Field note · 2026-07-03

The AI Mix Engineer That's Allowed to Say I Don't Know

Post Mortem reads one REAPER track, separates facts from guesses, and lets you preview one safe fix. The engine stays open; the panel is paid.

aimixingreapermusic-productionaudioopen-core

July 17 update: The first version described below grew into a docked paid panel with verified Fix Preview, one-step apply and undo, a native installer, and offline licensing. The engine and Reaper Daemon remain MIT licensed. The $39 Apple silicon early-access release is live with private installer and unique offline-license delivery.

I pointed it at one of my own guitar tracks expecting to catch it in a lie.

It came back and told me there was a resonance sitting around 50 Hz, six or seven dB louder than the rest of the low end, underneath the lowest note that guitar can actually play. It noticed there was no high-pass filter engaged anywhere in the chain. It gave me one move. Turn the Lo-Cut on the amp from off up to somewhere around 80 to 100 Hz. Then it did the thing I did not expect. It told me it might be wrong. The resonance could be the cabinet impulse instead of the amp, it could not tell which from where it was standing, so it rated its own confidence medium and left it there.

I soloed the track and listened. It was right. The 50 Hz was exactly where it said, doing exactly what it said. And the hedge was honest too, because gun to my head I could not have told you off the top of my head whether it was the amp or the IR either.

That moment is the whole reason the tool exists.

The thing I could not find

I have been producing instrumental death metal for about ten years. I know how to mix. What I do not have is a second set of ears at two in the morning when the low end feels wrong and I have been staring at the same eight bars long enough that I cannot hear them anymore.

So I went looking for something that could listen and tell me the truth. Most of what I found could listen. Almost none of it could tell the truth. You feed a track to an AI and it hands you a paragraph of confident advice, and half of it is generic, and none of it tells you how sure it is. It will tell you your mix is muddy the same way a horoscope tells you this is a week for change. Technically it might be right. It has no idea, and worse, it will never admit it has no idea.

That is the part that makes me close the tab. Not that a tool is wrong sometimes. Everything is wrong sometimes. It is that the tool sounds exactly as certain when it is guessing as when it knows, and that leaves you doing the one thing you came to it to stop doing, which is not trusting your own ears.

What it actually does

Post Mortem reads the real state of one track. Not a vibe, the actual state. Every plugin on it and where the knobs are currently set, the sends and receives, the bus it feeds into, the fader, the pan. Then it renders a ten second stem of that track exactly as it sounds after all the processing, and it measures that stem. Integrated loudness, true peak, crest factor, a third-octave picture of the whole frequency range.

Those numbers, the real ones, go to a model that answers like an engineer instead of a brochure. What it sees in the measurements. What is probably causing it. One concrete move, with the plugin named, the parameter named, the current value, and the value to change it to. And how confident it is. The raw WAV stays on the computer.

Here is a real one, run on a kick drum.

The kick is hitting -0.1 dBFS sample peak with -8.7 LUFS integrated and an 11.65 dB crest factor, a level that leaves essentially no headroom for the parent Drum_Buss to do its work. Tonal shape is otherwise reasonable, a strong fundamental at 50 Hz, a clean scoop through 200 to 1000 Hz, and a healthy click region at 2 to 5 kHz. The problem is level and dynamics, not EQ.

Confidence: Medium. The over-hot level and the Kontakt receive are directly verifiable. The polarity read fits the metrics but is still a hypothesis until the phase-flip test is done.

Read that last paragraph again. It separated what it knows from what it is guessing, out loud, without being asked.

It is not allowed to bluff

This is the part I care about more than any of the measurements.

A Post Mortem Track Check sees one track. It does not see your mix. So it is forbidden from telling you your guitars are masking your vocal, because it cannot possibly know that, and it says so instead of pretending. Every diagnosis carries a confidence rating, every time, no exceptions. When the data is thin the diagnosis says the data is thin. An honest I am not sure beats a confident wrong answer every day of the week, and that rule is written into the core of the thing where it cannot be talked out of it to make the output sound more impressive.

That was a choice. It would have been easy to let it swagger. Confident tools demo better. They feel smarter in the first thirty seconds. But I have spent enough years around plugins that promise to be the last one you will ever need to know exactly how that movie ends. The person who actually knows a subject talks in specifics and caveats. This is good for that, but if you are doing this other thing you want the other tool. The person one video deep says this changes everything. Certainty is the tell. I did not want to build another thing that swaggers.

One track. One move.

It gives you one move on purpose. Not five. Not a checklist.

I know what it is to open a plugin with forty parameters and feel your eyes cross and walk away from something that might have been perfect for you. The problem was never that you are not smart enough. The problem is that you have two thousand plugins and half of them have not been opened since you installed them, and one more wall of options is not help, it is noise wearing a lab coat. So Post Mortem picks the single thing most worth doing and hands it to you with the exact number. The paid panel can try that value temporarily, capture the result, and restore the original. You listen. You decide. It is a second opinion, not a driver.

What it will not do yet

I want to be straight about the edges, because the opposite is exactly the overclaim I just spent four paragraphs complaining about.

The 0.1.1 paid panel has Track Check and Fix Preview. It does not have Mix Check, history, hosted credits, real-time monitoring, automatic plugin insertion, routing edits, item edits, or automation writing. The verified-isolation path is reliable for item-less routing and bus tracks; ordinary tracks containing media items may be refused because the render could contain the full mix. Windows and Linux paid installers stay withheld until their clean-machine customer paths pass.

That is narrower than the finished product I have in my head. It is also the honest edge of what this version can prove.

Built on the folder

Post Mortem can read a track and render a stem because of the bridge underneath it. That is Reaper Daemon, a free file bridge that lets a program read and drive a REAPER session with nothing but files in a folder. I wrote about why that design won on being boring here. Post Mortem is what happens when you build something real on top of a foundation you can actually trust.

The engine, command-line tools, and Reaper Daemon are free and open source under the MIT license. The docked panel, installer, and offline licensing are the paid product. Apple silicon early access is $39 one-time, with no required subscription. It includes the current Track Check and Fix Preview release, permanent use of version 1, and 12 months of updates. Mix Check, local Sonic Memory, history, hosted checks, and automatic updates are planned work, not features in version 0.1. There is a page for it here, and the public engine is on GitHub if you want to read the part that measures and reasons before you run it.

The best thing it does is the thing most tools are too proud to do. Sometimes it looks at your track and tells you it is not sure. That is not the tool failing.

That is the tool being worth trusting.

Related

Keep reading.