An AI tells you to cut 2 dB around 300 Hz on the bass and squeeze the drum bus. You apply both, hit play, and the mix sounds better. Or does it sound louder? That gap is where most AI mix advice gets accepted by mistake, because a small level increase reliably reads as "more detail, more punch, more polish."
This guide gives you a repeatable test you can run in any DAW, against any AI assistant, to separate a better mix from a louder one.
Why a new tool does not change the test
A forum thread headline describes EYEPIECE as a free Mac app that reads your DAW project for your AI (thread). This article draws nothing else from the thread and does not evaluate the app or its suggestions.
The point is general. Once an AI can see your project, it will hand you EQ, compression and balance moves with confident reasons attached. Confident reasons are not evidence that the move helped. Your ears, at matched levels, are the evidence.
What this article does not contain: No listening tests were run for this piece, so it reports no results. It includes no audio examples unless the editor adds them. Every setting below is a conservative starting point, not a measured outcome.
Before you start: set up the test
Pick a 15-to-30-second section of the song that includes the part the AI commented on. A chorus or a dense verse works better than an intro. Set a loop on it and leave it alone for the whole session.
Put a loudness meter on the master that reads short-term LUFS. Short-term uses a 3-second sliding window, which suits a short loop. If your DAW lacks a meter, use one from any loudness plugin you own. Use the same meter for every comparison.
Add a simple gain or trim plugin at the end of the changed version's signal path. It switches in and out together with the change, so you can level-match without touching the processing you are testing.
Step-by-step test protocol
-
Save a duplicate of the mix. Use Save As and name the file with the date and "original." Never test on your only copy. Then work in a second copy, or duplicate the track or bus you are about to change, so the original stays one click away.
-
Ask the AI for specifics, then pick one. Vague advice like "brighten the vocal" cannot be tested. Ask for the track, the frequency, the amount in dB and the reason. Choose a single suggestion. If the AI offers eight, rank them and test them one at a time, in order.
-
Apply the change at the AI's size or smaller. If the AI suggests a 4 dB boost, start with 1 to 2 dB. A smaller move is easier to judge honestly and easier to push further later. Record exactly what you did.
-
Build an instant A/B. Put the original and the changed version on a switch you can flip during playback. Bypass the plugin, mute and unmute duplicate tracks, or use a bus selector. Switching must not interrupt the loop or introduce clicks.
-
Match the levels. Play the loop with the original active and note the short-term LUFS reading once it settles. Switch to the changed version and compare. Use the trim plugin to bring the louder version down until both read within about 0.5 dB. Aim for 0.2 dB if your meter is steady enough.
-
Cross-check by ear. LUFS is a good guide, but a strong low-end or high-end change can still seem louder or quieter than the number says. If one version still feels louder after matching, nudge the trim a few tenths of a dB and re-check. Write down the final offset.
-
Make it blind, or blind-ish. Ideally a helper flips the switch while you look away and does not tell you which version is playing. Alone, you can label the two versions X and Y, flip a coin to assign them, and hide the plugin window. The goal is that you cannot see which version is playing.
-
Listen at two volumes. Run the comparison once at a low monitoring level, quiet enough to talk over comfortably, and once at your normal mixing level. Low levels expose balance and weight problems. Normal levels expose harshness and dynamics.
-
Switch several times and log your choice. Make at least five switches, and pick your preferred version each time without peeking. Consistency matters more than any single pick. Then reveal the answer and fill in the decision log.
-
Decide: keep, adjust or reject. Keep a change you preferred consistently at both volumes. Adjust one you liked at only one volume, or only in part, then retest the revised version. Reject one you could not distinguish or that you disliked.
Starting settings by change type
These are conservative first moves. Treat them as a place to begin, not a recommendation for your particular mix. We cover related ground in how to design a uk hardcore reese bass (daw guide) explained.
| Change type | Starting setting | Re-level note | |---|---|---| | EQ cut or boost | 1 to 2 dB, moderate Q, one band at a time | Boosts raise loudness; trim the changed version to match | | Compression | Ratio at or below 2:1, aim for 1 to 2 dB of gain reduction | Turn off or reduce makeup gain until levels match | | Balance (fader) | 1 dB moves on a single track | Any fader change shifts overall level; re-match | | Stereo width | Small increase or decrease on one element or bus | Check in mono and re-match level afterward |
What to listen for with each change
EQ
Listen to what happens to the neighbors, not just the target. A 300 Hz cut on bass may clear the kick, or it may thin the whole low-mid body. Check whether the vocal still sits in the same place. Check whether the cymbals suddenly seem brighter because the mud has gone. For more on this, see gamessoundcon 2026: worth it for composers?: the details.
If you can hear a difference only when you hunt for it, the change is probably too small to matter.
Compression
Watch what the compressor does to transients and sustain. With 1 to 2 dB of gain reduction, a good result sounds steadier and more present without feeling smaller. A bad result sounds duller, pumpier or flatter on the attack of snares and plucks.
Level matching is critical here, because compression with makeup gain is the classic way to fool yourself.
Balance
Balance moves answer one question: does the song's emotional focus improve? Check the vocal against the lead instrument, then the kick against the bass. Listen at low volume especially, since weak elements vanish first. Run the loop through a few full passes, because a change that helps the chorus can hurt the verse.
Stereo width
Switch to mono for a few seconds. If the element loses level, hollows out or disappears, the added width has a cost. Then return to stereo and ask whether the widening opened up space or just pushed elements away from the center, where the song needs them.
When to trust your ears over the AI
Trust your ears when you prefer the original at matched levels at both volumes. The AI may be right about what a frequency chart shows and still wrong about what your song needs.
Also read: see phobos vs polaris vs phoenix: which bt synth to buy
Trust them too when the AI's reason does not match what you hear. If it says a region is harsh and you hear nothing harsh in the loop, the claim does not apply to this mix. Likewise, if the change helps only after you turn it up, you have found a level effect, not an improvement.
Give the AI the benefit of the doubt when you struggle to hear a difference but the suggestion fixes a problem you already knew about, such as a masked vocal. In that case, keep the smaller move and retest in context on the full song.
Treat the AI as a source of hypotheses. Level-matched listening is the experiment.
Printable decision log
Copy this table into a document or onto paper. Add one row per suggestion.
| # | Suggestion (and AI's stated reason) | Change applied | Level offset (dB) | Preference (blind, low / normal volume) | Keep, adjust or reject | |---|---|---|---|---|---| | 1 | | | | | | | 2 | | | | | | | 3 | | | | | | | 4 | | | | | | | 5 | | | | | |
Under each row, add a notes line for the loop's timestamp, the meter reading of each version and the number of switches that favored each side. After a few songs, the log can reveal a pattern, such as an AI that keeps suggesting low-mid cuts you consistently reject.
What to expect next
Your first few runs will feel slow, and some suggestions will end up as "reject" or "adjust." That result tells you how well an assistant fits your genre and your monitoring. After a handful of songs, the log will show which kinds of advice earn a place in your workflow.
To add real audio examples to this process, record your own matched A and B bounces and keep them next to the log.
