Threshold constants are where an embedding model's geometry leaks into application code. A cutoff that reads as "same story" in one vector space reads as "same broad subject" in another, because the two spaces were trained to different objectives. RSSMonster stacks three decisions on top of those numbers: events group articles covering one real-world story, topics link related events into broader subjects, and interest islands are longer-lived clusters inferred from what a user tends to read [10]. Each layer consumes the layer beneath it. A looser event boundary does not stay at the event layer, which is why a change that looks helpful at the bottom of the stack shows up as noise at the top [11].
Both models passed the existing semantic regression suite while the pipeline behaved differently [7], and that combination tells you what the suite actually tests. It checks that embeddings are well formed and stable enough to pass fixed assertions. Nothing in it asserts that the count of event and topic associations lands in the same place. The diagnostic fixture is 23 hand-labelled article-like inputs against 23 taxonomy candidates [8], so at most 529 input-candidate pairings [18], with close neighbours such as Football, Tech and F1 included deliberately. That is a good confusion harness. Cluster-count drift is a different measurement, and a fixture that size will not surface it. Codex was handed the two generated log files and asked to identify the differences and explain what they meant for the application [14].
The money is worth stating in plain arithmetic. Around EUR 10 a month [3] is about EUR 120 a year [17]. That is the entire budget for the migration, and the migration does not end with one commit: similarity thresholds and fallback behaviour are now model-specific values rather than shared defaults [13], which is a second set of constants to keep calibrated whenever either model moves. At any plausible hourly rate, the first afternoon of tuning has already outspent the year. The reason to do it anyway is the one the author gives, which is whether commercial inference should be the default for a self-hosted reader at all [3].
On the question of where small models are good enough, this experiment measures one thing. The benchmarked comparison is the embedding process [16]. ModernBERT's jobs, information quality, promotional content and emotional tone, arrive with a role description and no reported accuracy against the hosted baseline they replaced [6][16]. The author reports that Qwen performed better than expected [19], and also that no model clearly won the embedding benchmark [9], which is the most useful kind of result and the least quotable.
What transfers here is the location of the work: a per-provider threshold table, and a downstream check that counts events and topics before and after the swap. Any pipeline that reads similarity numbers as fixed application constants is running the same experiment whether or not it is measuring it.
