Can cleaning recover text already replaced by U+FFFD?

Measured inputs and outputs

These are locally constructed test strings, not evidence that Claude inserts these characters. We executed the saved homepage script snapshot with a minimal DOM harness and compared exact strings. The invisible-checkbox state and dash mode appear in each row. The observations below test this page's specific question. File API and event-lifecycle tests include an additional stage described in the analysis. We make no detector-score or statistical-watermark removal claim.

Different original bytes can produce one string

We decode a one-byte buffer containing FF and another containing FE with the default UTF-8 TextDecoder error policy. Both produce the same U+FFFD string. The cleaner retains that string in both cases. A mixed input loses its separate U+200B but keeps U+FFFD; the intentional-literal reference also remains. The observer counts replacement characters and records the output UTF-8 encoding. Those output bytes encode the replacement character itself; they are not a reconstruction of either original invalid byte. This demonstrates an information boundary rather than the wrong-UTF-16-decoder scenario measured on another page.

A retained symbol cannot identify its history

Once two distinct source byte sequences have become the same string, a cleaner receiving only that string cannot distinguish their origins. U+FFFD can also have been entered deliberately, as our final fixture illustrates. The mere presence of the symbol therefore does not prove a particular encoding fault. We save the explicit byte preparation in the downloadable reproducer so the two decoding cases can be checked independently. No real file was corrupted for this test, and we do not infer that Claude, an editor or a server caused a replacement in a user document.

Return to original bytes when they exist

Preserve the original file or response buffer before trying another decoder. Determine its declared and actual encoding through the originating system, then decode a working copy with the intended rules. If only the already-replaced string remains, state that recovery is unavailable from that representation alone. Replacing every U+FFFD with a guessed letter creates new content and cannot be verified as restoration. Rerun the destination check after any evidence-based repair. The homepage can remove its selected formatting characters but cannot reverse a lossy decoding decision or infer missing source bytes from a glyph.

Reproduce this test

Save reproduce.cjs and tested-app.js in the same folder. Run the command below with Node.js. The harness prints its runtime, script SHA-256 and every measured row. Compare those rows with the original record. Using a newer script or runtime creates a new experiment; retain the version information with your rerun.

node reproduce.cjs