The cabinet is a 14 cubic foot Danby wine fridge with its compressor removed and a Peltier stack in its place. The job is to keep printer filament dry. For five sessions the spec said keep the cabinet cold, and every thermal decision downstream of that phrase inherited it — radiator sizing, duty cycling, the setpoint, the sensor placement. It was wrong, and it was wrong in the quiet way that survives review: nobody questions a premise that isn’t being argued.
Cold is not the goal. Cold is a byproduct of how a thermoelectric dehumidifier removes water — you chill a surface below the dew point, water condenses on it, you drain it. The chilling is the mechanism. The dryness is the product. Confusing the two meant I was tuning the machine to maximize its side effect.
Worse, the side effect is actively harmful. Filament pulled from a cold box condenses on the way to the printer, and a spool’s thermal mass makes it the worst case in the room — it stays cold long after it leaves the cabinet, collecting exactly the moisture the cabinet just spent four hours removing.
What the envelope costs
The correction forced a real estimate for the first time. The box is roughly 396 L with about 5.3 m² of surface, a square meter of which is tinted glass — call it ~5 W/°C. That parks the interior five to six degrees below ambient on about thirty watts of pumping, which is comfortably inside what four tiles can deliver and, more importantly, is no longer a number anyone needs to hit.
The setpoint was the other casualty. The firmware carried a 10% relative-humidity off-point, inherited from a spec sheet and never sanity-checked. At 20°C, 10% RH is a dew point near −12°C. A Peltier stack in a wine fridge is not reaching negative twelve on the cold sink, which means the setpoint is unsatisfiable, which means the tiles never shut off.
That is a firmware problem wearing a hardware costume. Nothing needs to be rebuilt. But it does mean the only thing standing between this build and continuous 100% duty operation is a duty cap I guessed at — 45 on, 15 off — with no measurement behind it. A guessed cap holding back an unsatisfiable setpoint is not a safety margin. It is a coin flip with a schedule.
The intuition runs backward
Doubling from two tiles to four raised an obvious question I got wrong on instinct: why not one bigger tile instead? Larger module, deeper cold, fewer parts.
Maximum ΔT is a material property of bismuth telluride — roughly 65–70 °C at zero thermal load — not a size property. A bigger module doesn’t reach deeper. It pumps more heat into a radiator whose rejection capacity is fixed, which raises the hot side, which by definition warms the cold side. Scaling the tile scales the problem.
Kill table
what this session falsified, and what killed it
| Claim | Verdict | Why |
|---|---|---|
| Keep the cabinet cold. | Wrong objective | The goal is dry. Cold is a byproduct of the mechanism, and filament leaving a cold box condenses en route to the printer. Struck throughout the spec. |
| One larger tile beats four small ones. | Physics | ΔT is a material property, not a size property. A larger module raises the hot side against a fixed radiator and therefore warms the cold sink. Rejected before purchase. |
| Heat the whole cabinet for a dryer mode. | Killed, not deferred | Replaced by a partitioned dryer chamber. An independent hardware over-temperature cutoff is non-negotiable — a resistive heater in a printed box running unattended is the one configuration here that can burn the house down, and a software limit is not a thermal fuse. |
| Distribution blocks are on hand and specified. | Never existed | The third part this spec invented. The row carried the weakest provenance tag in the whole reconciliation and was still being propagated in prose as owned inventory. |
| Separate blocks, so no stray strand can bridge to ground. | Withdrawn | Written for busbar terminals with exposed screws. A lever connector is a spring cage with 15 mm insulated bores. The objection was correct for the hardware it was written about and got carried to hardware it doesn’t describe. |
Two of those were my own prior claims. One was a part that existed only in a document. The reasoning error worth keeping is the last one — a true objection outliving the context that made it true.
What actually transfers
Three things from this session are worth more than the cabinet:
- A product’s name is not its topology. A connector sold as “2 in, 8 out” reads as ten electrically common holes. It is two isolated 1→4 banks. Meter what’s common; don’t parse the part number.
- Read the rating off the part, not the listing. The fuse listing gave amperage, speed, size and material — no voltage. Ceramic slow-blows ship in both 250 V mains and 32 V automotive and look identical in the hand.
- Cycling a dehumidifier re-evaporates its own condensate. Every time the cold sink warms, water sitting on it goes back into the air. Continuous low-power beats intermittent full-power — an argument the duty cap has never had to answer.
The cabinet still isn’t finished. But it is finally being built toward the right objective, which is the only kind of progress that survives the next session.
