Does a third object-level false-friend note also cosine-match the Kelly-Luccioni/Garfield-Seglen-Steck confirmed-bridge note and resolve false friend by the same mechanism?
observation-kelly-luccioni-garfield-seglen-steck-is-a-cosine-similarity-retrieval-magnet argues, from two data points (Olsder-Greenhalgh on 2026-09-17, Falcon-Helmholtz on 2026-09-29), that observation-kelly-luccioni-garfield-seglen-steck-cosine-pairing-is-confirmed-bridge may itself be an unusually retrieval-hot note — one that cosine-matches any object-level "embedding false friend" instance at ~0.87–0.88 because of its own dense self-describing vocabulary ("cosine," "confirmed bridge," "new species," "unlinked neighbors"), rather than because of anything specific about either object-level partner.
Two data points is a candidate pattern, not a confirmed property. What would settle it: a cosine check between the Kelly-Luccioni/Garfield-Seglen-Steck note and a third, independent object-level false-friend instance not yet paired against it. The 2026-09-29 capture that raised this question named several untested candidates: observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend, observation-gates-jevons-cosine-pairing-is-embedding-false-friend, observation-liang-1979-olsder-1975-cosine-pairing-is-embedding-false-friend, observation-ach-abduction-lucky-widrow-cosine-pairing-is-embedding-false-friend, observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend, observation-nips-founders-kwakernaak-twente-founding-cosine-pairing-is-embedding-false-friend, and observation-fujitsu-fairchild-toshiba-kongsberg-cosine-pairing-is-embedding-false-friend — any one of them, checked directly against Kelly-Luccioni/Garfield-Seglen-Steck, would serve.
What's needed: a working mcp__seek__vault_bridge or vault_novelty call that surfaces a direct pairwise cosine score between the target note and a chosen candidate, rather than taking a value as given from a task framing. The 2026-09-29 session found the tool itself loading and running (an improvement over the standing permission-error gap logged 2026-08-22 in 00-meta/seek-flags.md) but not returning a pairwise score for either note in its top-5 nearest-neighbor output — so even a working tool call may need a more targeted query shape than "nearest neighbors of this title" to answer this specific question. If a third instance lands at a comparably high cosine and resolves false-friend by the same recursion-level/asymmetric-grounding test, the "retrieval magnet" claim graduates from candidate to established, mirroring question-verify-puckette-petiska-retrieval-magnet-third-instance's parallel check on the false-friend side of the taxonomy. If it does not, the two-data-point pattern was coincidence and the claim should be corrected in place per the operating spec's correction-history convention.
Next move: once a vault_bridge/vault_novelty call that returns direct pairwise scores is available, run it against one or more of the candidates named above.
Progress — 2026-09-30 (promotion of 10-inbox/raw/2026-09-30-does-a-third-object-level-false-friend-note.md, headless): partial only; question stays open. This session found vault_bridge unblocked (unlike the standing permission-error gap several sibling sessions hit) and ran all eight calls the question calls for — the confirmed-bridge note's own title plus each of the seven named candidates', each queried for top-5 nearest neighbors. Result: negative in both directions, recorded in observation-kelly-luccioni-garfield-seglen-steck-seven-candidate-sweep-is-negative. What did NOT resolve: the tool still returns only per-topic top-5 lists, not a direct pairwise score for two named notes, so the negative is an absence from mutual shortlists rather than a measured sub-threshold cosine for any specific pair. The retrieval-magnet hypothesis stays at two data points. This question closes only when a vault_bridge/vault_novelty call shaped to return a direct pairwise score becomes available — see the [defect] entry appended to 00-meta/seek-flags.md 2026-09-30 — or when a candidate outside the seven named here is swept and matches.
Progress — 2026-10-01 (Cali's review, via Claude Opus 5.5): the pairwise query this question was waiting for now exists — mcp__seek__vault_pair (seek-code df62bcc), document-to-document cosine on the same basis as constellation's seed pairs. Run on a copy of the live index (2026-09-30 evening): the two confirmed instances score 0.9495 (Olsder-Greenhalgh meta-pairing, rank 2 in KL-GSS's neighbour list) and 0.9464 (Falcon-Helmholtz meta-pairing, rank 5); the seven candidates score Gates-Jevons 0.9187 (rank 16), Hawks-Jeffress 0.9135 (17), Brooks-Werbos 0.9061 (21), ACH-Widrow 0.8906 (32), NIPS-Kwakernaak 0.8734 (43), Liang-Olsder 0.8664 (46), Fujitsu-Toshiba 0.8236 (75), of 2,669 indexed notes. So the negative holds relative to the instances (no candidate sits as close as either confirmed one, and none is in KL-GSS's top 15), but none of these is a low cosine in absolute terms — this whole lineage lives in a dense pocket of the index. Whether that settles the question (and what "retrieval magnet" should mean against that density) is yours to call; left open.