Judge Battle

Overview

The Judge Battle node retrieves every accepted submission for the specified battle and sends each one to an AI judge model for evaluation against a defined set of scoring criteria. Each criterion is scored independently on a numeric scale and the raw per-criterion scores are emitted for downstream aggregation. When parallel_evaluation is enabled, all submissions are evaluated concurrently to minimise wall-clock time. The node transitions the battle's status to judging at the start and to judged upon successful completion.

Configuration

Field	Type	Required	Default	Description
battle_id	string	Yes	—	UUID of the battle to judge. Can be a static value or a template expression referencing an upstream node's output (e.g. {{nodes.battle_create.output.battle_id}}).
model_key	string	Yes	—	Identifier of the judge model to use (e.g. "claude-sonnet-4-6", "gpt-4o"). Must be a model available in the workspace's AI provider settings.
criteria	array of objects	Yes	—	Scoring dimensions. Each element must have a name (string) and a description (string) that is passed verbatim to the judge.
max_score_per_criterion	integer	No	10	Upper bound for each criterion's score. The judge is instructed to produce a score in [0, max_score_per_criterion].
parallel_evaluation	boolean	No	false	When true, all submissions are evaluated concurrently. Increases throughput but also increases token consumption and may hit rate limits for large battles.

Inputs

Port	Type	Description
input	object	Optional upstream data. A battle_id key here overrides the static config value.

Outputs

Port	Type	Description
output	object	Contains battle_id, judged_at (ISO-8601 timestamp), and evaluations — an array of objects, one per submission, each with contender_id, submission_id, and scores (an object mapping criterion name to numeric score).

Example

{
  "nodeType": "judge_battle",
  "config": {
::: v-pre
    "battle_id": "{{nodes.battle_create.output.battle_id}}",
:::
    "model_key": "claude-sonnet-4-6",
    "criteria": [
      {
        "name": "correctness",
        "description": "Does the solution produce the correct output for all standard inputs?"
      },
      {
        "name": "efficiency",
        "description": "How efficiently does the solution use time and memory relative to optimal?"
      },
      {
        "name": "readability",
        "description": "Is the code clean, well-named, and easy to understand without comments?"
      }
    ],
    "max_score_per_criterion": 10,
    "parallel_evaluation": true
  }
}

Notes

• Each criterion description is injected directly into the judge's system prompt; write descriptions that are unambiguous and self-contained so the model does not need to infer intent.
• parallel_evaluation: true can significantly reduce latency for battles with many submissions, but each parallel call counts independently against the workspace's AI provider rate limits.
• If the battle has no accepted submissions when this node executes, the node succeeds and emits evaluations: []; downstream nodes should handle this empty case to avoid errors.
• BYOK (Bring Your Own Key) workspaces resolve model_key against their own provider credentials; if the key does not match a configured provider, the node fails immediately with a configuration error.