Score Aggregator

Overview

The Score Aggregator node reads all evaluation records for a battle — whether produced by a single judge_battle node or by multiple independent judge runs — and reduces the per-criterion, per-contender scores into a single aggregate score per contender. The resulting leaderboard is written back to the battle record in the database and is emitted downstream. The aggregation algorithm is controlled by aggregation_method, and the node enforces a minimum quorum via min_evaluations before producing results.

Configuration

Field	Type	Required	Default	Description
battle_id	string	Yes	—	UUID of the battle whose evaluations to aggregate. Supports template expressions.
aggregation_method	"mean" | "median" | "weighted_mean"	No	"mean"	Algorithm used to combine multiple scores into one.
weight_map	object	No	—	Maps criterion name to a numeric weight. Only used when aggregation_method is "weighted_mean". Weights are normalised to sum to 1 before application.
min_evaluations	integer	No	1	Minimum number of evaluation records required per contender. Contenders with fewer evaluations are excluded from the leaderboard and flagged in the output.

Inputs

Port	Type	Description
input	object	Optional upstream data. Accepts battle_id to override the static config. Can also accept the evaluations array emitted by a judge_battle node to avoid a redundant database read.

Outputs

Port	Type	Description
output	object	Contains battle_id, aggregated_at (ISO-8601 timestamp), leaderboard (array of objects sorted by total_score descending, each with contender_id, total_score, rank, and score_breakdown by criterion), and excluded_contenders (array of contender IDs that did not meet min_evaluations).

Example

{
  "nodeType": "score_aggregator",
  "config": {
    "battle_id": "{{nodes.judge_battle.output.battle_id}}",
    "aggregation_method": "weighted_mean",
    "weight_map": {
      "correctness": 0.5,
      "efficiency": 0.3,
      "readability": 0.2
    },
    "min_evaluations": 2
  }
}

Notes

• When using "weighted_mean", any criterion present in the evaluation records but absent from weight_map is assigned a weight of 0 and does not contribute to the total score. Log a warning appears in the run trace if this occurs.
• The node writes the finalised leaderboard to the battle_results table and updates the battle status to "scored". Downstream nodes such as series_advance read from this persisted record rather than the node's in-memory output.
• With "median" aggregation, ties in the final score are broken by the submission timestamp (earlier submission wins), preserving a stable ranking.
• If min_evaluations is greater than 1, this node is most useful in hybrid judging pipelines where multiple AI models or multiple human reviewers each produce a separate evaluation run.