Pkl objectives (scoring)

Pkl owns scoring math on evidence. Go builds evidence and validates output.

Concern	Path
Objective schema	configs/schema/SearchBenchObjective.pkl
Example module	configs/rounds/local-ic-vs-jcodemunch/scoring/localization-objective.pkl
Go types	src/searchbench-go/internal/pure/score/
Runner	src/searchbench-go/internal/adapters/scoring/pkl/

Example objective

File: configs/rounds/local-ic-vs-jcodemunch/scoring/localization-objective.pkl

objectiveId = "localization-v1"

local challengerQuality =
  1.0 - (if (current.localizationDistance.goldHop.challenger < maxHop) ...)

local finalScore = base * regressionPenalty * invalidPredictionPenalty

values = new {
  helpers.intermediate("challengerQuality", challengerQuality)
  helpers.penalty("regressionPenalty", regressionPenalty)
  helpers.finalValue("final", finalScore)
}

final = "final"

Piece	Role
current.*	Evidence fields from the round report
helpers.intermediate / helpers.penalty	Named intermediate values
helpers.finalValue("final", …)	Selected score
final = "final"	Which value is the round score

Referenced from the round manifest:

scoring = game.objective("scoring/localization-objective.pkl")

Lifecycle

matches → round-report → evidence.pkl → Pkl objective → objective.json

Output example: configs/rounds/local-ic-vs-jcodemunch/evidence/bundle/games/code-localization/rounds/round-001/objective.json

{
  "objective_id": "localization-v1",
  "values": [
    { "name": "challengerQuality", "value": 0.833..., "kind": "intermediate" },
    { "name": "final", "value": 0.864..., "kind": "final" }
  ]
}

Invalid objective output fails the round in Go.

Non-goals

Objective Pkl does not define round manifests, datasets, evaluator backends, or optimizer wiring.