Version 3.2.1

Local CLIP and SAM 3 workflows, managed model downloads, and review-interface improvements.

Historical record dated 2026-09-10. Requirements and release status below refer to this version. For usage instructions, see the documentation.

Changes: v3.1.1 -> v3.2.1

Updated on Apple App Store.

RawCull 3.2.1 requires the latest macOS 27 (Golden Gate), an Apple Silicon Mac, and Xcode 27 when building from source.

Commit range: 3b1c41d..3caf6ba
Latest commit: 3caf6ba (Upgrade PhotoAIKit and CoreAI (#11), 2026-09-10)
Current version: RawCull 3.2.1 (Build 360)

🚀 Major Features

• Released the RawCull 3 local AI workflow.
• Added natural-language semantic photo search using DataComp CLIP.
• Added DataComp CLIP-powered visual similarity analysis and burst grouping.
• Added SAM 3 Deep Review with subject isolation, detail scoring, and burst-winner recommendations.
• Added managed downloads for verified DataComp CLIP and SAM 3 model packs.
• Added persistent storage for AI embeddings, masks, scores, signatures, and burst-review decisions.
• Modularized the complete AI application layer while preserving the RawCull 3.1.1 workflow, ranking policies, cache formats, preferences, and migration behavior.

🧩 Modular AI Architecture

• Added one application-owned RawCullIntelligenceRuntime that creates and retains stable AI feature objects for the app session.
• Replaced settings callbacks with revisioned, typed AI configuration snapshots that reject stale updates and avoid unnecessary restarts.
• Split model discovery, downloads, license acceptance, progress, cancellation, retry, removal, and installed-model locations into a focused model-management component.
• Added dedicated semantic-search and similarity feature surfaces so views no longer inspect providers, embeddings, repositories, or backend descriptors directly.
• Moved burst computation into a dedicated coordinator with typed pipeline values, cache restoration, compatibility migration, grouping, ranking, lifecycle handling, and review-state preservation.
• Added a focused Deep Review controller for availability, request construction, operation lifecycle, cancellation, and burst-specific winner application.
• Reorganized the AI source files under RawCull/Intelligence into Composition, Contracts, Similarity, SemanticSearch, BurstAnalysis, DeepReview, ModelManagement, Persistence, and Presentation areas.
• Kept the intelligence boundary application-local after evaluating a separate Swift package; the current code still depends intentionally on app-owned catalog snapshots, paths, resources, policy callbacks, and Background Assets wiring.
• Removed temporary view-model constructors, runtime model exposure, and semantic-search, similarity, and Deep Review forwarding APIs after callers migrated to the focused features.
• Added exact import-boundary enforcement for all concrete AI backends and PhotoAIKit products. SwiftUI views and general model code no longer import restricted AI modules.
• Kept persistence modularization Phase 9 deferred to avoid mixing storage API changes with the completed structural refactor. Existing on-disk formats remain unchanged.

🧠 AI Models and Downloads

• Added dedicated AI model settings and model-management views.
• Added model bundle discovery, validation, and resource management.
• Added support for DataComp CLIP and SAM 3 models.
• Added model-location controls and validation feedback.
• Added model-download and Managed Background Assets infrastructure.
• Added license-acceptance handling for externally distributed models.
• Added model provenance, license notices, manifests, and installation documentation.
• Published the V3 managed model release containing DataComp CLIP and SAM 3 asset packs.
• Added archive size and SHA-256 verification for the published CLIP and SAM 3 downloads.
• Added provenance validation that rejects incomplete, mismatched, blocked, or malformed production model records.
• Added EfficientSAM license and provenance records for future use; EfficientSAM is not included in the production download manifest.
• Improved model-download validation, retry behavior, error handling, and test coverage.
• Fixed modular-runtime and model-download integration after the Intelligence file reorganization.
• Documented the version coordination required across RawCull, the downloader extension, Background Assets manifests, and model bundles.
• Updated PhotoAIKit to v2.4.0 and the matching apple/coreai-models revision.
• Updated the resolved xgrammar dependency to its released 0.2.2 version.
• Retained the V3 compiled DataComp CLIP and SAM 3 assets. A V4 model release is not required unless a future CoreAI export changes the verified model artifacts or model contract.

🔍 Grid and Similarity View

• Added clearer sharpness-scoring controls with scoring, re-scoring, sorting, and cancellation actions.
• Added contextual scoring targets for selected images, filtered images, or the complete catalog.
• Added combined Index & Find Similar behavior when DataComp CLIP index entries are missing.
• Improved similarity backend status and progress presentation.
• Reserved consistent space for semantic-search activity messages to prevent layout shifts.
• Improved selection coordination and keyboard navigation.
• Refined grid controls, button styling, and rating interactions.
• Routed similarity state and actions through the focused similarity feature instead of the central application view model.
• Added RawCull-owned semantic-search backend presentation values for UI and accessibility without exposing PhotoAIKit types.
• Added Deep Review availability indicators and overlay controls to thumbnail presentation.

🔎 Loupe View

• Added a redesigned metadata panel for the selected image.
• Added expandable and collapsible metadata sections.
• Added keyboard control for showing or hiding metadata.
• Added filename, folder, camera, exposure, rating, and culling information.
• Improved source-image switching and one-to-one zoom rendering.
• Fixed one-to-one zoom so it displays actual image pixels.
• Improved thumbnail and image navigation while using Loupe View.
• Added keyboard actions for the zoom overlay.
• Added optional focus-mask and subject-outline presentation to the zoom workflow.
• Preserved native preview detail for focus-mask analysis instead of using sharpened thumbnail output.

🎞️ Burst View and Deep Review

• Redesigned the burst-culling workspace around a photo-first review experience.
• Redesigned the Burst Groups home screen with a guided Next up card, review queues, recent groups, and completion progress.
• Added staged catalog-preparation status for semantic indexing, sharpness scoring, and burst-group discovery.
• Added clearer progress and estimated-time feedback while preparing a catalog.
• Added separate queues for bursts that need review, completed bursts, and all burst groups.
• Added maintenance actions for scoring parameters, catalog re-indexing, and similarity indexing.
• Added clearer burst numbering and frame-position indicators.
• Added navigation back to the burst list.
• Added reviewed and unreviewed burst controls.
• Added keyboard-action guidance for navigation, zoom, source selection, focus, rating, picks, and rejection.
• Added burst-aware metadata to the zoomed image view.
• Added culling status, rating, sharpness, focus, review, and recommendation information.
• Improved comparison-grid navigation and selected-frame handling.
• Added SAM 3 subject-mask outlines for Deep Review and focus workflows.
• Added mask-outline controls and availability feedback.
• Fixed focus-mask source selection for burst JPEG previews.
• Preserved manual winners and review decisions when compatible burst groups are restored or recomputed.
• Isolated burst analysis from catalog navigation and rating actions while retaining the existing user workflow.

🗂️ Sidebar, Navigation, and File Operations

• Refined the main catalog sidebar and file presentation.
• Added clearer empty, filtered, and missing-file states.
• Made the empty-state catalog icon an accessible button that opens the catalog folder picker.
• Improved scan progress by reporting the discovered-file count directly.
• Added accurate completed/total progress for thumbnail creation and JPEG export.
• Added image-sorting progress and status indicators.
• Moved additional navigation actions into application menu commands.
• Improved toolbar organization and reduced duplicate controls.
• Refined rating controls, filters, and rating-pin behavior.
• Corrected rated-photo counts so only files with an explicit rating are counted.
• Improved copy-file startup, completion reporting, source and destination presentation, and recovery after interrupted work.
• Improved JPEG extraction selection handling and per-file failure reporting.

⚡ Performance and Reliability

• Added replacement-safe thumbnail cache identities.
• Prevented stale cached thumbnails from being reused after files are replaced.
• Added bounded thumbnail preloading to reduce CPU and memory contention.
• Improved thumbnail loading, scanning, and disk-cache coordination.
• Simplified thumbnail and shared-memory cache paths after removing obsolete diagnostics instrumentation.
• Reduced CullingGridView rendering work identified during performance tracing.
• Reduced image-downscaling work identified during performance tracing.
• Made histogram loading use the latest request when selections change quickly.
• Reduced histogram-processing cost by sampling large previews to a maximum dimension of 512 pixels.
• Improved JPEG extraction progress, cancellation handling, selection behavior, and per-file failure reporting.
• Improved AI artifact persistence and cache-boundary handling.
• Improved catalog loading, sorting, saved-file persistence, and culling-state restoration.
• Improved similarity, sharpness, focus, and burst-analysis consistency.
• Preserved generation-token and cancellation behavior so superseded AI work cannot publish stale results.
• Kept DataComp CLIP and SAM 3 artifacts separated by compatible model identity.
• Used PhotoAnalysisKit focus-mask morphology defaults and removed obsolete RawCull visibility overrides.

♿ Accessibility

• Added bounded accessibility labels, values, and hints across major workflows.
• Improved accessibility for ratings, image controls, focus controls, model management, and comparison views.
• Added accessible descriptions for image sorting and metadata controls.
• Added accessible labeling and guidance to the empty-state Add Catalog button.
• Improved keyboard-accessible culling and navigation actions.
• Added accessible progress descriptions to the redesigned Burst Groups workflow.
• Decoupled semantic-search accessibility from PhotoAIKit backend types.
• Added accessible state and controls for Deep Review mask availability and subject outlines.

🧪 Testing and Quality

• Added tests for semantic search, Deep Review, similarity features, burst coordination, and intelligence-runtime identity and configuration.
• Added model-download and model-validation test coverage, including retry and modular-runtime integration.
• Added provenance-validation tests covering malformed records, missing evidence, archive mismatches, blocked distributions, and invalid notice hashes.
• Added thumbnail cache identity and contention tests.
• Added histogram request-ordering and bounded-sampling tests.
• Added typed AI persistence and cache-boundary tests.
• Added accessibility presentation tests.
• Added Loupe and Burst metadata tests.
• Added focused tests for Burst catalog-preparation presentation and JPEG-export selection behavior.
• Added tests for comparison-image source selection, focus-mask resolution policy, subject-mask outline rendering, zoom-overlay keyboard actions, copy completion, and quit recovery.
• Updated culling, burst queue, navigation, concurrency, and thumbnail-provider tests for the focused AI feature boundaries.
• Added isolated test factories so AI and view-model tests do not depend on production caches, settings, or user data.
• Added checked-in smoke and performance enumeration verification with duplicate-identifier protection.
• Expanded the smoke manifest to 208 unique test selectors and retained two performance tests.
• Built RawCull successfully after resolving the PhotoAIKit v2.4.0 and CoreAI dependency upgrade.

🧹 Diagnostics and Code Cleanup

• Removed the Diagnostics menu and the memory, RAW-file, similarity, and semantic-search diagnostics screens.
• Removed the supporting diagnostics logs, view models, cache statistics, and obsolete contention instrumentation.
• Removed unused production APIs and their obsolete test-only coverage.
• Retained active operational logging and intentional test-support APIs.
• Removed obsolete AI compatibility shims after all callers moved to focused feature APIs.
• Completed an unused-code and import audit across RawCull/Intelligence with no findings in the modular AI boundary.
• Removed superseded modular-AI planning documents after consolidating the active model-release and provenance guidance.

🛠️ Build and Documentation

• Updated the project for the RawCull 3.2.1 release.
• Advanced RawCull to Build 360.
• Added a dedicated model-downloader target and supporting entitlements.
• Updated Swift package dependencies and project configuration.
• Updated PhotoAIKit to v2.4.0 and CoreAI models to revision cc8120.
• Added model installation, provenance, licensing, and macOS Background Assets release documentation.
• Added detailed modular-AI architecture, migration, dependency-boundary, rollback, and validation documentation.
• Added a high-level modular-AI guide for contributors who are new to the architecture.
• Updated the README with the version-3.2.1 release requirements, AI workflow, model-download guidance, and build instructions.
• Updated export, release, smoke-test, performance-test, import-boundary, and model-provenance verification.

📦 Version Progression

• Released the initial RawCull 3.0.0 implementation.
• Stabilized AI persistence, thumbnail loading, accessibility, and release validation.
• Advanced the project through RawCull 3.1.0 and RawCull 3.1.1 Build 345.
• Started RawCull 3.2.0 as the modular-AI development branch from the 3.1.1 behavior and data-format baseline.
• Completed modular-AI Phases 0-8, 10, 11, and 12; intentionally deferred persistence Phase 9.
• Published the V3 DataComp CLIP and SAM 3 managed model packs.
• Added Deep Review mask outlines, focus-mask improvements, copy-file reliability fixes, and performance improvements.
• Updated RawCull to PhotoAIKit v2.4.0 and its matching CoreAI revision.
• Released RawCull 3.2.1, Build 360.
• Updated this changelog through commit 3caf6ba on 2026-09-10.

• ←Previous
• Next→