Between the Spokes: What the Embeddings Can't See
A paper that applies algebraic number theory to planar geometry combines vocabulary and concepts from both communities; a representation learned from titles and abstracts should reflect that. The natural test of an automated bridge-discovery pipeline is whether a known bridge sits between its two field clusters — or at minimum closer to the source-technique field than a pure-target-field paper would sit.
Will Sawin's same-day disproof of the Erdős unit-distance conjecture is exactly such a paper — a planar geometry problem lifted into algebraic number theory using a Hajir–Maire–Ramakrishna tower-cutting technique the discrete-geometry community had never touched. In SPECTER2 — AllenAI's standard scientific-paper embedding — Sawin sits firmly inside the geometry neighborhood: d(Sawin, AMP small-unit-distance) = 0.103, d(Sawin, Pach–Raz survey) = 0.120, d(Sawin, Erdős-2D baseline) = 0.126. The algebraic-number-theory anchor HMR sits at d = 0.159 — a clearly separable band further out. Gemini's embedder shows the same directional pattern (Pearson r = 0.80 with SPECTER2 on pairwise distances): Sawin's neighbors are geometry-dominant in both embeddings; the algebra anchors are further out in both.
The same surface-reading shows up in coverage of OpenAI's same-day disproof: a fair amount of writeups still read the result as AI doing novel math. The actual move was the bridge — the number-theory community had the technique, the discrete geometers (Spencer, Szemerédi, Trotter, Guth, Katz, Alon) had the problem, five years between publication and disproof. Between the Spokes argued the cost of walking into those inter-field gaps has dropped — a model carries the cross-domain breadth a human career can't accumulate. The embedding calibration says that drop isn't reaching the standard representations downstream.
Why the embedding hides it
Embedders learn surface signatures from titles and abstracts; technique provenance lives in the body of the paper. Sawin's abstract is about unit distances, so the embedding scores high on unit-distance neighbors. Two different embedders from two different vendors give the same answer because they're both reading the same surface.
The scan, run anyway
If the embedding can't see bridges, a scan that relies on embedding geometry will surface non-bridges. I ran the asymmetric MVP — theory papers scanned against ML/applied-CS papers across the math/CS axis — on 2026-05-24 to measure what it does surface. Full run trace, costs, the calibration set with arxiv IDs, and the three surviving pairs are on a companion page.
Two thousand candidate pairs after a SPECTER2 nearest-neighbor cross-axis scan over math and CS arXiv categories; 95 surviving slot extraction; 3 surviving a cheap LLM judge; all 3 verdict-FOLKLORE on independent Opus review — textbook results or specialist common knowledge. Total external API spend: $0.20.
The density math is consistent with that outcome. arXiv has roughly half a million ML/applied-CS papers and half a million math/theory papers in the categories drawn from. Suppose ~10³ real bridge pairs exist in the joint corpus. A uniform sample of 800 empirical and 1377 theory papers contains both endpoints of an expected 0.004 of them — vanishingly few real bridges should be reachable at this scale. Nearest-neighbor selection would lift that yield only if the embedding preserved bridge geometry, which the Sawin calibration says it doesn't. When the selection criterion is orthogonal to the target, the cascade defaults to surfacing standard domain neighbors.
The deeper objection
Scaling the corpus from 2200 to 1.9M papers would fix the density problem. It wouldn't fix the geometry problem. If novel bridges manifest as new directions in representation space — axes that didn't previously exist, in the Singular Learning Theory framing where new knowledge appears as singularities resolved by “blowing up” — then no flat-embedding nearest-neighbor scan can surface them. The bridge isn't a point between two clusters; it's a direction. The Sawin calibration is consistent with both readings; the data here doesn't separate them.
What's next
Anchor-seeded sampling is the cheap discriminator. Pick five to ten known or suspected non-folklore bridges, take their SPECTER2 neighborhoods, run the asymmetric cascade against those neighborhoods. If non-anchor cousins surface at non-trivial rate, the geometry holds enough signal to justify the 1.9M-paper production build. If the cascade still returns folklore neighbors of the anchors, the SLT objection is doing real work and the mechanism needs to change before the corpus does.
Cost of the discriminator: about a dollar.