// project log
A chronicle of engineering by repeated mistakes.
There are plenty of websites that show the finished project.
This isn't one of them.
This is where the ugly prototypes live. The burned components. The backwards connectors. The notebook pages. The scope captures. The measurements that contradicted the assumptions.
Projects are intentionally unfinished. The goal was never to build the thing — it was to become the engineer capable of building Version N+1.
Iteration isn't just the journey.
Iteration is the destination.
// site conventions
Every project documents the evolution of understanding — not the evolution of hardware. Six blocks, every time.
ACTIVE
A three-stage coil gun — three coils, three optical gates, and a capacitor bank that took four versions to stop being the interesting problem. It is a timing project wearing a physics project's clothes.
MOTHBALLED
A dual-axis pan-tilt platform that has to hit what the camera is looking at, inside a constraint set written by a city council. Shelved at v06 — but its pan-tilt controller is next door running PROJECT #001, which is the actual result.
Mothballed before it was started, which is a personal best. The parts are in a box with a label on it and that counts as documentation.
A three-stage coil gun, built to find out where the velocity actually goes. It is a coil gun and not a rail gun — establishing that took one whole version and a mislabelled keycap. Four versions in, the answer has moved from "the capacitors" to "the gate timing" to "the coil resistance," which is three different projects wearing the same enclosure. The coil gun is the tuition.
-- is not broken; it is reporting. Gate spacing was shorter than the sensor's own response time, so the transit was over before the second edge existed. The measurement chain was the experiment and nobody had tested it.
Four versions spent on switching and gate timing, and the discharge was never oscillating to begin with. At ζ = 1.52 the current creeps up and decays with no reversal, so the field peaks late — after the slug has already passed the coil centre. That is the suck-back, and it is caused by resistance, not by leaving the switch on a moment too long.
→ v05: rewind stage 1 in heavier wire for R < 2.8 Ω and re-run the same 11-shot series.
For four versions this was filed as a timing problem. It is a resistance problem. At a damping ratio near 1.5 the discharge never oscillates — current rises slowly, peaks late and decays without reversing — so the field is still pulling backwards when the slug is past the coil centre. No amount of switch timing fixes a waveform shaped like that.
Stored energy has not been the constraint since v02. Going from 10,000 µF at 35 V to 1000 µF at 450 V bought sixteen times the joules with a tenth of the capacitance, and exit velocity barely noticed. That is the clearest evidence that the loss is downstream of the bank.
The measurement chain has to be commissioned before it is trusted. Eleven runs were spent firing into an instrument that could not, in principle, produce a reading. That was not a sensor fault — it was never testing the sensor against something of known speed.
What is still unknown: whether the surviving false trigger is conducted through the shared sensor supply or radiated into the gate line. The v04 isolated supply is the experiment that answers it, and it went in this morning.
If stage 1 is rewound to bring the discharge from overdamped to roughly critically damped, does exit velocity move at all — or has the sequencer been solving the wrong problem accurately since v02?
A dual-axis pan-tilt platform that has to put a soft projectile where the camera says it is looking. The pun is load-bearing: the binding constraint here has never been physics, it has been a municipal ordinance and a property line. Seven versions in, the projectile got softer every single time and the pan-tilt pointing got better every single time, in that order. Mothballed at v06 — not because it failed, but because the part worth keeping turned out to be the pan-tilt controller, and that has moved next door to PROJECT #001.
The resolution is sixteen times finer than the error. Every microstep past the first sixteen is the controller confidently commanding a position the gear train has not agreed to yet. Raising the gain to close that gap faster is how you build an oscillator instead of a tracker — which is precisely what happened, and it looked like a tuning problem for three weeks.
→ v07: preload the pan gear train, re-measure θb, then re-tune. In that order.
This was never a controls project. It is a gear-train project that has been wearing a controls project's clothes since v02. Backlash of 1.8° on the pan axis against a microstep of 0.1125° means the pan-tilt's position resolution is sixteen times finer than its position error, and no amount of loop tuning closes a deadband.
Mounting the camera to the tilt axis was the single most expensive decision in the project. It makes the pointing error structurally invisible to the only instrument on board, which is why the error survived four versions without ever being measured properly.
The payload does not matter and never did. Pneumatic launcher, gel blaster, foam noodles — each swap took a weekend and none of them changed the accuracy. The platform is the project; the ordnance is a fixture.
The genuinely useful output of this project is the controller. It has revised once in fifteen months, it now drives a completely different machine on PROJECT #001, and it was built as scaffolding. That is why this platform is mothballed rather than finished: it answered its question, the answer was "you were building the wrong thing," and the right thing walked off to another bench. What is still unknown is how big the pointing error actually is, and that cannot be answered until there is a reference the turret is not bolted to. The rig is on a shelf, intact, waiting for someone to build that reference.
Measured against a reference the platform cannot move, how much of the pointing error is backlash — and once the gear train is preloaded, is what remains small enough that a residential property line stops being the binding constraint?
// log
Reverse chronological. Every entry is something that happened at the bench, not something that got finished.
A 15 V 50 mA dual rail with battery protection and a 0.1 A ultra-fast-recovery diode, built and labelled this morning. Both dead v2 controllers failed on the sensor side rather than the power side, which is the whole argument for suspecting a conducted path rather than a radiated one.
Four payloads, one controller revision, fifteen months. The v2 board is now sequencing coil stages on PROJECT #001, which means the scaffolding outlived every single thing it was holding up. Platform goes on the shelf intact.
Stage 1 measures 4.3 Ω against a characteristic impedance of about 1.4 Ω. That is a damping ratio near 1.52 — overdamped, no ring, current peaking late. Four versions of gate timing spent on a waveform that was never oscillating.
Everything into one enclosure, one connector standard, XT60 throughout. Two of the three intermittent faults disappeared on the way in and I cannot tell you which connection fixed them, which is the price of changing six things at once.
Scope permanently on the gate outputs, log viewer up on the second monitor, bank rebuilt at 1000 µF and 450 V. A sixth of the capacitance for sixteen times the energy, and exit velocity barely noticed — which was the first real clue that the bank was never the problem.
The sequencer armed, fired and logged eleven complete runs. Gate 1 transit velocity read two dashes for every one of them. The gates were closer together than the sensor's own rise time, so the measurement was impossible before the first shot.
The kickback clamp works. It works on a solderless breadboard six inches from cables I had personally labelled FIRING CURRENT with three exclamation marks. Both facts are in the same photograph and it took a week to notice they were arguing.
It is not a rail gun. No rails, no armature bridging them — it is a coil and a tube, which is a coil gun operating on reluctance rather than Lorentz force. A month of wrong intuitions came from reasoning about the wrong machine. About 6 J stored, and the slug moved roughly as hard as a strong exhale.
Pool-noodle segments from colour-coded printed magazines, quick-change interface, stamped JK-47. Payload number four and the softest thing this platform has ever pointed at anything. Every version has made the projectile safer and the pointing better, in that order, entirely by accident.
The trigger valve off a pneumatic nailer on the shop wall is faster than the solenoid I specified, sourced, wired and mounted. Some parts of this hobby are solved and the skill is noticing which ones.
Gel blaster on the yoke, pan and tilt on separately labelled drivers, everything on one carryable board. The labels exist because chasing a wiring fault across two identical unlabelled cables happened exactly once and lasted an afternoon.
A commercial gel blaster feeds from a drum, fires repeatably, and had already solved the sealing problem I was three weekends from solving worse. Two months of pneumatics went on a shelf. The beads evaporate, which is correct behaviour for anything that might land in a neighbour's garden.
First run on battery and inverter. A simultaneous pan and tilt move dropped the rail far enough to reset the controller mid-motion — a fault that had existed for a month with no symptom, because the bench supply had amps to spare and absorbed every inrush without comment.
Measured backlash on the pan axis. The microstep is 0.1125°, so the resolution is sixteen times finer than the error — every microstep past the first sixteen is the controller commanding a position the gear train has not agreed to. Raising the gain to close it faster is how you build an oscillator.
Two barrels cut before choosing a projectile. One of them is still on the shelf, perfectly correct for a foam rocket that was never adopted. Bore diameter is downstream of the payload and I did it in the other order.
The valve was never the limiting part. The limiting part was that a charged chamber behind a single momentary switch means the only interlock in the system is attention. It fired twice on a bump rather than a press before that sank in.