adk_eval_metric behaviour (erlang_adk v0.7.0)

View Source

Metric and judge adapter behaviour for deterministic evaluation.

A judge backed by an LLM implements the same callback but declares kind => judge in its descriptor. The evaluation engine bounds it within the containing case worker's heap and absolute deadline.

Summary

Types

score/0

Callbacks

score/4

Types

score/0

-type score() :: number() | {ok, number()} | {ok, number(), map()} | {error, term()}.

Callbacks

score/4

-callback score(Expected :: term(), Actual :: term(), Context :: map(), Config :: map()) -> score().