Six word changes against our published watermark

Each replacement below comes from the same authored candidate slot as the word it replaces. We selected changes with access to the detector, then reviewed the result as an ordinary piece of technical writing. The table reports scalar offsets in the original passage, before any replacement changes later positions.

The repeated words matter. A second "value" begins at original offset 327 and a second "passes" at 1030; neither is changed. The last replacement begins at offset 997 at the moment it is applied because the first five changes have added three characters. Recording original offsets makes every target unambiguous.

After five changes, the score is 1.7698, only 0.0302 below the 1.80 cut-off. With 358 effective pairs, one green-pair change moves the z score by about 0.1221 here, so the margin is less than one scoring step. The sixth change reaches 1.5257, leaving 0.2743 below the cut-off.

The detector scores a previous-token and current-token pair. Replacing one word can alter the pair ending at that word and the pair beginning there. For that reason, each selected change above reduces the green count by two even though only one visible word changes.

A guided editor has to recompute the entire passage. A replacement can change neighbouring contexts, and the detector counts a repeated pair only on its first occurrence. A local highlight is a place to investigate, not a promise of a fixed score reduction.

The cleaner asks the user to supply a replacement, previews the new whole-passage score and applies nothing until the user confirms. The editor does not generate synonyms or silently rewrite text. After an edit, the context table is rebuilt rather than pretending that the previous ranking still holds.

Reproduced observation. Six human-reviewed, detector-aware substitutions move this fixed passage from 2.9904 to 1.5257 while keeping 358 effective pairs. The registered candidate slots constrain the available wording, and the exact result is reproducible from the published profile and reference implementation.

The experiment does not establish that meaning is mathematically unchanged. Semantic review is a human judgement recorded for this fixture. Nor does the experiment show that six edits will clear another passage, another seed, another detector or Claude's mark. A different word can increase the score, change the effective pair count or damage the writing.

Crossing the cut-off does not make the result "human". In the frozen evaluation, 2 of 96 passages by named public-domain authors crossed this site's threshold. The score answers one narrow question about one published pattern. It does not answer who wrote a passage or how they worked.

The controlled fixture makes every choice inspectable. Its registered passage, existing candidate slots and fixed expected result form the reproducible experiment.

The guided mode lets a person explore their own sufficiently long text, with fewer guarantees. Below 120 distinct pairs there is no verdict and no attack table. Above that minimum, the editor can show contributing contexts and preview a user-authored replacement. The editor cannot promise that an acceptable alternative exists or decide that a semantic change is harmless.

Publishing a detector makes detector-aware research possible. An automatic rewriting service would be a different product with a much broader claim.

Try the controlled experiment and guided editor, or read what current cleaners can and cannot change.