Module config 

Evaluator configurations, rustev.evaluator-config/1 and /2 (spec 004, 3.5.3; spec 014).

The configuration is the definition a report’s numbers mean anything under: the task adapter and its version, how labels are read, which proposal payloads define agreement, which step’s probabilities are scored, subgroup and reliability-bin boundaries, latency quantiles, cost unit comparability, exclusions, gate thresholds and the temperature grid. Its canonical digest is the report’s EvaluatorConfigId. Fractional values are Decimal strings, so the identity uses the existing form.

A gate states only metric, direction, tolerance and minimum coverage: the baseline and candidate cohorts it compares (dataset, split, config) are the ones their reports name, since a configuration cannot name its own identity.

Version 2 binds the task adapter’s rules: its adapter reference carries rules, a digest of the document that determines the adapter’s label check and correctness, or opaque. Version 1 has no rules, keeps its bytes and identity, and counts as unbound for every gate (spec 014).

Version 3 (spec 041, 3.1.3) adds the digest of the adapter’s measure declarations to the adapter reference, and declares scorers, rubric metrics, human scales, agreements and origin families, so every one of them enters the identity. Versions 1 and 2 produce no new metric.

Structs

Agreement

Which proposal payloads define an outcome: the action, and these parameters. Whole-judgment equality is never the agreement metric.

ConfigAdapter

The configuration’s task adapter: name and version, and in version 2 its rules.

ConfigError

Why a configuration was refused.

EvaluatorConfig

Gate

MetricFormula

Enums

AdapterRules

What determines a task adapter’s label check and correctness (spec 014, 3.1): the content digest of a canonical document that fully determines them, or opaque when no document does.

Direction

Constants

EVALUATOR_CONFIG

EVALUATOR_CONFIG_V2

Version 2: the adapter reference binds the adapter’s rules (spec 014).

EVALUATOR_CONFIG_V3

Version 3: measures, scorers, rubric metrics, agreements and origin families (spec 041).

MAX_GRID

The most temperatures one grid may hold (spec 004, 3.6.2).