clipguard / copy-paste DLP

clipguard prevents sensitive work data from leaking into unregulated destinations (personal email, social media, ChatGPT). It fingerprints text when copied from a regulated workspace, then compares pasted text against those fingerprints before allowing it through.

Why a minimum length gate?

• Single words like "Hello" or "Thanks" are not data leaks.
• Short strings produce unreliable fingerprints — too few bits to compare meaningfully.
• Passwords and secrets are short but handled by a separate PII/regex scanner, not fuzzy matching.
• Threshold: 50 characters or 8 words — roughly one sentence. Anything shorter passes without running the engine.

Why strip stop words?

• Words like "the", "is", "and", "hello", "thanks" appear in every document. Zero signal about content origin.
• Without stripping, "Thanks for your email" would match work documents containing those same words — false positive.
• After stripping, only domain terms remain: "revenue", "142M", "falcon", "technova". The real fingerprint.
• If nothing remains after stripping, the text was boilerplate and is allowed immediately.

Why Shannon entropy?

• Entropy measures how unpredictable text is. Common phrases ~2.5 bits/char, API keys and financial data ~4.0+.
• Problem: two boilerplate texts score high on fuzzy similarity because they share common vocabulary.
• Solution: low entropy raises thresholds to 90% fuzzy / 70% fragment. Only near-exact copies of boilerplate trigger a block.
• High-entropy text (the stuff we care about) uses normal, more sensitive thresholds.

Why SimHash for fuzzy matching?

• Catches rephrased content — swapped words, reordered sentences, slightly changed numbers.
• Locality-sensitive hash: similar inputs produce similar hashes. O(n) time, no corpus needed.
• Alternative considered: cosine similarity with TF-IDF. Rejected — requires corpus maintenance, too heavy for client-side.
• Threshold: 70% — high enough to avoid random collisions, low enough to catch paraphrasing.

Why 5-word shingle fingerprints?

• SimHash measures whole-text similarity but misses partial leaks — one paragraph from a 50-page doc.
• Shingles break text into overlapping 5-word sequences. A paragraph from the source still matches via its shingles.
• Smaller set as Jaccard denominator: pasting 1 paragraph from 50 pages still scores high.
• 5 words is the sweet spot: 3-word shingles trigger too often, 7-word shingles miss short passages.

Why require cluster density?

• A single shingle match is coincidence — two docs might share one 5-word phrase by chance.
• We require ≥35% of shingles to match — meaningful structural overlap, not a stray hit.
• Combined with stop-word stripping, shingles contain only signal words. 35% match on signal = strong evidence.

Why a 60-minute vault TTL?

• Bounds memory usage and limits the false-positive window.
• Stale fingerprints from an hour ago cause confusion as context shifts.
• Configurable per policy via Settings panel or config.json.

Why client-side only?

• Prototype. In production, the engine runs in a browser extension with real tab context and clipboard APIs.
• No data leaves the browser — vault lives in memory, dies with the tab.
• Server-side would add latency to every paste. This math runs in <1ms.