Veil by thecodekaizen

Veil finds what in your video could identify you, before you post it. A 25-second street clip can show a stranger your full name in a notification, your surname and flat number on a door plate, and a phone number you say out loud in Hindi. Standard redaction looks for faces, plates, screens and documents; on our test clips it found none of the four planted private details. Veil asks a different question. Muse Spark reads keyframes and proposes what in this specific video could identify the creator, each with a written reason. SAM 3.1 tracks every proposal through every frame. Muse Voice Transcribe hears spoken numbers in Hindi or English. Then deterministic code redacts, decodes the exported file and checks it: covers verified pixel by pixel, local OCR, a face detector, and re-transcription of the audio. A mask check uses OCR to reject masks that landed on the wrong object. Every finding ends as COVERED, UNRESOLVED or NOT_CHECKED, never "safe". A natural mode re-renders the text with fictional names instead of black boxes, and audits that file too. Discovery can also run on Muse Glimmer, Meta's open-weight model, locally, so keyframes never leave the machine. Results (3 composited test clips, fictional data): standard detection 0 of 4; Veil 11 of 12 across three full runs, then 12 of 12 after one post-gate fix. We pre-registered a pass/fail gate before any model ran. It failed, and we publish that alongside every later result. Audio timing still fails our own test. Live demo: https://thecodekaizen.github.io/veil