Why can a marker reappear when I export cleaned text as JSON?

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. Destination observations run independently on the exact input and cleaned strings in the recorded Node runtime. We make no detector-score or statistical-watermark removal claim.

The exporter deliberately creates a new character after cleaning

We build a fresh object with the tested string in value and an authored toJSON method. That method returns an exported property containing the value followed by a new U+200B. The marked input is cleaned to AB, and a plain object serializes that exact value. The hooked version instead produces AB followed by the appended marker. A nested export has the same added character under box. The ordinary AB control repeats the effect without any cleanup, and even the empty control becomes a one-marker exported value. This rules out an old marker somehow surviving deletion as the explanation. The method is deliberately present in our fixture; we did not discover it in a real Claude exporter or third-party application.

The hook chooses representation while the source stays unchanged

The observer saves the hook calls and the source value after serialization. Root export receives the empty key, and nested export receives box. Those calls happen on the fresh fixture object, while its original value property remains the supplied string. The destination creates a different representation without mutating that property. Our previous JSON text guide uses ordinary values and shows preserved serialization; our replacer guide filters property names. Neither demonstrates an object supplying its own exported form through toJSON. We do not test recursive hooks, network transport, class inheritance or an actual export plugin. Calling stringify again from a hook can introduce separate recursion behavior, but this observer returns an ordinary object and does not perform that unmeasured operation.

Inspect final export bytes as well as the cleaned source

If an export contains characters absent from the cleaned input, review the transformation stages that run afterward. Compare the plain source property, hook implementation and final serialized payload before blaming the cleanup counter. The retained-joiner control also separates a character the cleaner keeps from the newly appended selected marker. When you control the exporter, define its output contract and verify it with an exact payload check; do not rely solely on the earlier input status. Our downloads show every generated string and hook key with the actual saved cleaner script. They do not remove arbitrary application methods or certify an entire export pipeline. Preserve both representations so a later change can be traced to the stage that actually introduces it.

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