Supra AI 1.3.4

Full Changelog: v1.3.3...v1.3.4

Supra AI 1.3.3

Full Changelog: v1.3.2...v1.3.3

Supra AI 1.3.2

What's Changed

• 1.3.2: automatic document classification by @cadespivey in #47

Full Changelog: v1.3.1...v1.3.2

Supra AI 1.3.1

What's Changed

• 1.3.1: chat rendering, citation styles, guided model setup, document-import fix by @cadespivey in #46

Full Changelog: v1.3.0...v1.3.1

Supra AI 1.3.0

What's Changed

• Ship a notarized DMG in releases by @cadespivey in #38
• Add marketing website and GitHub Pages deploy by @cadespivey in #39
• Fix CI typecheck for static image imports by @cadespivey in #40
• Restyle the site with Typography-for-Lawyers principles by @cadespivey in #41
• Remove confusing illustrative blocks from the homepage by @cadespivey in #42
• Hero: single-block headline + equidistant Download link by @cadespivey in #43
• Chat context, jurisdiction-bound research, deterministic model picks, UI polish (1.3.0) by @cadespivey in #44
• Default high-quality reasoning to DeepSeek-R1-Distill-Qwen-32B by @cadespivey in #45

Full Changelog: v1.2.0...v1.3.0

Supra AI 1.2.0

What's Changed

• Sign with Developer ID (stable container + foundation for auto-update) by @cadespivey in #35
• In-app update check + notarized-release pipeline by @cadespivey in #36
• Bump version to 1.2.0 by @cadespivey in #37

Full Changelog: v1.1.0...v1.2.0

Supra AI 1.1.0

Models

• The legal-model-plan's role models are now downloadable. The guided download in the Models tab now offers the three role models from local-legal-model-setup.md (verified on HuggingFace, all 4-bit MLX):

  • Qwen3-30B-A3B-Thinking-2507 → legal reasoning (/legal, /research)
  • Qwen3-30B-A3B-Instruct-2507 → drafting (/draft)
  • DeepSeek-R1-Distill-Qwen-32B → critique (/critique)

  A downloaded model auto-resolves to its chat route (its repo name normalizes to the route's configured identifier), so no manual assignment is needed.

• Per-route model recommendations. Each chat route in the Models tab now suggests the best model among what you've downloaded — the plan's model if you have it, otherwise the best fit by type (thinking/reasoning models for legal reasoning and critique; instruct models for drafting) — with a one-tap Use to assign it.

Note

The planned 6-bit high-quality legal-reasoning model isn't published by mlx-community, so that route has no curated download but remains manually assignable.

If you want to support the work of CourtListener or the Free Law Project, or create your own account for access to their databases, please visit https://free.law/membership/.

Supra AI 1.0.0

Supra AI is a local, MLX-powered macOS research and drafting assistant for legal work. It runs large language models entirely on-device and pairs them with source-grounded legal research and a document-intelligence pipeline — with citation verification built in so the model never presents unverified authority as settled law.

⚖️ Not legal advice. Every citation, quotation, and proposition must be independently verified by a qualified attorney before any reliance.

✨ Highlights

• Fully local generation — models load and run in a sandboxed MLX XPC service; no prompt, document, or query leaves the device for generation.
• Source-grounded legal research — /research and /legal retrieve authority from CourtListener, rank it (jurisdiction, court level, recency, precedential status), and constrain the model to the retrieved packet. Fabricated or unsupported citations are flagged and quarantined behind a "do not rely" banner instead of being shown as good law.
• Document intelligence — import PDF/DOCX/XLSX/RTF/EML/images with OCR, chunking, on-device embeddings, hybrid retrieval, source-cited Q&A, fact chronologies, and exportable structured outputs, all scoped per matter.
• Matter workspace — organize chats, research sessions, authorities, documents, and outputs by matter, with an audit trail.
• Privacy by default — privileged query terms are redacted to stable fingerprints in logs and diagnostics unless explicitly enabled; the CourtListener token is Keychain-stored and device-bound.

🔒 Pre-publish hardening

This release went through four rounds of adversarial security/correctness review (converged to zero findings), hardening legal-citation safety, path-traversal and zip-bomb handling in document import, XPC parameter validation, generation concurrency, and data-integrity on delete. 242 package tests pass and the full app scheme builds.

📋 Requirements

macOS 15+ on Apple Silicon · Xcode 16+ (Swift 6) · local MLX model weights · optional free CourtListener API token for research.

See the README for setup and configuration.

---

What's Changed

• Milestone 1: model selection, persisted chat, cross-process file access, validation runner by @cadespivey in #1
• App shell: model downloader, Settings, sidebar fix; remove Tasks by @cadespivey in #2
• Matters: folders that group chats by @cadespivey in #3
• Surface load-failure detail; guard downloads by model_type by @cadespivey in #4
• Fix XPC runtime service registration (missing XPCService Info.plist key) by @cadespivey in #5
• Handle reasoning-model output: suppress + strip chain-of-thought by @cadespivey in #6
• Validation run reliability: fail-fast on service drop + reconcile stranded runs by @cadespivey in #7
• Make the generation preset actually affect sampling by @cadespivey in #8
• Milestone 2: legal research layer (checkpoint — WO 19–24) by @cadespivey in #9
• Milestone 2 polish: close out all caveats / minor items by @cadespivey in #10
• Silence unused-try? warning (zero-warning build) by @cadespivey in #11
• Add Milestone 3 plan (Matter Document Intelligence) by @cadespivey in #12
• Silence unused-try? warning in ModelDownloadController by @cadespivey in #13
• Milestone 3: Matter Document Intelligence by @cadespivey in #14
• UI/UX quick wins: tab order, All Documents, chat list, import-failure banner by @cadespivey in #17
• App shell overhaul: sidebar restructure, runtime reconcile, drop Limited Mode, toolbar by @cadespivey in #18
• Explain Structured Outputs + clarify Context (#10, explainer half) by @cadespivey in #19
• Fix M3 validation OCR scenario (decodable fixture image) by @cadespivey in #16
• Seeded 3-matter test corpus + validation suite (incl. live CourtListener) by @cadespivey in #15
• Research/Authorities XCUITest: target, a11y IDs, seeding hook (#4) by @cadespivey in #20
• Ground Structured Outputs in chosen documents (#10 follow-up) by @cadespivey in #21
• Unify matter tab layouts (shared MatterTabScaffold) by @cadespivey in #22
• Model load state: real Loaded vs Default + auto-load selected model on startup by @cadespivey in #23
• Remove in-app Milestone 1/3 validation triggers by @cadespivey in #24
• App logo, launch splash, and "See Supra" brand (§) by @cadespivey in #25
• Show CourtListener API token status in Settings by @cadespivey in #26
• Chat status bar: model/processing state + generation settings switcher by @cadespivey in #27
• Global chat attachments: files/images/screenshots into context (OCR-to-text, ≤10) by @cadespivey in #28
• Add Assistant Profile ("soul document") in Settings by @cadespivey in #29
• Adversarial code audit: fixes for bugs, dead code, and abandoned branches by @cadespivey in #30
• Audit follow-ups: fix/remove remaining identified issues by @cadespivey in #31
• Pin ad-hoc signing to keep the sandbox container (and its data) stable by @cadespivey in #33
• Collapsible reasoning view in chat by @cadespivey in #32

New Contributors

• @cadespivey made their first contribution in #1

Full Changelog: https://github.com/cadespivey/Supra-AI/commits/v1.0.0