Entropy

Muse Shield lets people use Muse Spark on their real work (customer records, payslips, medical notes, inboxes, contracts, photos) without handing over their real data. Not a redaction tool. A privacy engine that lets AI work on your private data without ever seeing it. People already paste confidential information into chatbots. Today the choice is "share everything" or "don't use AI". Muse Shield adds a third option: a privacy entity engine that sits between you and the cloud model. It isn't find-and-replace masking. Our fine-tuned Muse Glimmer 30B (er-v3a, a LoRA we trained) reads each message, document, image or tool result and builds an entity graph. It finds every person, organisation and private value (money, IDs, contacts, addresses, secrets, health details) and links every mention of the same person ("Rohan Majji", "Rohan", "Mr. Majji", his email). It also works out who owns what (whose salary, whose account) and the relations between people. Each value becomes a stable placeholder kept in an AES-256-GCM encrypted local vault. Muse Spark only ever sees <PERSON_1>, <AMOUNT_2>, <ID_1>. Its reply is restored on your device. When Spark needs a hidden value, it has to ask. You get one consent card with every item it needs and approve each one. For a photo, Spark can ask for just the licence plate while the face stays blurred. A pre-send guard refuses to send any real value you haven't approved. It also works as an agent. You can connect Gmail, Calendar, Contacts, Telegram and a local folder from the app. Results are masked before Spark reads them. Spark plans tool calls using placeholders, you confirm the real values on a card, and a local executor sends them. Recipients must be your own contacts, never an address planted in an email. It also handles PDF, Word and PowerPoint, and can produce verified masked copies. Results on official test sets: on Nemotron-PII our 30B model has 2.0% core leak vs Muse Spark's 1.9% (base Glimmer: 5.6%). On TAB court judgments it protects 98.9% of direct identifiers vs Spark's 82.6%. Coreference F1 is 0.974 vs 0.891. 20/20 agent scenarios pass.