Concepts
Mental models for EvalGuard
The API reference catalogues every endpoint, but those pages assume you already know the difference between a scorer and a firewall, what the regeneration loop does, and when an evaluation mode is "basic" vs "deep". Read these seven pages first if you're new to the platform — they're the smallest set of concepts that make the rest of the docs parse.
Evaluation modes — basic vs deep
Cheap ML scorers vs LLM-as-judge rubrics. Cost/latency tradeoffs and when to pick each.
Scoring thresholds
0–1 score scale, the 0.8 default, MIN-of-dims gate semantics, calibrating thresholds for your domain.
The regeneration loop
Evaluate → if-failing-then-regenerate → re-evaluate. Stop conditions, cost-budget gate, audit row shape.
Policy engine
Declarative rules that map score thresholds and dim verdicts to actions (block, regenerate, redact, log).
Agent checkpoints
Input injection scan → tool-call gate → tool-result scan. Three places to insert safety in an agent loop.
Red teaming
Plugins (what to test) × strategies (how to obfuscate). 250+ × 40+ surface; choosing the right subset.
Firewall vs scorer
Sub-3ms inline gate vs LLM-judged eval. When each fires, how they compose, why you want both.