VaultSort 3.0.0

What's New

On-device semantic understanding. A small (~470 MB, downloaded once) Apple-Silicon-native embedding model reads PDFs, ePubs, Word, plain text, Markdown, code, and more — entirely on your Mac via the Apple Neural Engine.
Categories from your content, not a generic list. Smart Categorize clusters your real files and proposes names derived from the patterns it actually finds — Audits, Invoices, Officeholders, Presets — instead of forcing your library into pre-baked buckets.
Live wizard. Pick a folder, watch the analysis stream in (300-file reservoir sample, ~2 seconds on M-series), then accept, edit, merge, or drop each proposed category before any file moves.
Saved Category Sets. Curate once, reuse across Advanced Organize jobs.
Composable with the full Rule Builder. The new semanticCategory condition combines with filename, size, date, tag, and every other Advanced Organize predicate. Example: "if semanticCategory is Receipts AND size > 500 KB, move to /Receipts/Large."

Order-independent results. Re-running on the same folder set in any selection order now produces byte-identical proposals.
Tighter junk-drawer behaviour. Prose-heavy clusters require ≥ 0.7 cohesion to surface; loose groupings merge into Uncategorized rather than littering the proposal list.
Smarter naming. Original casing preserved for acronyms (USAH stays USAH, not Usahs); irregular plurals (Children, Knives, Mice) handled correctly; calendar tokens, copy markers, and filename modifiers (last, latest, prev) excluded from cluster names.
Filename-evidence wins on disagreement. When body content and filenames genuinely disagree about what a cluster is "about", the filename signal now wins — matching how users actually read their own folders.

Embedding model integrity verified via SHA-256 checksum at install.
Embedding cache scoped per-user, never synced, never transmitted.
New smart_categorize_derive_completed and _prepass_completed events surface health metrics (silhouette, cache hit rate, sanity counts) — all aggregate, no file paths or content.

Fully redesigned Advanced Organize interface to support the Smart Categorize wizard alongside the existing job and rule system.
Three-button save-collision dialog when accepting a Category Set whose name already exists (Replace / Save as copy / Cancel).
Schema-version stamp on every saved Category Set with forward-compatible migration shell, so older sets continue to load and newer sets are gracefully skipped on older builds.
L2-norm sanity guards in both deriver and prepass paths — degenerate embeddings can no longer crash a derive run.
Embedding-cache dimension validation on open and on each write.

Exemplar deduplication: clusters no longer surface multiple files that share the same basename.
Fewer "Cluster N" fallbacks; singleton clusters now derive names from their own filenames when TF-IDF has no signal.
Prose-heavy junk drawers (loose .md / .txt / .docx groupings) no longer surface as named proposals at low cohesion.
Telemetry events now include dropped_non_finite and renormalized counts so embedding-pipeline health is monitorable.

macOS 12+ on Apple Silicon. Smart Categorize requires Apple Silicon; the rest of VaultSort 3 runs on Intel as before.
Existing rules, jobs, and Category Sets from VaultSort 2.x load unchanged.

VaultSort 3 ships with a price increase on new licenses. Existing license-holders receive the upgrade as part of their plan.