A passing pilot build is what sets up a failed first lot
A passing pilot build is not evidence that the line can build the pack. It is evidence that a small group of engineers can, with their hands inside the process. What actually protects the first production lot is a run at normal takt, staffed by production operators, with engineering off the floor, and with five things already frozen and written down before that run starts: the weld schedule, the fixture design, the torque specification, the BMS firmware revision, and the end-of-line limits. Then the first lot has to be read as a trend, unit one against unit last, because the drift that ruins a ramp is invisible to a first-article inspection.
The last unit of a pilot build I sat through came off around nine at night. The line was quiet enough to hear the contactor in the end-of-line tester open and close. There were four people on the floor and three of them were engineers. The process engineer was crouched at the weld station with the fixture open, and the operator's hands were flat on the bench, waiting. That is the detail I kept. On a line that is actually running, nobody's hands sit flat on a bench while a machine is indexed.
That unit passed. So did everything else that week. The pilot report said the build was clean, and it was, and it meant less than everyone in the room wanted it to mean.
What the pilot is actually measuring
Stand where I was standing, maybe four feet back from the weld head, and watch what happens between units.
A unit comes out of the welder and the process engineer picks it up before the operator does. He looks at the nuggets. He sets it down, taps the screen, and changes one number in the schedule — energy, up a step. Nobody writes it down, because it is a small change and he will remember it. Two units later a busbar does not seat cleanly, so he finds a burr on the fixture locating pin, and he deburs it with a file. The operator re-seats the connector on the next pack with a little wiggle that takes an extra second and a half, and it works every time.
None of that is cheating. That is what a pilot is for. You learn faster with an engineer at the station than you ever will from a report, and the whole point of the exercise is to find out what the design and the process do when they meet real parts.
But every one of those moments is a compensation, and compensations are precisely what disappear the week the line runs at takt.
There is also a physical mechanism underneath the supervision, and it is the one people underestimate. At pilot pace the line might index every ninety seconds when it will eventually index every thirty-five. That extra time is not idle. It is cooling time. The weld horn sheds heat between welds. The aluminum fixture comes back toward room temperature instead of heat-soaking through a shift. The operator has three spare seconds to seat a busbar squarely. Run the same recipe at production pace and the horn soaks, the fixture grows a few thousandths, the spare seconds are gone, and the same setpoints now deliver a different weld. That is variability created by pace, not by skill, and no amount of pilot data collected at pilot pace will show it to you.
So the question at pilot exit is not "did it pass." The question is "what was holding it up, and will that thing still be standing there in three weeks." Usually the answer is: an engineer, and no.
Run at rate means something specific
Run at rate is a build at the takt the line is designed to hit, with production staffing, over enough consecutive units that the line has to survive its own housekeeping — a shift change, a break relief, a horn dress or tip change, a reel of tape running out, a fixture wipe-down. It is not a demonstration that the line can produce good parts when everyone is watching. It is a demonstration that the line produces good parts when nobody in particular is watching.
What it proves that the pilot cannot: that the takt is achievable by the actual crew, that the consumable change interval does not land in the middle of a critical step, that the tooling holds its geometry across a thermal cycle, and that the settings on the machines are the settings that make good parts.
What it does not prove, and I want to be honest about this, is that the design is right. A run at rate on a bad design will happily produce bad packs at the correct rate. It only tells you whether the process is capable of repeating itself without a babysitter.
Keep a count of every time someone touches a setting
This is the cheapest instrument in the whole exercise. Put a clipboard or a shared log at each station and write down every intervention — who, when, what, why. Do not ask people to judge whether it was significant. Just count them.
At the end of the run, the number of interventions is your pilot's real yield figure. A run with zero setting changes across a full shift is a very different animal from a run with eleven, even if both produced nothing but passing units.
The five things that have to be frozen
Frozen does not mean the number is written in a document somewhere. Frozen means the value is on the machine, locked at a level the operator cannot reach, and recorded as the value that produced the retained sample units.
Weld schedule
Not the recipe name — the actual energy, time, and force, or the amplitude and collapse distance if it is ultrasonic, plus the condition of the horn and anvil at the moment of the freeze. A schedule tuned against a horn with two hundred welds on it is not the same schedule on a fresh horn. Record the dress count and the planned change interval, or the schedule is only valid until the first tip change.
Fixture design
The drawing revision and the physical fixture that produced the pilot units, identified by its tag. If the pilot fixture has tape on a locating surface, or a shim under a clamp, that is a variable, not a detail. The production fixture has to be built to the same revision, and someone has to compare locating surfaces and clamp travel side by side rather than comparing drawings.
Torque specification
The spec plus the tool that delivers it: driver model, program number, clutch setting, and the joint stack-up it was validated against. A pilot bench driver and a production fixtured driver are different instruments, and torque values do not transfer by copying the number.
BMS firmware revision
The exact build, identified by revision and hash, plus the flashing procedure and the fixture used to flash it. Not "latest." Not "the one the firmware engineer had on his laptop." The pilot units were probably flashed with something, and whatever that something was is now the reference against which every production pack gets compared.
End-of-line limits
The numeric limits and the measurement system that produced them. This is where pilots quietly cheat, because wide limits let the pilot ship and nobody argues with a passing unit. Freeze the test fixture, the probe arrangement, the stabilization time, and the limits themselves, and freeze the gauge study behind them — a limit set on a tester with poor repeatability is not a limit, it is a coin flip with a number attached.
Unit one against unit last
A first-article inspection is a photograph. What you need is a film.
Take the first unit off the first production lot and hold it. Take the last unit of that same lot and hold it next to it. Then measure both the same way, with the same gauge, if you can manage it with the same person measuring. Dimensions, torque audit on a sample of fasteners, weld cross-section on destructed units from both ends, end-of-line results compared as distributions rather than pass/fail, and the per-station cycle time from the line's own data, not from a stopwatch someone ran during the pilot.
What you are looking for is not a failure. You are looking for a slope. A dimension that walks three thousandths from the first unit to the fortieth is not a defect, and it will be a defect by the four hundredth. Torque that starts at the nominal and ends at the bottom of the window is a driver that is drifting or a joint that is settling, and either one is a conversation you want to have in week one rather than at ramp.
Keep the first and last pilot units sealed and labeled with the frozen settings register. Those two packs are the only honest baseline the production line will ever have, and they are worth more than any report written about them.
The handover document that actually matters
The pilot team hands over a settings register, and most teams do that part. The section that gets left out is the list of compensations — every place the pilot team did something the production line will not do. The deburred pin. The extra second and a half on the connector. The second EOL run after re-seating a probe. The hand-tightened screw on the fixture that was never torqued because it kept backing out.
Write those down as a list of known soft spots, and then for each one, state what the production line is supposed to do about it and who verifies that they did. Not "monitor." Verify. If the answer is "the operator will notice," the answer is wrong.
The rest of the handover is mechanical. The frozen settings with their revision numbers and physical locations. The retained sample units. The raw end-of-line data from the pilot, not the summary. The intervention log. Photographs of the weld horn and fixture at the freeze point, taken that day. The name of the one person who can authorize a change to any of it, and the rule that a change means a recorded re-qualification rather than a quiet tap on a screen at two in the morning.
Exit criteria belong on the line, not in a meeting
A sign-off meeting will confirm anything you put in front of it, because everyone in the room wants the program to move. That is not a character flaw, it is what meetings are for.
So write the exit criteria so they can only be satisfied by a physical event on the line. Instead of "the team agrees the pilot demonstrated capability," write something like: one full shift at production takt, production crew, engineering present only as an escalation contact, zero setting changes logged, and first-and-last-unit measurements within the agreed window. That is a criterion a person can stand on the floor and watch happen. It cannot be talked into existence.
One caveat, because it is real: this does not fit every program. On very low volume, or on a product that is deliberately hand-built by skilled technicians, heavy supervision is the process and freezing it is the correct answer — you freeze the technician's method, not a machine's setpoints. And on a program where the design is still moving, freezing five variables in week six is premature and will just generate re-qualification work you have to redo. The freeze belongs at the point where the design is stable and the process is about to be handed to people who were not in the room.
What you can never do is freeze nothing and call the pilot a success. That is how a clean pilot build turns into a first lot that fails in month two, with everyone genuinely surprised.
