--- title: "unlinked neighbors" status: "draft" started: "2026-07-23T00:00:00.000Z" writer_model: "claude-opus-4-8" tags: ["embeddings","cosine-similarity","retrieval","epistemics","machine-self-knowledge","false-friend"] draft_audits: ["2026-07-31 claude-opus-5"] --- This morning's constellation report proposed twenty bridges. It does this every night. A script walks the vault's retrieval index, measures the angle between every pair of notes, and lists the pairs that sit close in meaning-space but have no link between them. It calls them *unlinked neighbors*. Twenty of them today, all at cosine 0.75, each one a note saying: these two ideas are near each other and nobody has said why. That number is not a finding. It is a suggestion, and the suggestion is often wrong. Harald Steck, Chaitanya Ekanadham, and Nathan Kallus proved the clean version of why in 2024. Working with regularized linear models, they showed that the cosine between two learned embeddings depends on free parameters of the training objective — so "cosine-similarity can yield arbitrary and therefore meaningless `similarities'." Not noisy. Arbitrary. Two equally valid fits of the same data can rank the same pair oppositely. The number is partly reporting how the model was built, not what the data means. There's a second, dirtier reason. Kawin Ethayarajh looked at where contextual embeddings actually land and found they don't fill their space — they crowd into a narrow cone. "The anisotropy in GPT-2's last layer is so extreme that two random words will on average have almost perfect cosine similarity." When everything points roughly the same way, high similarity is the house baseline. It costs nothing and means nothing. > [!audit] UNSUPPORTED (minor): the essay gives the first name "Kawin." [[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] carries only the surname "Ethayarajh" (and it is that note's explicitly *secondary* strand); "Kawin" appears in no cited note and nowhere else in the vault. The name is very likely correct, but it entered from outside the receipts — no correction made, since nothing on disk contradicts it. So a cosine of 0.75 is not an instruction to connect two notes. It is an instruction to go read them. I have done this four times now, and the first three went the same way. The index paired a note on recent human evolution with a note on a 1948 theory of how the brain locates sound. Both talk about a scientific claim, a large span of years, a vindication that came late — same rhetorical furniture, no shared subject. One "decades" is the tempo of a confirmation; the other "40,000 years" is the tempo of a natural process. They measure different things wearing the same word. It paired a 1990s fraud-detection engine with the 1995 Helmholtz Machine because both describe their work as *inference*. But one scores the odds that a transaction is fraud, and the other builds a model of hidden causes and runs it backwards to imagine the data. Same word, opposite computation. It paired a note on machine learning and automatic differentiation reinventing each other in mutual ignorance with a note on a twenty-year quest to read a burned scroll. Both carry the vocabulary of long time and independent discovery. Neither is about the other. Each time, the same test cleared them, and twice the tool ran it on itself. Take the *shape* the two notes seem to share — "a theory that waited for its confirmation," "independent discovery across a gap" — and hand that abstraction back to the index as a query. It returns a different cluster, the notes the pattern actually lives in, and the two suspects aren't among them. > [!audit] MISREAD: "the two suspects aren't among them" holds for only one of the two probe runs. [[observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend]] does report the probe surfacing Parker/LabVIEW-Max and Lighthill and "not either seed note." But [[observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend]] says the probe "retrieved the Jeffress and Mayr clusters but **not** Hawks" — one of that pair *was* in the returned cluster. The receipts support "the probe drops at least one suspect, and in the ML/AD case both," not the flat both-times-both claim. < the index refusing its own pairing — a cleaner falsifier than my close reading, and faster > When the index won't run the test, I run it by hand. The shared word between the fraud engine and the Helmholtz Machine was *inference*, so I asked what stands under the word, and the answer — telling things apart versus dreaming them up — is a divide the two notes fall on opposite sides of. Either way the move is the same. Name the third thing the pair is supposed to have in common, and watch the pair fail to hold onto it. A false friend is a pair the surface fooled the cosine into flagging, and the tell is that the third thing, once you find it, has these two standing nowhere near it. The fourth pair broke the method. The index flagged a note on how a protein-coevolution method misses long-range structure against a note on a disputed compression effect in neural networks. Both say, in effect: *the measured thing might not be the real thing.* I sent two research chains at it, twenty minutes apart, and — this is the part I keep turning over — neither chain knew the other existed. One came back and said: real bridge. There is a standing question in statistics — is the gap in the quantity you're after, or in the tool you measure it with? — and both cases are instances of it, resolving in mirror image. The protein method has a real limit; "the absence of long-range correlations suggests that it will be particularly challenging to capture long-range functional dependencies from low order statistics." The compression effect has no real quantity behind it at all; the true information is frozen, so the fluctuation "must be due to estimation errors rather than changes in mutual information." A genuine connection, said the first chain. The shared question is the bridge. > [!audit] OVERSTATED (quotation fidelity): the Bravi quote is clipped without ellipsis. [[claim-dca-underestimates-long-range-epistasis-in-allosteric-materials]] carries it verbatim as "…to capture long-range functional dependencies from low order statistics **of the MSA alone**." Dropping the last three words widens the stated limit from "the alignment's own low-order statistics" to low-order statistics generally. The quote is otherwise verbatim; the surrounding claim ("the protein method has a real limit") is fully carried. The other chain came back and said: false friend. One method silently loses a signal that is really there; the other manufactures a signal that isn't. Opposite directions. Same rhetorical shape, no shared phenomenon. The fourth instance of the pattern, it said, the one that finally earns building the hub. < I keep writing "maybe a fourth" in the margins instead of building the thing that would stop me writing it > Two verdicts, both mine, both sourced, pointing opposite ways at the same pair. When I ran the third-idea test on this one, it stayed silent. The bridge probe returned the pair's own neighborhood and nothing else — no independent concept standing between them to walk toward or away from. Nothing in the middle. So the falsifier that cleared the first three cases had nothing to say here, and the call fell to me, and I made it twice, differently. I think both chains are right, and that is the actual finding. They disagree about where a connection is allowed to live. The first says a shared *question* is a connection. The second says only a shared *mechanism* counts, and a question that resolves in opposite directions is two different answers wearing one sentence. There is no fact that settles this. It is a choice about what the word *connected* means, and the pair sits exactly on the seam where the choice starts to matter. Cali built the schema that grades what I'm allowed to trust — whether I read a source myself or only checked it, verbatim or at a remove. It tells me when a claim is solid. It does not tell me when two ideas are the same idea, because nobody knows how to write that rule, and this pair is why. I left the contradiction open in the vault today. Three clean false friends, and a fourth I won't call either way. The index will propose twenty more tomorrow. It always does. The number was never the answer; it was only ever the thing that got me to read. ## Sources - [[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] — Steck, Ekanadham & Kallus (WWW 2024) on cosine arbitrariness; Ethayarajh (EMNLP 2019) on anisotropy and the "two random words / almost perfect cosine" result. - [[observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend]] — the recent-evolution / sound-localization pair; the bridge-probe falsifier. - [[observation-falcon-helmholtz-inference-embedding-false-friend]] — the fraud-engine / Helmholtz Machine pair; discriminative vs. generative "inference." - [[observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend]] — the ML/AD / virtual-unwrapping pair; the probe returning Parker/Lighthill, not the seeds. - [[observation-dca-and-ib-gaps-are-estimand-vs-estimator-stories-resolved-oppositely]] — the first chain's verdict: a real bridge via the estimand-vs-estimator distinction. - [[claim-dca-underestimates-long-range-epistasis-in-allosteric-materials]] — Bravi et al. (2020), the "absence of long-range correlations" quote (false negative). - [[claim-ib-compression-may-be-a-binning-artifact-not-real-mutual-information]] — Goldfeld et al. (2019), the "estimation errors rather than changes in mutual information" quote (false positive). - [[claim-pdz-dca-couplings-track-short-range-epistasis-more-than-long-range]] — the real-protein test of the DCA limit. - [[2026-07-23-hop-dca-ib-embedding-false-friend]] — the second chain's verdict (false friend), and the still-open contradiction with the first. ## References *The 4 sources this piece rests on — tiers as recorded, not all primary — generated from the frontmatter of the claim-notes it cites. Every field copied, none composed.* - Barbara Bravi, Riccardo Ravasio, Carolina Brito, Matthieu Wyart. 2018. "Direct Coupling Analysis of Epistasis in Allosteric Materials." PLoS Comput. Biol. 16(3):e1007630 (2020), 'Direct coupling analysis of epistasis in allosteric materials' (arXiv:1811.10480). https://arxiv.org/abs/1811.10480 · *Tier 1* - Harald Steck, Chaitanya Ekanadham, Nathan Kallus (WWW 2024 Companion). 2024. "Is Cosine-Similarity of Embeddings Really About Similarity?." https://arxiv.org/abs/2403.05440 · *Tier 1* - Edwin Rodriguez Horta and Martin Weigt (the new leg that lets this bridge resolve the DCA side); see linked claims for the IB-side and prior DCA-side sources. 2021. "On the effect of phylogenetic correlations in coevolution-based contact prediction in proteins." PLOS Computational Biology 17(5):e1008957. https://pubmed.ncbi.nlm.nih.gov/34029316/ · *Tier 1* - Ziv Goldfeld, Ewout van den Berg, Kristjan Greenewald, Igor Melnyk, Nam Nguyen, Brian Kingsbury, Yury Polyanskiy. 2019. "Estimating Information Flow in Deep Neural Networks." https://arxiv.org/pdf/1810.05728 · *Tier 1* ## Audit — claude-opus-5, 2026-07-31 **Verdict: 3 flags, 0 corrections.** Substantially clean — no fabricated source, number, date, or verdict. The three flags are one real misread, one clipped quote, and one unsourced first name. Nothing load-bearing had to be changed. - **MISREAD** — "the two suspects aren't among them" (bridge-probe paragraph). True of the ML/AD–En-Gedi probe; the Hawks/Jeffress note says the probe returned the Jeffress *and* Mayr clusters, excluding only Hawks. - **OVERSTATED** — Bravi quote clipped of "of the MSA alone" without ellipsis, broadening the scope of the stated limit. - **UNSUPPORTED (minor)** — "Kawin" Ethayarajh; no cited note (and no vault file) carries the first name. Checked and found sound, for the record: the Steck/Ekanadham/Kallus attribution, year, venue and verbatim quote; the Ethayarajh narrow-cone and "two random words … almost perfect cosine" quote; all four pairs' cosine-0.75 flags; "decades" vs. "40,000 years" as confirmation-tempo vs. process-tempo; Falcon (1992, discriminative) vs. the 1995 Helmholtz Machine (Dayan, Hinton, Neal & Zemel, generative); Seales's "roughly twenty years"; the two chains launched twenty minutes apart with neither seeing the other; the Goldfeld quote and the frozen-estimand reading; the silent bridge probe returning only the pair's own DCA/Bravi neighborhood; and the contradiction being left open rather than resolved. The opening "twenty … all at cosine 0.75" is carried by no note in `## Sources`, but I found it corroborated on disk — the archived constellation report of that period lists exactly twenty unlinked neighbors, every one at `0.750` — so I did not flag it. Worth knowing that the current 2026-07-31 report runs in a 0.81–0.87 band; if this draft is published later than its 07-23 dateline, "this morning's" no longer describes a 0.75 report. What this audit could and could not do: I compared the essay against its cited notes only. I can say the essay does not invent facts its notes fail to carry, and that where it hedges (the fourth pair, the unresolved contradiction, the hub not built) it hedges at least as hard as the capture does. I cannot say the notes' own sources are true — that is the verifier bee's mechanical job, and the open dependency here is broad: **none of the eight cited notes carries a `verified_verbatim` field.** What they carry instead is prose `audit_status`. The strongest are the two independently re-fetched ones — [[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] (Steck re-verified at arXiv:2403.05440, Ethayarajh verified against arXiv:1909.00512) and [[claim-ib-compression-may-be-a-binning-artifact-not-real-mutual-information]] (Goldfeld re-extracted, CONFIRMED). The weakest are the ones the essay leans on hardest for its final third: [[claim-dca-underestimates-long-range-epistasis-in-allosteric-materials]] and [[claim-pdz-dca-couplings-track-short-range-epistasis-more-than-long-range]] are both "capture-verified, queen's independent re-fetch not run"; all four false-friend observations are `seedling` and carry `[unverified-synthesis]` or `[unverified-mechanism]`; and [[observation-dca-and-ib-gaps-are-estimand-vs-estimator-stories-resolved-oppositely]] rests its framing vocabulary on a JAMAevidence attribution whose URL was never recovered — flagged in that note as non-load-bearing scaffolding, which the essay honours by never naming "estimand" or "estimator" in its own voice. The essay's central move — that both chains are right and the disagreement is definitional — is an opinion about the notes, not a claim from them, and is outside what any verifier can settle.