Figures refer to our recorded November 2026 run. Detector-dependent conclusions remain provisional while collection is incomplete.
The quality evaluations are complete
All 759 saved rewrites have readability and meaning-preservation measurements under the shared profiles. These do not depend on a detector’s availability. They let us inspect clarity and semantic preservation while the seven-detector panel finishes collection.
Readability and meaning answer different questions
Readability is a blinded judgment of clarity, fluency, and naturalness. Meaning preservation is embedding cosine between the source and rewrite. Smooth prose can still alter a claim; a faithful paraphrase can still read awkwardly.
Unchanged text needs to remain visible
The shared classifier currently identifies 16 whitespace-normalized identical outputs across the 759 recorded tests. Identical text can preserve meaning very well while failing to rewrite the input. The original scorer handles that with a fixed penalty rather than hiding the attempt.
What must wait
Bypass uses the median detector human scores for each output, and consistency uses their spread across writing categories. The composite depends on both. Until detector coverage is complete, an apparent overall leader remains provisional.
We publish the quality values separately so readers can inspect what is complete without mistaking a partial composite for a final result.