A LoRA on Qwen3.5-9B closed 24 of the 27 points between the base model and Jev

The six specs describe one interface, built three ways. Laya adds two transformer layers on top of an encoder and scores the options a caller supplies ^[4]. Jevlike turns each candidate into a query that reads from a shared context representation and hands back a score ^[10]. SemIf hangs a tiny three-class NLI classifier on the last token of a Qwen3.5 backbone ^[9]. Encode the context once, score N supplied candidates against it, return a number for each. @hxiao argued this class of model could pull tool calling, routing and MCP-style decisions back from small generative LMs toward discriminative ones ^[15], and @ankrgyl said Jev is already available as an eval model in Braintrust at roughly 400x lower scoring cost than prior setups ^[16].

One reproduction published numbers. The distance between the stock checkpoint and Jev was 27 points on Bespoke Nimble's own curated eval, and the LoRA plus synthetic contrastive curation plus constrained decoding closed 24 of them, about 89 percent of the way ^[18]. The last three points are what staying closed still buys. For that 90 percent to mean anything on your traffic, your candidate sets and decision mix have to resemble the eval @madiator curated, and @abacaj noted there is no standard benchmark for the category and that the demos leaned on speed more than quality ^[13].

The harder substitution problem is calibration. @gabepereyra pointed at routing, citation selection, escalation and legal ops as the calibrated-probability use cases ^[14]. Laya emits probabilities from 0.0 to 1.0 ^[4], and latent.space reports its confidence as entropy-based, not calibrated ^[5]. Both scores sit on the same zero-to-one axis, so an escalation threshold fitted against one of them will fire at a different rate against the other. Refitting it means owning a labelled decision set.

latent.space says the data behind this work is acknowledged to be 100 percent synthetic, and that not enough people are talking about the data side ^[12]. If teams starting from unrelated architectures all land near the same score, the performance is coming from the curation and the objective, not the backbone. The record supporting that is one published eval table ^[8] plus the newsletter's own characterisation of DiffusionGemmaJev as pretty close on benchmarks ^[6] and Bespoke Nimble as slightly lower ^[7]. latent.space also says Laya's author claims RLCD without justification ^[17]. Five of the six reproductions have no published score in that account.

The cheap end is already local. Kev-0.5B is a LoRA adapter plus a small readout head on Qwen2.5-0.5B, and @jaredpalmer says it runs on a MacBook Pro ^[11]. Bespoke Nimble reports 100ms on an H100 and local usability ^[8]. For notification ranking or an escalation route, the 0.5B adapter is the one that runs on the laptop you already have.