Excessibility.Review.Judge behaviour (Excessibility v0.15.0)

Turns a Excessibility.Review change into a release-risk verdict.

A judge takes one view's change (regions + newly-introduced findings) and returns a tier, a human-readable blast-radius summary, and structured risks. Two judges ship:

Excessibility.Review.Judge.Heuristic (default) — transparent, no dependencies; derives the verdict from the change's findings.
Excessibility.Review.Judge.LLM — asks a model to reason about the change. excessibility owns the prompt, schema, and parsing; the host app supplies the actual model call (a completion function), so the library takes on no HTTP dependency.

Select a judge per call (judge: MyJudge) or globally:

config :excessibility, review_judge: Excessibility.Review.Judge.LLM

Verdict shape

%{
  tier: :auto | :review | :block,
  blast_radius: String.t(),
  risks: [%{severity: String.t(), area: String.t(), detail: String.t()}],
  confidence: float() | nil,
  source: :heuristic | :llm
}

Summary

Functions

verdict(change, opts \\ [])

Judge a change with the configured (or explicitly given) judge.

Types

verdict()

@type verdict() :: %{
  :tier => Excessibility.Review.tier(),
  :blast_radius => String.t(),
  :risks => [%{severity: String.t(), area: String.t(), detail: String.t()}],
  :confidence => float() | nil,
  :source => :heuristic | :llm,
  optional(:judge_tier) => Excessibility.Review.tier()
}

Callbacks

judge(change, opts)

@callback judge(change :: Excessibility.Review.change(), opts :: keyword()) :: verdict()

Functions

verdict(change, opts \\ [])

@spec verdict(
  Excessibility.Review.change(),
  keyword()
) :: verdict()

Judge a change with the configured (or explicitly given) judge.

Resolution order: opts[:judge], then config :excessibility, :review_judge, then Heuristic.

Tier floor

A judge may raise the tier freely, but may not fully green-light a deterministic :block: when the change's heuristic tier is :block (new critical/serious findings from the rules engine) and the judge says :auto, the verdict is floored to :review and the judge's raw tier is kept in :judge_tier. :auto means nobody looks at the change — rendered page content feeds LLM prompts, so that is not a downgrade a model is allowed to make on its own. :review still cuts block-level false-positive friction while keeping a human in the loop.