Skip to content

v0.2.0 — Redaction + MCP Server + Layer 3 + one-click bundle

Choose a tag to compare

View all tags
@davidoladeji-ogenticai davidoladeji-ogenticai released this 07 May 14:20
v0.2.0
a6945be
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's new

The full v0.2 milestone — 15 tickets, 5 PRs, 329 tests passing. Plus a one-click MCP Bundle for non-technical install.

Highlights

  • Redaction API — Shield.redact() / Shield.unredact() with deterministic tokens. Numbers and percentages survive; identifiers get masked; restoration is byte-perfect (OGE-308 / OGE-309).
  • MCP server — five tools (shield.analyze, shield.analyze_batch, shield.redact, shield.unredact, shield.profiles) over stdio + SSE. Now async natively (OGE-310 / OGE-311 / OGE-312 / OGE-318 / OGE-319).
  • Layer 3 (Ollama, opt-in) — localhost-only LLM classification with profile-tuned prompts. Currently disabled by default — measurement showed it trades 15-30pp precision for marginal recall on the OGE-51 benchmarks. See benchmarks/MOE_COMPARISON.md (OGE-313 / OGE-314 / OGE-315 / OGE-320).
  • Cross-layer calibration framework — pluggable per-layer calibration (linear / Platt / isotonic), packaged JSON defaults, refit script (OGE-321).
  • shield-therapy-pro profile — DSM-5-TR named diagnoses, CPT mental-health billing codes, expanded DOB patterns, minor-client markers, trauma indicators (OGE-355).
  • Custom recognizer SDK — shield test-recognizer CLI command, template + GDPR example, CONTRIBUTING update (OGE-322).
  • AsyncShield + analyze_batch — coroutine-friendly variant with analyze_stream async generator, parallel batch with per-item failure containment (OGE-318 / OGE-319).
  • ModelRegistry — Shield.required_models() quality tiers (fast / quality / comprehensive) so downstream consumers know what to pre-pull (OGE-315).

Easiest install — MCP Bundle

For Claude Desktop / Goose / Cursor / any MCPB-compatible client:

  1. Download ogentic-shield-0.2.0.mcpb (attached below).
  2. Open Claude Desktop → Settings → Connectors → + → Install from file → pick the .mcpb.
  3. First launch downloads spaCy's en_core_web_lg model (~600MB, one time).
  4. Five shield.* tools appear in the connector. Done.

No git, no Python, no terminal.

Developer install

pip install 'git+https://github.com/OgenticAI/ogentic-shield.git@v0.2.0#egg=ogentic-shield[mcp]'
ogentic-shield serve --profile shield-legal --profile shield-finance

Privacy contract (unchanged)

All processing runs on the user's device. Layer 3 is contractually localhost-only — LocalhostOnlyError is raised at config load and at client construction if anything else is configured. MCP tool responses default to shape-only entity reporting (no raw matched text leaks unless the caller opts in).

Honest caveat

No model in OGE-320 — including the L1+L2-only baseline — meets every PRD §8 precision target on the current OGE-51 dataset. The L1+L2 stack is reliable for routing decisions (LOCAL_ONLY / REDACT_CLOUD / CLOUD_OK); production callers should leave Layer 3 disabled for now. v0.3 work is prompt-narrowing + a bigger labelled dataset.

Assets 3
Loading