All notable changes to JudgeCalibrationKit will be documented here.
- Typed imports for xceval 0.4
reportdocuments, including their complete normalized sample rows and mapping diagnostics. - Canonical
artifactIDand exact-sourcebyteDigestprovenance on inspect and report imports when xceval supplies them.
- Updated the canonical
XCEvalFormatdependency to xceval 0.4.x. - Expanded the typed document boundary so every xceval 0.4 command is handled
explicitly. Non-sample command documents return a deterministic
unsupportedCommandimport error.
- Existing samples JSON, samples JSON Lines, and sample-bearing inspect JSON remain supported, including the checked-in canonical xceval 0.3.0 fixture.
- Metrics, evidence, selections, comparisons, gates, and dataset-lifecycle documents are not treated as calibration populations. They are summary-only or may represent filtered subsets, so importing them would bias agreement statistics.
- Typed normalized xceval imports through the upstream
XCEvalFormatproduct, including source provenance and deterministic mapping diagnostics. - Explicit single-key metadata slice analysis with distinct missing-value slices, bounded cardinality, and deterministic reports.
- Worst-slice judge-human CI gates with selection-aware clustered-bootstrap intervals, eligible-slice counts, and discarded-replicate attribution.
- A deterministic iOS 27 simulator harness for runtime verification of the typed Apple Evaluations adapter. The harness cross-target typechecks, but its local run found no installed iOS 27 simulator runtime; a live runtime PASS remains pending.
- Raised the package baseline to Swift 6.2 and macOS 14 while retaining the existing iOS 16, tvOS 16, watchOS 9, and visionOS 1 declarations.
JudgeCalibrationXCEvalnow delegates ownership of the normalizedxceval/v1envelope toXCEvalFormat0.3.x. Calibration mappings, stable sample identity, scale validation, and policy remain project-owned.- The package-level dependency statement is now scoped precisely: the
JudgeCalibrationcore target links no third-party dependencies, while the optional xceval bridge uses the canonical upstream format product.
- Existing
judge-calibration/v1datasets and data-import convenience APIs remain supported. - Existing analysis and gate configurations retain overall-analysis behavior when slice options are absent.
- Worst-slice
humanPanelgates are explicitly unsupported in this release; slice gates supportjudgeHumanandeveryJudgeHumancomparisons. - Worst-slice mean absolute and signed error results are explicitly unavailable because a minimum-value gate cannot represent their metric direction safely.
- Swift 6.2 and macOS 14 are now the minimum toolchain and macOS deployment baselines; the other declared Apple platform floors are unchanged.
- SwiftPM now resolves
Evaluations-Framework-CLI0.3.x across the package graph. TheJudgeCalibrationtarget itself still links no third-party module. importData,importJSON, andimportJSONLinesremain available. TypedXCEvalDocumentand[XCEvalSampleLine]import-result APIs add provenance and diagnostics without changingjudge-calibration/v1.- Configurations without slices retain 0.1.x overall-analysis behavior.
- Envelope/schema/command/sample-count and JSON Lines decode failures now use
XCEvalFormat's error surface; JCK continues to own mapping and calibration errors.
Thanks to Rudrank Riyam for publishing XCEvalFormat in xceval 0.3.0 and for
making the normalized interoperability boundary a shared public product.
- Dependency-free calibration schema, validation, dataset merging, metrics, reports, clustered-bootstrap intervals, and gates.
- Swift Testing failure helper.
- Typed Apple Evaluations adapter with dynamic Xcode discovery scripts.
- Optional importer for normalized
xceval/v1output, pinned to xceval 0.2.1. - Hand-authored examples, design documentation, and DocC guides.