A second registration
Published ; updated
Declawd v2 repairs two problems in the first experiment: candidate words that did not fit their sentences, and a minimum length below the shortest passages used to choose the cut-off. The home page, checker, cleaning lab and method now use v2. The older articles and their v1 evidence retain their original figures.
Sentences
The first template could produce “due the delay” and “carry with a run”. V2 reviews every candidate in its complete sentence before the seed is drawn. The review covers grammar and meaning, including words around each slot. The candidate list and review are registered inputs, so changing them requires another version.
The first v2 run still produced “a empty field” and “a additional check”. That run is preserved as an invalidated attempt. The candidate errors required another registration before another draw. The active profile is declawd-v2-r2, with the same corpus, minimum and rule for choosing the threshold. The earlier seed and measurements remain available, including the errors that caused the replacement.
Lengths
V2 requires 200 distinct token pairs for a verdict. Calibration uses three views of each eligible passage: the prefix ending at 200 pairs, the prefix ending at 201 pairs and the full text. The registered rule takes the lowest cut-off on its fixed grid that meets the target in all three groups separately. A 199-pair prefix is recorded as insufficient text, with no score or verdict.
The evaluation repeats those views after the threshold is fixed. The method reports every group, including any target miss. Prefixes overlap and passages share authors, so the groups are dependent observations. Their intervals describe this corpus and do not establish a false-positive rate for human writing generally.
The corpus and author split were already published for v1. V2 draws a new seed against those known inputs. Its evaluation authors remain separate from calibration authors. Both halves were already public before this registration.
One draw
For each registration, the inputs and procedure are committed first. The registration is then committed before one seed is drawn and recorded. The resulting marked passage, cut-off, evaluation and controlled word changes are retained whether or not the marked example clears the cut-off. The draw has a commit record and a timestamp, without an independent witness.
The cleaning lab tries up to six changes using a rule fixed before that draw. At each step it chooses the largest available reduction from the reviewed candidate words, with a fixed tie order. Its published vector records the actual number of changes and the resulting score. A successful reduction concerns Declawd's own public rule only, and says nothing about Claude's watermark or the person who wrote a passage.
Results
The corrected run's marked fixture scored 0.77 against a cut-off of 2.30. Six of 96 full evaluation passages crossed the cut-off, a rate of 6.25 per cent against the registered 2 per cent target. The 200-pair group had one crossing among 96 passages, and the 201-pair group had one among 94. Two passages could not reach 201 distinct pairs.
The six controlled edits reduced the prepared passage from 0.77 to -0.21. It was already below the cut-off before those changes, so this run demonstrates a score reduction without a threshold crossing. The method keeps the calibration results, evaluation intervals and exclusions together. Neither the seed nor the cut-off was changed in response to these outcomes.
The archive
The v1 profile, seed, template, scores and base64 experiment remain available unchanged. Historical articles are labelled so their numbers are read against the experiment that produced them. V2 has a different domain separator and seed, so a v1 score is not a v2 score.