org card: tighter copy, name Patronus Ark, animal model families

README.md

@@ -9,29 +9,45 @@ pinned: false
 
 # 🛡️ Patronus Protect
 
-An on-device AI firewall. It runs on your own machine, quietly watching the AI traffic going in and out of your apps and blocking what you'd rather not see leave: prompt injections, leaked secrets, the shadow AI that nobody signed off on. None of it is sent to a cloud to work, because a security tool that phones home never sat right with us.
+**An on-device AI firewall.** It watches the AI traffic going in and out of your apps —
+browsers, IDEs, native apps, MCP servers — and blocks what shouldn't get through: prompt
+injections, leaked secrets, shadow AI. Everything runs locally, nothing is sent to a cloud,
+and it works with any provider (Claude, OpenAI, Gemini, Copilot, Ollama, …).
 
-It covers the places people actually use AI (browsers, IDEs, native apps, MCP servers) and doesn't mind which provider is behind it, whether that's Claude, OpenAI, Gemini, Copilot, Ollama or a local Mistral.
+## Patronus Ark
 
-## The open-source library
+**Patronus Ark** is the scanning engine underneath the app: a Rust core with Python
+bindings that scans in three escalating layers, so most traffic is resolved before a
+transformer ever runs.
 
-Underneath the app is a scanning library written in Rust with Python bindings, and it's open for anyone to use. It's the part we're proudest to give back.
+| Layer | What runs | Cost |
+| --- | --- | --- |
+| **L1** | Native rule-based detectors | microseconds |
+| **L2** | Small classifiers sharing one encoder | milliseconds |
+| **L3** | Full quantized transformers, lazily loaded | only for the uncertain rest |
 
-The way it scans is layered, so it stays fast without cutting corners. A request first hits L1, plain rule-based checks that need no models and run in microseconds. If something looks worth a closer look it moves to L2, small classifiers that share a single encoder and answer in milliseconds. Only the genuinely uncertain cases reach L3, the full transformer models, loaded lazily and run by a background worker that schedules the heavy work. Most traffic never needs L3, so you get real detection without paying for it on every request.
+Across those layers it catches prompt injection and jailbreaks, PII, secrets and DLP leaks,
+sensitive documents, risky agentic tool use and MCP policy violations, plus routing and
+threat classification — all on the endpoint.
 
-Across those layers it catches prompt injection and jailbreaks, PII, secrets and other DLP leaks, sensitive documents, risky agentic tool use and MCP policy violations, plus routing and threat classification. Everything happens on the endpoint, and because the core is Rust it stays quick and light enough to sit in the request path without anyone noticing.
-
-The library is open source (AGPL-3.0, with a commercial license if you're shipping something closed). Issues and pull requests are welcome, and so is a note telling us it broke on your setup.
+Ark will be open source (AGPL-3.0, commercial license available). It is **not public yet**
+— the repository link will appear here at launch.
 
 ## The models 🤗
 
-We had to train our own detectors to build the firewall, and honestly it felt wrong to keep them to ourselves. Open detectors make everyone's AI a little safer, and we've leaned on plenty of open work to get this far. So here they are, with our thanks: use them, fine-tune them, benchmark against them, and please tell us where they fall short.
+Our detectors, trained for the firewall and shared openly — use them, fine-tune them,
+benchmark against them, and tell us where they fall short:
 
-- **Wolf-Defender**, our prompt-injection detector (English and German).
-- **Orca-Sonar**, a document classifier for DLP and sensitive-document routing.
-- A **unified multi-task model** with seven heads: injection, sensitive documents, tool class, tool action, tool tags, routing, threat.
-- **GLiNER Edge** for PII extraction.
+- 🐺 **Wolf Defender** — prompt-injection detection (EN/DE); threat classification *(coming soon)*
+- 🐋 **Orca Sonar** — document classification for DLP and sensitive-document routing
+- 🦁 **Lion Warden** — one unified seven-head AI-security model *(coming soon)*
+- 🐕 **Husky** (Sight / Paw / Nose) — agentic tool use: tool type, operation, data-flow risk *(coming soon)*
+- 🐆 **Panther Read** — user-intent and request routing *(coming soon)*
+- 🦈 **Shark Scent** — PII detection *(coming soon)*
+- **GLiNER Edge** — quantized zero-shot entity extraction (upstream GLiNER, Apache-2.0)
 
-They're deliberately small. ModernBERT under the hood, quantized to ONNX (int8 and int4), a few hundred MB of RAM so they run happily on a laptop instead of a GPU rack.
+Deliberately small: ModernBERT under the hood, ONNX-quantized (INT8/INT4), a few hundred
+MB of RAM — they run happily on a laptop instead of a GPU rack.
 
-The app itself, and downloads for macOS and Windows, live at [patronus.studio](https://patronus.studio). Come say hi.
+The app, with downloads for macOS and Windows, lives at
+[patronus.studio](https://patronus.studio). Come say hi.