Keep the evidence. Drop the leap.

This saved auditor compares two visible text specimens. It trims each input, computes an actual Web Crypto SHA two fifty six fingerprint and separately tokenizes words with Intl Segmenter. Word tokens are lowercased for counting. The fingerprint covers the trimmed text bytes, while lexical difference uses word counts. Reordering alpha beta to beta alpha therefore changes the fingerprint but adds and omits no tokens. Case and punctuation can also change bytes while leaving the counted words unchanged. Source alpha alpha beta has two alpha tokens and one beta. Generated alpha beta beta has one alpha and two beta. The source compares multiplicities, so it records one added beta and one omitted alpha. These two bars use two hundred forty pixels per added or omitted token and are both two hundred forty pixels wide. This is a count difference, not an inference that one text was copied, that an author intended deception, or that a model is conscious. The marker chips are fixed phrase-pattern matches in the generated text, including the user, AI collective and certain self-revision phrases. They flag visible wording rather than proving why it appeared. A successful comparison saves its full result locally and enables the complete JSON export. Editing inputs without comparing leaves that prior result. Invalid blank inputs clear the active result and disable export but leave the older saved result in storage. Reload can restore and rerun that last successful comparison. Reset clears inputs and saved result. The sample is illustrative supplied text, not recovered screenshot evidence.