Interface: LLMJudgeConfig

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:19

Configuration for LLM Judge

Properties

effort?

optional effort: "low" | "medium" | "high" | "xhigh" | "max"

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:31

Reasoning effort. When omitted, the resolver's default (max) applies only when the resolver's model was also selected — a pinned modelId gets no injected effort (zero request-shape change).

---

errorScore?

optional errorScore: number

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:42

Score returned when evaluation fails (provider error, parse failure). Defaults to 0 (fail-closed) so a judge used as a safety gate does not pass on infrastructure failure — matching EmergentJudge. Set to 0.5 to restore the prior neutral-on-error behavior for non-gating quality scoring.

---

llmProvider

llmProvider: AIModelProviderManager

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:21

LLM provider manager

---

modelId?

optional modelId: string

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:23

Model to use for judging.

Default

the central judge resolver (gpt-5.6)

---

providerId?

optional providerId: string

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:25

Provider ID. Pinning a non-openai provider requires an explicit modelId.

---

systemPrompt?

optional systemPrompt: string

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:35

Custom system prompt for the judge

---

temperature?

optional temperature: number

Defined in: packages/agentos/src/safety/evaluation/LLMJudge.ts:33

Temperature for judging (lower = more consistent)