More windows
Published ; updated
Searching a passage creates more chances to find a high score. Our registered mixed-passage experiment flagged 19 of 24 human evaluation documents in its longest group when the window search used the whole-document cut-off. A separately calibrated search cut-off reduced that count to one. It also missed most marked insertions.
Two cut-offs
The whole-document cut-off was 1.65. The maximum over the window search needed a separate cut-off of 3.40 to meet the registered calibration rule. Each procedure crossed zero times among 24 calibration backgrounds at each registered length. The backgrounds were split by author before the run.
Scroll horizontally to see all columns
| Background tokens | Whole document | Adjusted search | Uncorrected search |
|---|---|---|---|
| 128 | 2 | 0 | 3 |
| 384 | 1 | 0 | 11 |
| 768 | 4 | 1 | 19 |
The uncorrected search applies 1.65 to the highest window score. The adjusted search applies 3.40 once per document. Its one crossing among 24 human backgrounds at 768 tokens is 4.17 per cent, above the registered 2 per cent evaluation target. The whole-document method exceeded that target at every length. The thresholds and outcomes were retained.
Weak insertions
We inserted 64, 128 or 256 tokens of a marked template into the start, middle or end of each evaluation background. Across 648 marked mixtures, whole-document scoring flagged 125, the adjusted search flagged 19 and the uncorrected search flagged 597. Six adjusted detections located the insertion under the registered overlap rule. The other thirteen did not.
The isolated marked prefixes scored 1.405, 2.382 and 1.142 respectively. Each was below the adjusted search cut-off. These are outcomes from one key and template. The longer prefix supplied a weaker score than the middle one, and the stricter search missed nearly all mixtures. This run establishes no detection improvement from the adjustment.
Controls
The same insertion positions and lengths were tested with a neutral template that chooses the first candidate in each slot. Whole-document scoring flagged 66 of those 648 mixtures, the adjusted search flagged nine and the uncorrected search flagged 369. These authored neutral fixtures are separate from the historical human controls.
A high-scoring span can fall in the human background or across an insertion boundary. Searching more places changes the complete procedure that needs calibration. Individual windows are not the denominator for a document's false-positive rate.
Limits
The experiment reuses the historical public-domain corpus and reviewed template from Declawd v1/v2. Its 48 backgrounds supply 24 calibration and 24 evaluation backgrounds from different author groups. Prefixes, positions and passages within an author are dependent observations. The counts describe this corpus and the recorded educational key, with no population guarantee or measurement of a production watermark.
Reproduce
The plan, implementation and inputs were committed before the single public key, followed by calibration before evaluation. The registered experiment includes the commit sequence and reproduction commands. Its independent verifier reconstructed all 116,256 windows, thresholds, verdicts and location results.
The site publishes the evaluation, calibration, registration and served file hashes. The original v1/v2 contracts retain their bytes.