Model-free online evals (CEVAL): how it works and when to use it
Model-free online evals (CEVAL), grades output with deterministic evaluators (toxicity, PII, relevance, JSON validity and more) and tracks the results over time, no LLM-judge. Here's how Crowkis does it and why it matters for cost and safety.
Production LLM traffic is deeply repetitive, and repetition is exactly what a bill is made of. Model-free online evals (CEVAL) is how Crowkis grades output with deterministic evaluators (toxicity, PII, relevance, JSON validity and more) and tracks the results over time, no LLM-judge.
How it works
Crowkis grades output with deterministic evaluators (toxicity, PII, relevance, JSON validity and more) and tracks the results over time, no LLM-judge. It runs inside one Redis-compatible engine, so it composes with semantic caching, agent memory, and the other intelligence layers instead of being a separate service you wire together.
CEVAL relevance "q?" "answer" THRESHOLD 0.7
Why it matters
Repetitive LLM workloads are where the money is, and semantic caching can cut costs up to 60-70% on repetitive workloads. Model-free online evals (CEVAL) is part of what makes that reuse safe rather than reckless, the difference between a cache you trust in production and one you audit after every incident. It's one self-hosted binary, Redis-compatible, free to run.
Infrastructure earns the critical path one boring, verifiable feature at a time.