Inspect and change text locally

Try the prepared word changes against Declawd’s public detector, or see exactly which listed Unicode characters your text contains before deciding to remove any. Your text is not sent over the network.

This page does not detect or remove Claude’s watermark.

Anthropic says hidden characters are not its text-mark carrier. Unicode cleaning does not directly target that carrier, and neither it nor rewriting here can certify an Anthropic detector result.

This page tests Declawd’s published watermark and changes only the characters or words you select.

The separate SynthID-Text fixture explains a public token-scoring profile without accepting text.

Do not use these results to judge a person.

The cleaning controls need JavaScript because every inspection and change runs in your browser rather than on a server. With it switched off, the explanations remain available but no text is processed.

Prepared word changes

The marked passage begins at 0.77. Each of the 6 changes below is pinned to a scalar offset in the original passage, which keeps repeated words from being confused with one another. Applied together they move the full-passage score to -0.21, which is below threshold at the 2.30 cut-off.

The source passage already starts below the cut-off. These edits measure a further score reduction, with no threshold crossing.

One green context is worth about 0.12 on this passage. The registered procedure tries up to six changes, choosing the largest available score reduction each time from the reviewed template candidates. It reports the result even if it remains above the cut-off.

sensible@566 becomes worthwhile
continue@1942 becomes proceed
record@58 becomes log
value@75 becomes reading
flow@121 becomes rate
settle@162 becomes stabilise

Prepared passage

Set the reference flow to 40 litres per minute before you record the first value on the raw sheet. Hold the line at that flow for two minutes so the pressure can settle. Write down the meter value, the ambient temperature, and the clock time. Take readings at 60 and 80 litres per minute, and allow the same settling time at each rate. If a value changes by more than 2 per cent across repeats, stop the run and inspect the coupling for leaks before you continue. A small leak at the union will bias every later reading, so completing the check before proceeding is sensible despite the delay. Record the serial number of the reference meter and the date of its last calibration on the raw sheet. If that date lies outside the preceding twelve-month period, treat the run as invalid and obtain a reference meter with calibration dated within that period before proceeding. Keep the raw sheet even when a run is void, because the sheet still records what the instrument was doing at that time. When measurements at all three rates are complete, work out the mean error at each point and plot it against the permitted band. A meter whose error lies inside the band at every point passes the verification. A meter whose error falls outside the band at any single point fails the whole verification, and no mean across the three rates can alter that outcome. Sign the sheet, state the ambient conditions, and send the signed result to the quality file. Where a run is stopped early, write the reason on the sheet in plain words, because a blank field tells the next reader nothing at all. Store the sheets in date order so that a later enquiry can be answered using those records. If the meter is transferred to another bay, note the new position and the date of the relocation, because a meter that has moved may need a further check before its next use. Report any damage or faults affecting the reference equipment on the same day, and do not continue with a run on equipment you cannot trust. A doubtful reference makes every reading obtained with it doubtful too, and the cost of repeating a run is lower than the cost of correcting a bad result. When the file is complete for the month, check that every sheet has a signature and that no run is omitted from the sequence. A gap in the record is harder to explain later than a poor result recorded accurately at the time.

Replace words one at a time

Paste text to inspect the counted contexts in Declawd’s public test. Until the text reaches 200 distinct pairs, this mode shows only the count and how many more are needed. It does not show a score, verdict or editing table for shorter text.

Once there is enough text, select a counted green context, enter one replacement word and preview the full-passage result. Nothing changes until you approve that exact word.

Text to inspect

Full-passage score
Distinct pairs
Full-passage result

Scroll horizontally to see all columns

Counted green contexts in this full passage
Previous word	Current word	Scalar offset	Action

Listed Unicode characters

Finding or removing a listed character does not show that AI was involved. Joiners, direction controls, variation selectors and spaces can carry meaning.

This inspector uses a fixed, disjoint registry. It selects nothing by default and performs no Unicode normalisation, homoglyph folding or automatic replacement.

This browser mode removes listed characters only. The public command-line tool also supports exact, user-requested code-point replacements; this page does not replace or normalise spaces.

Language warning: joiners and direction controls can shape Persian, Arabic, Urdu, Hebrew and Indic text, or make it readable, so stripping them can damage that writing. Variation selectors can change how an emoji or ideograph renders.

Text to inspect

Scroll horizontally to see all columns

Listed characters found in the original text
Remove	Code point	Class	Position	Context