Traverse from the semiconductor business in a graph where Samsung Electronics and Samsung SDI have been collapsed into one node, and battery-supply disruptions come back as a result [2]. The traversal succeeds. Nothing in the response flags it, and the author's argument is that you cannot spot it without already knowing the answer you went to the graph to find [3]. The other failure leaves "Samsung Electronics" and "the largest memory chipmaker in the world" as separate nodes that accumulate edges independently, so a traversal from one never reaches the other and you get a partial answer instead of a wrong one [4]. Those two costs are not the same size, which is the stated reason the cutoff sits high [5].
The threshold only adjudicates part of the traffic. A mention already in the alias table resolves by lookup, with no model call, and is written under the existing canonical ID [12]. A mention the model scores as a probable match gets its surface form written into the registry, so the next article using that string also skips the model [13]. Anything without a confident match becomes a new canonical entry holding one mention and no resolved aliases [14]. The alias table is a cache, and the merge decision is its write path.
So the registry counts are worth arithmetic. 47,853 resolved entities against 47,883 aliases [15] is a gap of 30 [20], or 1.0006 aliases per entity [22]. The post does not say whether an entity's own canonical surface form occupies an alias row. If it does, then about 30 entities in nearly 48,000 have ever learned a second name. If it does not, nearly every entity has exactly one alias and 30 have two. Both readings describe a table that has consolidated very little, which is the shape a precision-first cutoff is supposed to have [6].
Blocking narrows it further. The sample request carries four candidate strings [8], and the response shows two scored pairs [9][10]. Four candidates make six unordered pairs, so four never appear at all [21]. They were dropped by the blocking step that DuckDB runs ahead of the full model [11][17]. Downstream, a pair absent from the response and a pair scored 0.12 [10] are the same thing: no edge. The recall ceiling is set by blocking rules, before the threshold gets a vote.
The cutoff itself is described only as a high confidence mark, with no number printed [23]. The two example scores are 0.96 and 0.12 [9][10], which is the easy region; nothing near the boundary is shown. A caller receiving the `decision` field cannot recompute it.
For this to transfer, your consumers have to resemble his: a person reading a small number of traversals to judge which news matters [6]. A missing edge is loud to a human who expected it. If instead a downstream system acts on every edge and nobody reads them, silence is cheap too, and the asymmetry narrows.
