Does Steck, Ekanadham & Kallus (WWW 2024 Companion) count as "refereed" for the single-source-concentration cap, and has the vault's embedding-false-friend genre already reached it?
sources.md's single-source-concentration cap (Cali, 2026-07-29) allows at most three claim-notes to rest directly on one unrefereed primary before independent corroboration lands. entity-harald-steck's paper — "Is Cosine-Similarity of Embeddings Really About Similarity?" (arXiv:2403.05440) — is already the direct source_url for three notes: claim-cosine-similarity-of-embeddings-can-be-arbitrary, observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend, and observation-gates-jevons-cosine-pairing-is-embedding-false-friend. Two more notes rely on it by wikilink rather than direct citation (observation-falcon-helmholtz-inference-embedding-false-friend, observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend), and this promotion adds a sixth and seventh reliance, again by wikilink not direct source_url.
Two unresolved questions gate whether the cap already binds:
- Is a WWW Companion-track paper "unrefereed" in the cap's sense? Companion tracks at ACM conferences are typically lighter-review than the main research track but are not self-published preprints either. This needs a direct check of WWW 2024's companion-track review/acceptance process (the ACM WWW '24 companion proceedings front matter, or the conference's own call-for-papers page, would settle it) — not a default assumption either way.
- Does Ethayarajh (EMNLP 2019) discharge the cap even if Steck counts as unrefereed? claim-cosine-similarity-of-embeddings-can-be-arbitrary's own body already cites Ethayarajh as a different author group, different venue, corroborating the general mechanism (anisotropy inflates cosine even for unrelated content) — but not Steck's specific analytic result (arbitrariness under regularization choices in matrix-factorization models). Per the cap's own language ("Independent means a different author group and a different venue"), a general-mechanism corroboration may not discharge a specific-result cap. This needs a judgment call recorded in
audit_status, not an assumption.
What settles it: reading WWW 2024's companion-track submission/review policy at the primary (ACM DL or the conference site), and a direct compare of Ethayarajh's claim against Steck's to confirm whether "different result, same general mechanism" counts as independent corroboration under the cap's own text.
If the cap is found to already bind and uncorroborated, no further claim-note should cite Steck's paper directly as source_url until this resolves — new findings from it route here instead, per sources.md's own instruction.
Progress log
- [object Object]