In April I signed this project up for the Liao 2022 open-source three-channel piezo drive board, Arduino Mega into DAC0800s into TDA2050s, roughly $50–80 in parts [11; project CLAUDE.md §2]. The design was published alongside a peer-reviewed paper, and anyone could download the schematic, the firmware, and the Gerbers, which are the copper-layer drawings a factory uses to make the board. I sent the Gerbers to a fab and got back a stack of professionally made boards that could not work. There was no ground pour on either copper layer, and sixteen holes the schematic said were grounded connected to nothing.
It took me weeks to believe that, because for most of them I assumed the broken thing was me.
Ordering the boards
The board's design files came from the authors' OSF repository. The boards that came back had my project's future printed on them in silkscreen. Populating the first one was my soldering final exam (§M3b): SMD 0805 passives, through-hole ICs, electrolytics, the works.
First signs of a problem
The suspicion started small. Checking the populated board against the published schematic, things didn't line up. Points the schematic said were ground didn't behave like ground under the multimeter, and continuity that should have existed didn't.
I want to document my honest first reaction, because I think it's the most universal part of this story. I assumed the mistake was mine. Obviously it was mine. I was a few weeks into electronics, holding a board designed by a research group and published alongside a peer-reviewed paper. The prior probability that I had botched a joint, misread the schematic, or misunderstood what a ground pour even was seemed overwhelmingly higher than the alternative. So I spent real time trying to find my own error, and the measurements kept disagreeing with my self-blame. One error on that board really was mine, a bad ground-end joint on R3 of DAC channel 1, caught and reflowed along the way [builder correction, 2026-08-06; earlier project notes had filed this joint under the v2 bring-up, see the corrections log]. But one reflowed joint couldn't explain missing continuity all over the board. The problem was bigger than my soldering.
Eventually the disagreement between the schematic and the physical board got sharp enough that I went to the source, the Gerber files themselves, opened in a Gerber viewer, layer by layer.
Sixteen floating holes
There was no ground pour, not on the top copper layer and not on the bottom one. And without a pour, every through-hole that was supposed to connect to ground through it connected to nothing. Counted against the schematic, sixteen ground-net through-holes were truly floating: U1 pins 1 and 2, U2 pin 1, and P2 pins 1 and 10, on each of the three channels, plus P1 pin 2 [project CLAUDE.md §2, inspection confirmed 2026-05-25].
That is not a subtle defect. The board's entire ground scheme was absent. As fabricated, the board could not function. The files everyone could download, the ones I had paid to have manufactured, were missing the one copper feature that made the circuit a circuit.
The feeling was two things at once: vindication (my measurements had been right for weeks) and something closer to unease. If the famous published files were wrong, then "published" stopped being a synonym for "correct," which is a real rite of passage.
Contacting the author
I wrote to Prof. Edwin Hwu, whose group published the board, on May 22, sending the evidence from the Gerber viewer. The full sixteen-hole count went into the log three days later, on May 25. Then I waited.
Nothing came back the first week. I didn't send a follow-up. One clear email with the evidence felt like the right amount, and pestering a professor didn't. But a project can't idle on hope, so the silence forced a decision that turned out to be one of the best of the project. If the files might never be fixed, I'd rebuild the board myself. That's how KiCad entered my life, and §M3c tells that story: the schematic re-drawn from the paper with the ground net done right, a partial layout, a real fallback. Waiting with no fallback is not a plan.
The reply
Eight days after my email, the reply arrived (I went back and checked the timestamps: sent May 22, answered May 30). It was brief, and it had corrected Gerber files attached.
I'll be honest that after the buildup, the brevity was almost funny. No drama, no explanation of how the broken files ended up published. Just the fix, from the person who could produce it in minutes where I would have needed days. Both things I felt about that are true: a longer reply would have been nice, and a working file beats a warm paragraph every single time.
The corrected Gerbers went to JLCPCB the way the originals had gone to fab before. The new boards came back with proper pours and, in one small mercy, the original designators on the silkscreen, which instantly obsoleted the designator-mapping table I'd built for my KiCad version (§M3c), a fate I've made peace with.
Second population: a working board
The v2 board got populated in one overnight session, June 11–12, all fresh parts, DAC0800s in sockets, fast because it was a redo of a board my hands already knew (§M3b). Then bring-up, in stages, against the project's written protocol, and this time the board came up clean. Nothing flagged, nothing to reflow [builder, 2026-08-06; the one bad R3 joint of this story belongs to the v1 board, a few sections up]. (The protocol itself wasn't above the fray either. Its first revision had one wrong row, telling me a compensation capacitor went to ground when it actually goes to −15 V. Even checklists have bugs, and the corrections log caught this one, which is what corrections logs are for.)
And then it worked. Three channels of drive electronics, built from open files via one broken fab, one detective arc, one unfabbed rebuild, and one brief email with the right attachment.
Lessons about open hardware
If you are about to fab a published board, open the Gerbers in a viewer first and continuity-check the ground net against the schematic. Ten minutes of that would have saved me a board order. The second thing I would tell you is to believe the multimeter over the PDF, including when the PDF is peer-reviewed and you are a few weeks into electronics and certain the problem is you. Mine was right the whole time.
Start the fallback during the silence, too. I never had to fab mine, and building it was still the right call.
And the loop does work. A broken file, a polite email with evidence, and a fixed file back, eight days later.
On redistribution. The corrected Gerbers came to me in private correspondence, so I asked before re-hosting them. The answer came back on August 9, "The online repository is not changeable, you can make yours so I can direct other people to use your fix." So the corrected files live in this project's repository, marked as the fixed version, credited to the paper's authors, with a note on what changed, under the same CC BY-SA 4.0 license the original design files carry on OSF. Anyone he sends there can find them right away.
The board now existed and powered up. It still had to prove it could turn firmware into a clean 15-volt sawtooth, which became its own adventure in measurement, correction, and one number the whole project had been quoting wrong (§M7).
Postscript (2026-08-20)
The loop closed better than this chapter dared to hope. In mid-August I sent Prof. Hwu a photo and a video of the motor his board now drives. He answered within days. He'd like more videos to share on his YouTube channel, linking back to this repository, and attribution without co-authorship is fine by his group. He also read the hardware from one photo and sent back two suggestions: a magnet-ball foot on the steel of a utility-knife blade for a durable contact, and his paper's rigid magnet coupling for the stack, with the one-line diagnosis that printed parts are soft and absorb the stack's motion. Both are being applied as I write this. A saga that started with broken published files ends with the original author advising the rebuild and offering to amplify it.