A third registration

Registration checks the template identifier and a lower bound of 286 distinct token pairs formed entirely inside literal text. Those pairs survive every candidate choice, so the marked passage can reach the registered 199, 200 and 201 boundaries for any seed. These checks run before the single draw.

The inputs, registration and seed have separate commits. One 32-byte seed was drawn on 2 November at 17:30:21 UTC and committed before scoring. The draw has a timestamp and a commit record, without an independent witness.

Every output carries the registration hash, seed and seed-record hash. Reproduction checks those bindings in every surviving output before writing, including when the calibration report is missing. Complete temporary files are published atomically, so an interrupted replacement preserves a complete earlier file or leaves the destination absent. Git checks now use the correct paths when the experiment sits inside a larger repository.

The scorer validates the whole profile before changing its global parameters. A missing field or invalid threshold therefore leaves the previous state intact. The domain separator is derived from the profile identifier, allowing a later profile to use its own domain without editing a list of accepted versions.

The minimum remains 200 distinct token pairs. The registered rule selected 3.30 from calibration authors, taking the lowest cut-off on its fixed grid that met the target in every eligible length group. That value was frozen by the algorithm before evaluation authors were scored in the same run.

The marked fixture scored 1.90 across 380 distinct pairs. The default control scored 0.33 across 377. Both were below 3.30. Full evaluation crossed once in 96 passages (1.04 per cent). The marked miss and the crossing remain part of the recorded outcome.

The registered edit rule tried up to six substitutions from the reviewed alternatives, choosing the lowest resulting score among unused slots with a fixed tie order. All six steps were retained. They reduced the marked score from 1.90 to 0.53, with both endpoints below the cut-off. The trace records a score reduction without a change from detected to undetected.

The 192 human passages and author split were already published and examined in earlier Declawd studies. Calibration and evaluation authors remain separate, but both groups were known before this registration. These counts describe the reused corpus and this key. They supply no population error rate or measurement of a provider's production watermark.

The root r2 files, first v2 attempt and earlier figures retain their bytes. A separate r2 archive has its own scorer and an exact file manifest. Future registrations use the separately bound guard described below. The second-registration article, method and active browser demonstration continue to report r2.

A further review found that the generic v3 length helper can overcount when adjacent literal strings split a token. The committed template has separated slot boundaries and at least 290 guaranteed pairs, above its recorded conservative bound of 286. Its fixture, seed and results remain valid. The future-registration guard joins adjacent literals, checks lexical boundaries and binds its source and preflight receipt before sampling. No new draw was made for that correction. The guard review records the counterexample, source hashes and 21 tests with mocked entropy. The frozen v3 tool is retained for reproduction.

The registered history verifier works on its preserved experiment history. A squash merge gives the main branch a different ancestry, so the separate history checker compares the current experiment bytes with the exact evidence tag and checks the original commit order there. The tag must resolve to the recorded commit. Missing or moved tags fail the check. The history review records the checked commit order and six regression tests.

The evidence tag preserves the original registration history, with the recorded outputs pinned at commit 06fac8a. The reproduction command reconstructs all five outputs from the recorded seed. A separate verifier checks the exact bytes and the order of the registration, seed and output commits. The 24 successor regression tests and 150 existing reference tests passed, and the r2 archive reproduced its historical outputs.

The site serves the registration, calibration, evaluation, edit trace and file hashes. The build independently checks the selected threshold, verdicts, scoring vectors and six edit steps.

This evidence accompanies the published v0.3.3 release. The downloads page records the published package and binary versions. The evidence tag identifies the experiment independently of that release process.