Parda

I publish screen recordings. One of them showed my tax ID document for four seconds; I found out after it was live. Parda is the step between "stop recording" and "publish". Muse Spark reads sampled frames and decides what is private (ID numbers, names next to IDs, keys, passwords, account numbers, amounts); a local OCR pass then pins each value to its exact pixels on every frame, so covers stay glued through scrolls, drags, typing and window jumps. Covers are a row of asterisks in the text's own colour and size: the `.env` still reads as a `.env`, the key names stay, only the values are gone. SAM 3.1 tracks non-text items such as an ID photo. Muse Voice Transcribe plus Meta's open MMS alignment bleep spoken secrets to the word. A verify agent re-scans the rendered output and fixes what it still sees, then writes `audit.json` with the verdict and a SHA-256 of the clean file. No blur or pixelation anywhere, because both can be reversed. A private mode runs the on-screen step on Muse Glimmer (open weights) on your own GPU, so frames never leave your control. On ten generated clips a planted value stayed readable on 0.0 % of its frames after Parda against 31.7 % after a regex-over-OCR sweep, with 2 of 3 spoken secrets bleeped.