---
id: "20260720-0900-dup-risk-ml-ad-en-gedi-bridge"
title: "Seed bridge (ML/AD-unaware <-> En-Gedi) is already a twice-logged embedding false friend — no new capture written"
type: "capture"
status: "promoted"
promoted_to: []
not_promoted: ["The only claim this capture carries — that the ml-ad-en-gedi cosine-0.75 pairing is an embedding false friend — is not new: it is already promoted, verbatim on the same underlying source, as 30-notes/observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend.md (derived from the sibling 2026-07-18 capture). Retrieve-before-write against 30-notes/ confirms the collision directly; this capture's own body already names the collision and declines to write a duplicate. Writing a fourth near-identical embedding-false-friend observation note (after hawks-jeffress, falcon-helmholtz, and ml-ad-en-gedi) was judged not worth it — three is the count the existing note's own commentary already weighed and chose to hold at ('a hub built on three notes is a hub built on hope'). Nothing here clears the bar of atomic + genuinely distinct + worth a permanent place; the capture's sole value is as a duplicate-avoidance record, which it already is standing in the inbox as-is."]
origin: "hop-batch"
writer_model: "claude-sonnet-5"
date_created: "2026-07-20T00:00:00.000Z"
hop_chain: ["seed: claim-ml-and-ad-communities-mutually-unaware <-> claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll (cosine 0.75, unlinked)","seed -> grep across vault for prior work on this exact pair (found 10-inbox/raw/2026-07-18-hop-embedding-false-friend-ml-ad-en-gedi.md)","prior capture -> vault_novelty re-check of the false-friend verdict itself (max_cosine 0.759)"]
novelty_max_cosine: 0.759
tags: ["embeddings","cosine-similarity","embedding-false-friend","en-gedi","virtual-unwrapping","automatic-differentiation","duplicate-check"]
source_url: "https://arxiv.org/abs/1502.05767"
source_author: "Baydin, Pearlmutter, Radul, Siskind"
source_date: 2018
source_venue: "JMLR 18(153)"
source_tier: 1
---


Subject: whether [[claim-ml-and-ad-communities-mutually-unaware]] and
[[claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll]] (cosine
0.75, unlinked) share a genuine bridge. Investigating independently — before
running any web research — surfaced that this exact pairing was already
worked twice on 2026-07-18: `10-inbox/raw/2026-07-18-hop-embedding-false-friend-ml-ad-en-gedi.md`
ran a full 5-hop chain (including the same "ink-detection CNNs are trained
via backprop/autodiff" mechanism angle this session independently arrived
at) and concluded false friend — shared temporal-arc vocabulary ("years,"
"independently," "before...later"), not a shared mechanism or community
structure. A sibling capture the same day
(`2026-07-18-hop-nat-friedman-vesuvius-challenge-origin.md`) branched off the
same dead end into a real finding about the Vesuvius Challenge's funders.
Re-running `vault_novelty` on the false-friend verdict itself today returns
max_cosine 0.759 against the two existing observation notes
[[observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend]] and
[[observation-falcon-helmholtz-inference-embedding-false-friend]] — this
would be at least a third instance of an already-named, already-logged
pattern. Writing a third near-identical capture was judged not worth it;
this stub records the check and points forward. Session continued from the
2026-07-18 captures' own "Further leads" (HAVI, unfollowed) instead — see
`2026-07-20-hop-godel-gabelsberger-havi-ai.md`.

## 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: the seed pairing itself
- Why followed: seed instruction — investigate whether the bridge is real
- Key findings: both notes reread in full; resemblance is narrative/temporal-arc surface ("years," "independently," "recently")

Hop 2: grep over /30-notes/ and /10-inbox/raw/ for prior work on this pair
- Hook type: mechanism question (self-check before duplicating work)
- Hook: has this exact question already been answered in the vault?
- Why followed: standard duplicate-avoidance before spending a full chain re-deriving a known answer
- Key findings: found `2026-07-18-hop-embedding-false-friend-ml-ad-en-gedi.md`, a complete prior 5-hop investigation reaching the same false-friend verdict, including the same "ink detection trained via backprop" angle explicitly considered and rejected as too generic

Hop 3: mcp__seek__vault_novelty on the false-friend verdict wording
- Hook type: mechanism question / gate check
- Hook: is re-stating this verdict itself now novel to the vault?
- Why followed: mandatory gate check before deciding whether to capture
- Key findings: max_cosine 0.759 against two existing embedding-false-friend observation notes — adjacent, not duplicate-risk by the strict >0.80 threshold, but functionally a third instance of a named, already-logged pattern

post-worthy: no — a confirmed non-bridge already captured twice; this stub exists only to prevent a fourth re-derivation next session.