Going in, the speed I had a right to expect from a working motor at 5 Hz was on the order of 0.6 µm per second, scaled from the source design's published performance [first_motion_playbook.md §1]. Watch that motor with perfect attention for thirty seconds and it accumulates about 18 µm of travel (Derived), a fraction of the width of a hair, against no reference mark an unaided eye could use anyway. A dead motor, watched the same way, accumulates zero. At this scale, to a human eye, a working motor and a dead one are the same sight.
An experiment whose outcome looks identical whether the hypothesis is true or false has a name. It is blind. Not blind in the good, controlled, clinical sense, but blind in the bad one. Whatever happened on that rail, I was going to see the same nothing. I ran experiments like that for about a month, mostly without understanding that I was doing so. This chapter is what that was like from the inside and what finally broke it, and I think the mistake in it is the one most likely to be somebody else's too.
A month without results
From the outside, and honestly from the inside too, it was ordinary [builder, 2026-08-07]. Power on, the stack ticking its dry 5 Hz tick (§M7). Set a preload with the calibrated screw, watch the slider, see nothing. Add a tick, watch again, nothing again. Over these sessions I walked the preload across the whole plausible working range, zero to 5.5 N [first_motion_playbook.md §1; my recollection said "about five," and for once memory and the record agree], which on paper is a systematic sweep of the parameter stick-slip physics is most sensitive to.
Every setting returned the same result, and every result felt like information. By the end I was close to certain the motor simply didn't work [builder, 2026-08-07].
That certainty is the trap, and it deserves a precise description. The nulls felt like data because they were tedious to produce, and effort feels like evidence. But given the arithmetic above, "no visible motion at this preload" was compatible with a broken motor and equally compatible with a working one. A null that cannot distinguish success from failure feels exactly like a null that can. Nothing about standing at the bench tells you which kind you're collecting.
One number decides how much of that is fair to me, and it is how long a single trial ran. Thirty seconds of watching really is blind at this speed. Watch long enough against a fixed mark and the same motion becomes visible, which is exactly how first motion was eventually caught (§M11). So some of what I am calling blindness was a detector problem and some of it may have been a patience problem.
Eliminating variables
What do you do when you can't interrogate the output? The only move I had was to eliminate suspects feeding the input, and §M9 told that story from the hardware side. The commercial rail, whose breakaway drag I had every reason to think exceeded the entire drive budget, was replaced with the bearing-bolt rail built to be as close to frictionless as printed parts allow. §M9 is blunt that neither rail's breakaway was ever actually measured, so this was a suspect removed on strong suspicion rather than one convicted on evidence. It still mattered here, for a specific logical reason. If nothing moved on the freer rail either, the drag explanation was finished [builder, 2026-08-07].
Which moved the drive to the top of the list, the suspect I had been vouching for on paperwork alone. It was not the only one left, as §M11 would eventually make clear, but it was the only one I could put a scope on.
Instrumenting the input
With the mechanics cleared as far as I could clear them, I put the oscilloscope on the drive signals at the stack, honestly expecting just to double-check that everything made sense [builder, 2026-08-07]. It did not make sense. That careful reading of the signals is what found the drive's faults, and per the records they had been there the whole time: the board's real output was half of what the documents claimed, its waveform fell apart if asked for anything much faster than the bench's 5 Hz, and one channel was quietly underperforming its neighbor. §M7 tells each of those discoveries properly. The summary that belongs here is harsher. The month of blind sweeps had been characterizing a motor whose input was broken, through an output nobody could see.
So the blindness ran in both directions, and naming both is the real lesson. Output-blind, because physics made micrometer motion invisible, which I understood in principle by mid-month, when the first-motion playbook formalized the arithmetic [first_motion_playbook.md §1, 2026-07-13]. And input-blind, because I had assumed the drive was healthy on its documentation's word, and §M7 is the chapter-length receipt for how that assumption went. Either blindness alone wastes evenings. Together they compound into something worse, testing a broken input against an unreadable output, an experiment that can only ever return silence.
The feeling
What I felt when the scope finally named a culprit was not defeat. It was relief, extreme relief [builder, 2026-08-07]. It had been a long time since the bench had given me an answer to anything, and here one finally was, even though the answer amounted to "your drive has been lying to you for weeks." Most project write-ups put their emotional payoff at the moment the machine works. Mine arrived earlier, at the moment I learned why it didn't. A diagnosis with a cause is worth more than a month of nulls without one. It felt like the project restarting.
Instrumented measurement
The formal lesson was already on paper by mid-month, and the playbook's own one-line verdict says it better than I can from this distance, "Your no-motion test is almost certainly blind, not broken." Fix the readout, not the speed [first_motion_playbook.md]. The playbook's two inequalities, µN > F_drag for the stick phase and m·a_flyback > µN for the slip [first_motion_playbook.md §2], double as a list of ways this motor can fail in complete silence, and as a reminder that a parameter sweep is only ever as good as the detector judging it. My preload sweep wasn't a bad experiment badly run. It was a fine experiment judged by a detector (me, squinting) with no sensitivity to the effect.
Two instruments ended up replacing the eye, and a third is built and waiting. The oscilloscope on the input, which found the disease. The Michelson interferometer on the output (§M8), built so that 260 nanometers of travel becomes one countable fringe, though it has not yet been pointed at the motor. And the humblest one, which needs no instrument at all, accumulation. Let a slow motion run long enough against a fixed mark and even a naked eye becomes an instrument again. That last one, a strip of masking tape and ten patient minutes, is how first motion was actually caught (§M11).
The definition of "experiment" I now hold is the one thing I'd ask a reader to take from this chapter. An experiment is not the act of doing something at a bench. It is an arrangement in which the outcome is capable of changing what you believe. For about a month, mine couldn't. Everything downstream of this chapter, the validation battery and the controls that go with it, exists so that no result in this project is ever that silent again.