Skip to content
Jevstiller

How Jevstiller compares with stuntd, TRACER, Distil Labs, routers, caches and open Jev models

As of October 2026. The first three have a detailed page each.

  • stuntd is the closest project. It is also a local Jev-compatible proxy that learns a head per question (on the Laya encoder or a stock sentence encoder) and checks 2% of live traffic. The differences (in full):
    • Guarantee: Jevstiller picks its threshold with a finite-sample bound on disagreement over all requests; stuntd uses a point estimate on a holdout.
    • Automation: Jevstiller trains, shadow-tests and promotes by itself; stuntd uses stuntd train / stuntd enable by default, with opt-in auto_retrain and auto_promote settings.
    • Training data: Jevstiller learns from Jev’s full probability distributions and gates unfamiliar inputs.
    • Question identity: Jevstiller identifies a question by its exact content; stuntd uses its name.
    • Keys: Jevstiller answers locally only for API keys Jev has accepted.
    • Deployment: Jevstiller serves many tenants from one server, on CPU.
    • Where stuntd goes further: it also speaks the OpenAI and Anthropic APIs, learns score questions, and can answer with no provider at all (zero-shot Laya).
  • TRACER (MIT, arXiv 2604.14531) is a Python library that fits a classical-ML student on your LLM traces with a learned acceptance policy and certifies it once with a Clopper–Pearson lower bound on agreement among the requests it accepts. It takes any teacher and is not a proxy: your application embeds, routes and falls back. Jevstiller bounds disagreement over all requests, re-certifies on every retrain, and keeps a 2% audit on live traffic (in full).
  • Distil Labs and cloud “distillation” (Amazon Bedrock, OpenAI, Azure) post-train a small replacement model from your traffic as a separate job, then switch traffic to it. Distil Labs’ endpoint falls back to your production model when the small one fails or times out, not when it is unsure; none of them states a bound, and none speaks Jev’s API (in full).
  • Routers (RouteLLM, Not Diamond, OpenRouter Auto) choose between existing models. Semantic caches (GPTCache, Portkey, jevcache) reuse answers to near-identical inputs. Neither learns to answer new inputs.
  • Open Jev-compatible models (Laya, Kev, jeff) replace Jev outright, at lower zero-shot accuracy.
  • Research:
    • OCaTS (EMNLP 2023), Cache & Distil (ACL 2024) and Online Cascade Learning (ICML 2024) train a student online from an LLM’s answers, without a guarantee.
    • BARGAIN (SIGMOD 2026) and vCache (ICLR 2026) guarantee agreement with the LLM, but without a student that keeps learning.
    • Jevstiller combines the two, with a permanent audit and automatic fallback on top.