Laya: an open, non-generative decision model pitched against TypeSafe's Jev

A developer at Convai Innovations is releasing Laya, an Apache-2.0 family of “System 1” decision models, framed partly as a priority claim against TypeSafe AI’s recently launched Jev. The core argument: most production AI pipelines waste generative LLMs on reflex classification tasks — routing tickets, flagging phishing or jailbreak attempts, scoring urgency — that need only a structured label. Instead of streaming tokens for 500ms–2s and then parsing free-form output, Laya runs bidirectional encoders in a single forward pass (~33ms on one GPU, 7.2ms batched) and emits only probabilities over three typed primitives: a categorical choice, an ordinal score, and a boolean (noul) with calibrated P(true). Because the output space is numeric, the model cannot hallucinate text or produce malformed JSON, and its confidence scores are meant to be mathematically calibrated rather than token-predicted.

The most substantive technical finding concerns multilingual routing. On a 51-language MASSIVE sweep, an English-vocabulary model (ModernBERT-large) collapsed on non-Latin scripts — 0.000 accuracy on Khmer, near coin-flip on Armenian and Hebrew — while still reporting 88–96% confidence. The lesson: confidence gating gives no warning when a model literally cannot read the input, so model selection must happen before inference. Laya ships a sub-millisecond Unicode-script router across 22 alphabets that adds under 2% overhead versus the forward pass, plus a preload mode to avoid multi-second cold swaps. Benchmarks cited include 0.993 accuracy on Enron spam and 0.980 on phishing, with weaker held-out jailbreak detection (0.755–0.762, rising to 0.931 at 50% selective coverage).

The significance is twofold. Technically, small calibrated encoders are a credible, cheap alternative to LLM-as-classifier for high-volume gatekeeping — including security-relevant filtering — where latency, cost, and honest uncertainty matter more than generation. Rhetorically, the post is also a familiar open-vs-closed grievance: the author contrasts published papers, open weights, and datasets against a well-funded frontier lab (founded by a ChatGPT co-inventor) shipping a similar concept with none of those, and claims 6–8x speed advantage. Readers should note the head-to-head numbers mix Laya’s own measurements with third-party figures for Jev, so the comparison is self-reported rather than independently reproduced.