On Crazy Stacy, CoProducer’s score moved 70 → 90. Clipped samples went 831 → 0. Integrated LUFS moved −12.61 → −13.63. True peak improved from +0.33 dBTP. That is a delivery pass, not a remix, not a new 808, not a vocal tune.

People hear “AI repair” and imagine a different record. These three public pairs are the opposite. Same song. Same start time. Forty-five seconds. Switch A/B on one playhead. The change is ceiling, clip count, and loudness lane. If the chorus still bothers you after that, the mix is the job — not another limiter.

Three public pairs

The table under the player colors a fail red and a pass green against the same gates as Analyze: score ≥ 90, LUFS near −14, true peak ≤ −1.0 dBTP, clips = 0. Crazy Stacy’s after true peak is still +0.04. It stays red. That leftover is more useful than a fake 100.

What the engine will not claim

Phase and stereo width are mix work. Auto-repair does not “fix” a negative correlation. Clipped transients cannot be reconstructed; the caution stays even when output peaks are limited. The web analyzer therefore projects post-repair stats instead of handing you a downloaded master. If you want the WAV, that is the desktop app — one instrument, one license, no monthly upload club.

Time saved: you stop rendering five “final_final” limiters to chase loudness the store will undo. You measure, you cap, you listen for whether the clip count was the actual problem. On Drippin it was. On BeatGoHard it was only the ceiling. Different medicine. Same meter.

Did the score floor cheat the after number?

CoProducer will not let a repair print a worse score than the pre-repair file. These jumps are still earned by clips and peak, not by a floor alone.

Can I A/B my own bounce on the website?

You can upload for measurement and a projected repair column. You cannot download a repaired mix from the web form. That is intentional.