at M's birth	FILIAL INFLUENCE	58/253	9/253	68/253	1.13
Lao Tzu × Zhuangzi Taoist succession	LINEAGE	31/253	51/253	207/253	1.46
Confucius × Lao Tzu Legendary opposition of Chinese traditions	PHILOSOPHICAL OPPOSITION	2/253	61/253	35/253	1.08
William James × John Dewey American pragmatist colleagues	COLLEAGUES	106/253	9/253	6/253	0.67
William James × Bertrand Russell Russell engaged James's pragmatism	CRITICAL DIALOGUE	1/253	46/253	27/253	1.03
Tyler Cowen × Alex Tabarrok Marginal Revolution co-authors 20+ years	COLLABORATORS	1/253	1/253	1/253	0.00
Ezra Klein × Matt Yglesias Co-founded Vox	CO FOUNDERS	3/253	54/253	16/253	1.21
Scott Alexander × Gwern Branwen Rationalist blogosphere core	PEER NETWORK	11/253	136/253	49/253	0.87
Robin Hanson × Eliezer Yudkowsky Famous FOOM debate; same field, opposed priors	ANTAGONISTIC SAME FIELD	2/253	2/253	36/253	0.84
Naval Ravikant × Balaji Srinivasan Crypto/sovereignty thought space	PEER NETWORK	41/253	49/253	91/253	1.16
Nassim Taleb × Nassim Nicholas Taleb (Incerto) Same person — expected rank 1 or 2	IDENTITY DUPLICATE	1/253	37/253	57/253	1.12

Negative Controls (4)

Pairs with no documented intellectual contact, drawn from different eras and disciplines. A working metric should rank these near the random baseline — not high, not low.

Pair	Label	Cos	Blend	Stylo	Δ
Confucius × Ali Abdaal Ancient Chinese / modern productivity YouTuber	UNRELATED CONTROL	173/253	240/253	215/253	1.52
Immanuel Kant × Bryan Johnson German idealism / longevity biohacker	UNRELATED CONTROL	181/253	118/253	—	—
Lao Tzu × Tomas Pueyo Taoist sage / viral infographic writer	UNRELATED CONTROL	167/253	56/253	100/253	1.21
Rabindranath Tagore × Shreyas Doshi Bengali polymath / product management coach	UNRELATED CONTROL	226/253	110/253	—	—

External Judge Cross-Check — inconclusive

A second opinion: an independent model from a different family (gemini-3-pro-preview) was given two redacted passages and asked to rate cognitive-style similarity 1–7. The judgments were correlated with each engine metric’s pair rank via Spearman ρ — pre-registered, with bootstrap 95% confidence intervals (B = 2000 resamples) and a published threshold for pass / fail / inconclusive.

Why the result is inconclusive. N=6 below 10-pair minimum for reliable ρ. The judge identified at least one author in 6/6pairs despite name-token redaction — famous philosophers are unredactable at this passage scale, and what looks like agreement or disagreement with the engine cannot be cleanly distinguished from training-data recall. A clean version requires contemporary writers absent from training data, or a per-dimension scoring protocol that does not ask the judge to identify the author.

Cos

-0.45

95% CI [-1.00, 1.00]

Blend

+0.15

95% CI [-0.80, 1.00]

Stylo

-0.32

95% CI [-1.00, 0.95]

The numbers above are reported for completeness, not as a verdict; with n = 6 every CI spans the full interval. One observation that survives the contamination caveat: the CTA-blend rank was the only metric whose ρ pointed in the same direction as the judge.

What this measures — and what it doesn’t

What it does measure. Whether the geometry is structurally non-random. If documented pairs cluster well below the 127-rank baseline, the metric is recovering real signal — not perfectly, but reliably above chance.

What it does not measure. Whether the matches are good in the felt sense — whether the writer would, on reading the suggested mind, recognize kinship. That is a human judgment the geometry approximates but does not replace. The fixture is also small (n ≈ 23) and curated; it tests whether the geometry behaves sanely on hand-picked positives, not whether it generalizes to every possible pair.

Why three metrics. The production matcher blends them. Cosine catches semantic neighborhood; the Jaccard+topology blend reads structural resemblance and surface motif overlap; Burrows’ Delta reads prose at the function-word grain. The three answer different questions about kinship, and a pair the geometry truly recovers should land well across more than one of them.

Generated 2026-06-11· refreshed daily

← Back to Methodology