v0.2.0 - System One Decision Models & Swift 6.1 Traits

Added

• Repository and Package Rename (SystemOneFoundationModels):

  • Renamed the repository from jev-foundation-models to system-one-foundation-models and the core SPM package and umbrella library product to SystemOneFoundationModels.
  • Reflects multi-model decision support across Laya on-device Core ML (Apple Neural Engine), Laya self-hosted HTTP (laya-serve), and hosted TypeSafe Jev cloud endpoints.
  • Modularized targets into focused libraries: SystemOneCore, LayaOnDevice, LayaFoundationModels, JevFoundationModels, and the umbrella SystemOneFoundationModels.
  • Updated all package dependency manifests, documentation, and Swift Package Index links.

• NutritionLabelScanner Reference iOS Application (Examples/NutritionLabelScannerApp/):

  • Added native iOS reference application showcasing real-time camera scanning and multi-backend decision modeling for nutrition safety.
  • Implemented live Vision OCR text recognition with spatial 2D row-clustering tolerance (NutritionLabelParser) to accurately reconstruct horizontal lines across multi-column tables.
  • Supports multilingual label parsing for European / DACH (Nährwertdeklaration, Brennwert, Zutaten), French, and FDA / US (Nutrition Facts, Serving Size, Ingredients) standards.
  • Strongly-typed dietary safety evaluation via @Generable DietarySafetyDecision and DietaryFlag models, computing allergen risk (0–3) and NOVA food processing classification (0–3).
  • Configurable dietary restriction profiles (vegan, gluten-free, lactose-free, seed-oil-free, low sodium, sugar limits).
  • Features OpenFoodFacts barcode fallback resolution, dynamic bounding box viewfinder overlays, and accessible haptic/audio feedback.

• Ergonomic Decision Shortcuts on LanguageModelSession (SystemOneCore/Ergonomics):

  • Implemented high-level ergonomic extensions on Apple's LanguageModelSession enabling ad-hoc queries without declaring boilerplate @Generable schema types:
    • session.probability(of:state:criteria:): Direct truth probability evaluation ($0.0 \dots 1.0$) for statements against state with optional criteria guidance.
    • session.choice(_:from:state:): Strongly-typed categorical selection across cases of any Swift enum conforming to Choosable.
    • session.choice(_:options:state:): Categorical selection among dynamic runtime string arrays.
    • session.score(_:levels:state:): Ordinal rubric scoring returning discrete winner indices and continuous probability-weighted mean scores.
  • Introduced the Choosable protocol (CaseIterable & Hashable & Sendable) with automatic default identifiers and optional natural-language option descriptions.
  • Introduced Choice<Option> container encapsulating winning value, calibrated confidence score, and candidate probability distribution.
  • Introduced ScoreResult container encapsulating continuous expected value, winning level index, level label, and complete rubric probability distribution.
  • Implemented DynamicDecisionSchema for runtime generation of GenerationSchema instances, ensuring ad-hoc queries undergo standard schema translation and retain backend parity across LayaCoreMLEngine, LayaHTTPBackend, JevBackend, and MockSystemOneBackend.

• Swift 6.1 Package Traits Partitioned by Model Boundaries (SE-0402 / SE-0450):

  • Upgraded Package.swift to swift-tools-version: 6.1 and adopted package traits.
  • Partitioned traits strictly along model boundaries matching developer intent:
    • Jev: Enables hosted TypeSafe Jev cloud decision API client (JevFoundationModels).
    • Laya: Enables on-device Core ML and Apple Neural Engine execution (LayaOnDevice).
    • LayaServe: Enables HTTP client for self-hosted private laya-serve instances (LayaFoundationModels).
  • Added convenient persona shorthands:
    • OnDevice: Activates ["Laya"] for zero-network, zero-secret local edge deployments.
    • Remote: Activates ["Jev", "LayaServe"] for networked clients without Core ML neural binaries.
    • All: Activates ["Jev", "Laya", "LayaServe"] for multi-backend showcase applications.
  • Configured .default(enabledTraits: ["Jev"]) for seamless backwards compatibility.

• Zero-Dependency ProxyTransport for Mobile Reverse Proxies (JevFoundationModels):

  • Added ProxyTransport conforming to JevTransport and Sendable in Sources/JevFoundationModels/Transport/ProxyTransport.swift.
  • Supports zero-trust mobile architectures routing through reverse proxies (e.g., Firebase Cloud Functions, Cloudflare Workers, AWS API Gateway) without embedding API keys in Mach-O binaries.
  • Provides dynamic closure-based Credential resolution:
    • .bearer: Dynamic OAuth 2.0 / OIDC / Firebase Auth token injection.
    • .header: Named header injection (e.g., X-Firebase-AppCheck with Apple App Attest / DeviceCheck).
    • .custom: In-place request mutations for dynamic HMAC signatures and replay nonces.
  • Evaluates credential resolution per attempt inside the retry loop, guaranteeing fresh tokens on retries for consumable single-use or rotating credentials.
  • Features jittered exponential backoff respecting RFC 9110 Retry-After headers and latency telemetry via x-envoy-upstream-service-time.

• Standalone Trait Sample Applications (Examples/TraitSamples/):

  • 01-OfflineLayaApp: 100% offline document classifier using LayaOnDevice and Choosable enum categories (traits: ["Laya"]).
  • 02-CloudJevWorker: Lean ticket triage worker querying hosted Jev API with zero ML linkage (traits: ["Jev"]).
  • 03-PrivateLayaServer: Internal VPC cluster ordinal vulnerability scoring (ScoreResult) over private HTTP (traits: ["LayaServe"]).
  • 04-HybridRouter: Local-first edge execution escalating to cloud Jev when confidence falls below 0.80 (traits: ["All"]).
  • 05-SecureAppCheckApp: SwiftUI and CLI application utilizing ProxyTransport with hardware-backed Firebase App Check (Apple App Attest). Includes production Cloud Function reference, automated deployment wizard setup-firebase-project.sh, and SETUP-PRODUCTION.md.

• Technical Notes Expansion (0003 through 0011):

  • Documented deep SDK findings, architectural designs, and platform nuances in tech-notes/:
    • Tech Note 0003 (tech-notes/0003-foundationmodels-dynamic-profiles.md): Declarative dynamic profiles and session adaptation in Apple Foundation Models (LanguageModelSession.DynamicProfile, DynamicProfileBuilder, @SessionPropertyEntry, turn isolation).
    • Tech Note 0004 (tech-notes/0004-secure-mobile-transport-appcheck.md): Production mobile architecture securing TypeSafe credentials using Apple App Attest / DeviceCheck hardware attestation, Firebase Cloud Functions proxy, and JevTransport.
    • Tech Note 0005 (tech-notes/0005-spm-dependency-isolation-and-vendor-transports.md): SPM dependency isolation and decoupling vendor-specific transports from core targets to prevent dependency creep.
    • Tech Note 0006 (tech-notes/0006-http-resilience-and-confidence-routing.md): HTTP resilience, RFC 9110 Retry-After parsing, exponential backoff, cooperative cancellation safety, and calibrated confidence routing for the undecided band (0.35...0.65).
    • Tech Note 0007 (tech-notes/0007-pluggable-system-one-backends-and-laya-serve.md): Pluggable SystemOneBackend protocol architecture, wire protocol parity between TypeSafe Jev /v1/systemone and self-hosted laya-serve, and dynamic header injection.
    • Tech Note 0008 (tech-notes/0008-on-device-coreml-decision-engine.md): On-device decision models via Core ML and Apple Neural Engine (ANE), ModernBERT/mmBERT compilation, option marker gathering, 8-bit quantization, and zero-dependency Swift tokenization.
    • Tech Note 0009 (tech-notes/0009-package-traits-and-ergonomic-shortcuts.md): Architectural analysis of AnyDecisionModel, runtime dynamic schema synthesis (DynamicDecisionSchema), ergonomic shortcuts (.probability(), .choice(), .score()), Choosable protocol, and model-bounded package traits.
    • Tech Note 0010 (tech-notes/0010-proxy-transport-and-dynamic-attestation.md): Zero-dependency reverse proxy transport architecture, dynamic Credential enum (.bearer, .header, .custom), per-attempt token acquisition inside retry loops, and secret-free mobile binaries.
    • Tech Note 0011 (tech-notes/0011-xcode-27-2-json-xcproj-format.md): Adoption of Xcode 27.2+ native JSON project format (project.xcproj) replacing legacy OpenStep .pbxproj for human and AI agent ergonomics.

Fixed

• macOS Keychain Password Prompts in KeychainService.swift:
  • Eliminated legacy macOS keychain queries (SecItemCopyMatching / SecItemAdd without kSecUseDataProtectionKeychain).
  • Prevents system password authorization dialogs when running un-entitled environments (CLI runners, test runners, Xcode previews, un-provisioned dev builds).
  • Adopted shared in-memory fallback store with thread-safe access and reset facilities for clean, isolated test runs.

v0.1.0 - Initial Preview Release

Initial Release (v0.1.0)

A lightweight, native Swift 6 bridge integrating TypeSafe AI's Jev System One decision model into Apple's Foundation Models framework.

Highlights

• Native Apple Foundation Models conformance (LanguageModel, LanguageModelExecutor).
• Automatic schema translation from @Generable structs/enums to Jev System One questions (noul, choice, score).
• 40–150ms structured evaluations with zero hallucinations.
• First-class calibrated probabilities and confidence metrics via response.metadata.
• Offline mock transport (MockJevTransport) for deterministic testing.
• Includes sample applications: ticket-triage-demo and duplicate-article-demo.
• Hosted DocC documentation configuration via .spi.yml.