A judge reads your conversation against a scenario’s scoring facts. Choose a judge to compare results across models while keeping the scenario and conversation fixed.

​
List supported models

Call GET /api/v1/models with your Labs bearer token. The response lists exact model identifiers, display names, and the default model for facts without a calibrated detector. The list includes active chat models available through Amazon Web Services (AWS) Bedrock, including Claude, Nova, DeepSeek, Gemma, Llama, MiniMax, Mistral, Kimi, Nemotron, OpenAI, Qwen, Writer, Grok, and GLM. Audio, video-only, embedding, reranking, safety-only, and legacy models are excluded. Model access still depends on the deployment’s AWS permissions and quotas. Use the returned id exactly. Short aliases such as opus are rejected.

​
Save an organization default

In the Labs portal, open Settings, find Judge model, choose a model, and click Save judge model. The searchable list uses the same supported model identifiers as the evaluation endpoints. The default applies to your organization’s evaluation, batch evaluation, comparison, and episode-turn requests when they omit a judge. A body field or header takes precedence over the saved default. The model list marks your saved model as the default when one is set. Choose Use existing calibrated judges and save to clear the organization default. Labs then uses each active trusted detector’s model and the existing platform default for remaining facts. Saved model choices require scenarios with fact-based scoring, just like explicit request choices. A scenario without fact-based scoring returns a 400 error when an organization default or explicit judge is applied.

​
Select a model for one request

Set judge_model in the request body, or send the X-Tacit-Judge-Model header. For example, use claude-fable-5-1, bedrock-qwen3-32b, or bedrock-gpt-6.1-sol. When you supply both fields, their values must match. Successful responses echo the effective request or organization choice in judge_model. For episodes, an explicit choice applies to that turn; later turns use the organization default unless they provide another choice.

​
Calibration and defaults

An explicit choice uses active calibrated judge detectors only when their model matches the requested identifier. Other judged facts use the chosen model without calibration. Deterministic calculations and persona self-reports keep their existing bindings. Selecting a model does not calibrate it or establish its accuracy. If you omit the choice and have no organization default, Labs keeps its existing behavior: active trusted detectors use their pinned models, and remaining facts use the existing default judge. Model selection requires a scenario with fact-based scoring. Other scenarios return a 400 error when you select a judge. Batch evaluation reports that error on the affected item. Invalid model identifiers and conflicting body/header values return 400 before evaluation starts.