It fit my notes, not the phone

Last week I designed a 3D-printed case for a brand-new phone. Before anything reaches my printer, the design has to pass a set of automated checks. Think of them as gates. The computer draws a picture of the finished part: clean. It checks that the digital shape is one solid piece with no holes or cracks in its skin: pass. It checks the printing rules, like no wall thinner than a credit card: pass. Every gate opened. I logged "it fits."

The phone could not physically have entered the case.

The gates were not wrong. The gates were checking the design against my numbers. The problem was my numbers.

The rest of this is a 3D printing story. Until it isn't.

I do not sculpt parts by hand in a design program. I write a description of what I need, and an AI agent turns it into code where every dimension is a number: the case is this long, the walls are this thick, the camera opening is this wide. Change a number and the shape updates. Nothing is eyeballed.

Every design then climbs the gate ladder before it goes near a printer. My hands and the printer's hours are the last check in the chain, not the way I find out something is wrong.

That ladder exists because of one bad afternoon. A 3D printer runs on a file of machine instructions, thousands of tiny commands like "move here, squeeze out this much plastic." Early on I wrote one of those files by hand and sent it straight to the printer. It sat on "preparing" for thirty minutes and never printed a single layer. Machine instructions are never written by hand now. The software generates them, or the job does not run.

So production is fast, checked, and mostly hands-off. Which is exactly what made the next failure so convincing.

To design a case, you need the phone's exact measurements. Those come from the manufacturer's official drawing: a many-page document, dense with numbers. I took notes from it. Somewhere in the note-taking, I copied the measurements for the camera bump (the raised block around the lenses) off the wrong page: a page about the phone's antennas, whose numbers look exactly like the ones I needed.

The design matched my notes perfectly. Every gate agreed. I called it a fit.

Then a plain look at the design said the camera opening seemed off, and I went back to the drawing itself instead of my notes about the drawing. The camera bump would have pressed into solid plastic. The phone could not seat. My "it fits" had checked the design against my copy of the truth, and the copy was the thing that was broken.

The drawing is the truth. Verify against the source, not your notes.

There is a second layer to this, and it cost a reprint to learn.

Even correct numbers from the source are only correct on paper. A case has to be slightly bigger than the phone it holds, and that little gap is called clearance. I started with the textbook gap, a quarter of a millimeter per side. The first time a printed case met a real phone, the fit was wrong, and the gap had to shrink to a tenth. The book number was not wrong; it just was not my printer's number. Every printer squeezes plastic a little differently, so the same file comes out a hair tighter or looser on different machines.

So now I print a tester first: a small throwaway part with a row of gaps, from very tight to very loose, and I check by hand which one my printer actually gets right. The tightest gaps come out jammed, or fused into one solid piece. The widest are sloppy. Somewhere in the middle is the gap that slides together the way a case should. That ten-minute test turns a book number into my number before it gets trusted in a real design.

The ladder of trust runs: my notes, then the source document, then a real-world measurement. Each level beats the one below it. The phone case failed because I let the bottom level stand in for the middle one.

Everything I delegate inherits my references with perfect fidelity. An AI agent building from my summary of a document will honor every error in the summary, beautifully, through every automated check that measures against it.

That is the trap of good tooling: perfect execution against a corrupted reference is the most convincing form of wrong. Every check passes, because every check sits downstream of the reference. Nothing in the pipeline re-reads the original unless a person decides that re-reading the original is a step.

It is now a step. Any verdict that matters gets checked against the original document, not my notes about it. And where a real-world measurement is possible, the measurement outranks the document.

Notes are copies. Copies drift. Check the verdict against the source.

When was the last time something passed every check you had and was still wrong? Hit reply. I read everything.

Christopher