One framework for every saved rewrite
Both cohorts use the same original source texts, measurement profiles, detector routes, classification heuristic, and HumanizeBench v2 weights, with the original metric calculations and penalties preserved. Historical detector and quality scores do not fill gaps in the new comparison.
No output selection by score
We preserve the recorded generation for each model/sample pair. Detector outcomes do not select a better rewrite. Technical measurement retries keep the same text, and successful measurements are reused for identical text and profile.
Corrections must be traceable
Artifact changes, route amendments, technical failures, and revised measurements must retain their provenance. A correction should identify the entry, sample, relevant raw evidence, and affected scoring rule.
Use the local submission form to record a review request for this preview. It is not a live public support inbox.