The ML/AD-unaware <-> En-Gedi cosine-0.75 pairing is another embedding false friend

The vault flagged claim-ml-and-ad-communities-mutually-unaware and claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll as unlinked neighbors at cosine 0.75. On inspection the proximity does not hold as a genuine bridge — it is the vault's now-familiar embedding false friend.

The ML/AD claim: "Until very recently, the fields of machine learning and AD have largely been unaware of each other and, in some cases, have independently discovered each other's results." (Baydin et al., JMLR 2018, Tier 1). That is a claim about disciplinary non-communication: two academic communities maintained separate citation graphs and independently reinvented the same technique. The En-Gedi note's arc is different in kind — Brent Seales's virtual unwrapping was developed by one researcher over roughly 20 years before its first legible result; no second community was independently reinventing the technique in parallel. A check of Seales's own history and Youssef Nader's Vesuvius Challenge ink-detection writeup found no "two fields didn't know about each other" story on the archaeology side — ML people joined openly, by invitation, during the 2023 crowdsourced contest.

The vault's own retrieval geometry agrees: probing the abstracted "mutual unawareness / independent discovery" pattern surfaces the already-linked Parker/LabVIEW-Max and Lighthill cluster inside moc-backpropagation-origins — not either seed note. The 0.75 is shared temporal-arc vocabulary ("years," "history," "independently," "before... later"), the exact failure mode claim-cosine-similarity-of-embeddings-can-be-arbitrary describes, already logged twice: observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend and observation-falcon-helmholtz-inference-embedding-false-friend. No wikilink added between the two seed notes.

Why this was hop-worthy

The seed explicitly asked for a bridge verdict, and the computable assists (vault_bridge, vault_novelty) both independently confirmed the false-friend read rather than just my own judgment.

Further leads

• Consider a moc-embedding-false-friends hub to collect this recurring pattern (now 3+ instances) instead of re-discovering it per chain.
• Youssef Nader's TimeSformer-based ink-detection pipeline is an unexplored mechanism hook.

Entity candidates

• (none new — Baydin, Seales, Nader all already have vault presence)

Hop chain

Hop 1: claim-ml-and-ad-communities-mutually-unaware & claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll (vault notes, cosine 0.75, unlinked)

• Hook type: cross-domain bridge (per seed assignment)
• Hook: two unlinked notes at 0.75 cosine spanning ML/AD historiography and biblical-scroll imaging
• Why followed: seed instruction — investigate whether the bridge is real
• Key findings: both notes read in full; surface resemblance (temporal arc, "years," "independently," "later formalized") apparent immediately, mechanism identity not yet tested

Hop 2: mcp__seek__vault_bridge probe, topic = "communities mutually unaware... independent discovery"

• Hook type: mechanism question (testing the bridge computationally)
• Hook: does the vault's own retrieval geometry place either seed note near the abstracted pattern it supposedly instantiates?
• Why followed: the computable-assist amendment for bridge checks
• Key findings: surfaced Parker/LabVIEW-Max (0.684), Barlow Twins (0.684), Lighthill 1973 (0.675) — neither seed note appears in the top-5 at all; bridge_candidate=true, but for a cluster that has nothing to do with En-Gedi
• Surprise: expected the tool to at least weakly implicate the En-Gedi note — found it absent entirely from the top-5 for the pattern it was supposedly an instance of

Hop 3: claim-lighthill-1973-named-general-problem-solving-disappointing & observation-parker-and-labview-max-are-opposite-poles-of-multiple-discovery (vault notes)

• Hook type: unfamiliar name / existing-cluster check
• Hook: are these already the vault's real "mutual unawareness" cluster?
• Why followed: confirm the tool's redirection was meaningful, not noise
• Key findings: yes — an already-elaborated multiple-discovery cluster sits under moc-backpropagation-origins with its own Merton-vs-Chandler framing; the ML/AD note belongs to a family the vault has worked hard on, just not paired with En-Gedi

Hop 4: web research — Brent Seales / virtual-unwrapping history (UKNow) and Youssef Nader's "The ink detection Journey of the Vesuvius Challenge"

• Hook type: surprising claim (testing for a hidden two-communities narrative)
• Hook: does the archaeology side have its own "mutual unawareness" story mirroring the ML/AD one?
• Why followed: only an object-level check, not embeddings, can confirm or refute the bridge
• Key findings: no — it's a single researcher's 20-year solo program; ML/computer-vision people joined openly and were credited explicitly during the 2023 contest, not an independent rediscovery
• Surprise: expected a real mechanism connection (ink-detection CNNs are trained via backprop, so autodiff literally trains the models solving En-Gedi's sequel problem) — found that link, while true, is too generic (basically all deep learning uses backprop) to be the specific "mutual unawareness" bridge the seed asked about

Hop 5: mcp__seek__vault_novelty on the false-friend verdict itself

• Hook type: mechanism question / self-reference
• Hook: is this diagnosis itself novel to the vault?
• Why followed: mandatory gate check before capture
• Key findings: max_cosine 0.764 against observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend and observation-falcon-helmholtz-inference-embedding-false-friend — this is at least the vault's third documented instance of the identical failure mode
• Surprise: expected to be reporting a novel meta-finding — found the vault had already named and logged this exact phenomenon twice before, with its own prior commentary predicting a "fourth or fifth" case

Saved hooks not followed:

• Nat Friedman & Daniel Gross's funding origin story for the Vesuvius Challenge — from Scientific American — reason saved: genuine person-behind-the-thing / cross-domain-bridge-landing-on-AI hook, captured separately (see 2026-07-18-hop-nat-friedman-vesuvius-challenge-origin.md)
• A moc-embedding-false-friends hub note — from this chain's own gate check — reason saved: infrastructure decision for the queen, not a hop

post-worthy: maybe — a clean negative result that also surfaces a recurring pattern (this vault keeps producing embedding false friends on delayed-recognition narratives) worth the queen's attention for possible MOC consolidation, though it is the vault's Nth instance of the same diagnosis rather than a fresh one.