The Lab Next to the Line

You cannot run FDE on the line if the only place a process engineer is allowed to try is a glass room off the factory tour. Hardware organisations already have "innovation labs." They are usually a walk-through, a cobot behind a rope, and a quarterly showcase that never touches a qualification lot. A lab that empowers sits next to the line: access to the allowed slice of OT, a harness that can graduate into the cell, a witness who unblocks safety and quality in hours.

The software version of this argument is hunt change-makers and give them tools now. Same OS. Heavier floor. Idea-to-execution is short in CAD and firmware too. If you do not give them a sanctioned bench, they will paste the drawing into a consumer chat. Talent on the plant has never been this valuable. Acting like you have time for a 2027 enablement plan is how you donate it.

Hunt. The town hall will not volunteer them.

Change-makers in hardware are busy and slightly allergic to programmes. Look at who already:

• Keeps the Python notebook that is the real MES for one cell.
• Automated iLogic / NX / Vault and was told to use the official process.
• Diffs twin against live telemetry without a ticket.
• Owns the unofficial pinout that is more accurate than PLM.
• Is the person three teams already ask before they ask engineering.

Anger with a spec is a gift. Spare time is not a proxy for taste. Open-door "anyone can 3D-print" fills the lab with people who have calendar, not picture.

Access, harness, witness: with OT constraints

Three jobs. Only three.

• Access, identity, the model route that is allowed on this site (including the air-gapped one), a data product that is classified, a sandbox that looks like the cell. Six weeks for an account means you have a brochure.
• Harness , skill folders, evals, logging, a ladder that can fail metal. Not a chatbot. A path that can graduate.
• Witness, safety, quality, OT security on hour-scale latency. If they only attend steering, the lab is a waiting room. Empowerment is a latency number. In a plant it is also a liability number.

Starter kit this month, not a roadmap to a kit: managed agentic IDE, role bundle of skills, classified slice, budget for tokens and bench time, a stop they will actually use. If they have to steal nights and USB sticks, you have told them this is a hobby. Hardware talent does not stay in hobbies.

The lab is not the line

Prototype in the lab. Graduate into the residency or the process owner's system of record. If everything lives forever in the lab, the lab is a hospice, or worse, a parallel factory that nobody qualified. Labels, in writing, by a date:

• Graduate, running on the cell, owner named, evals on, FDE or platform holding the harness.
• Library, the UI dies, the skill / connector / eval stays. Reuse is the point.
• Sunset, wrong problem, wrong owner, unsafe. Delete it where people can see. Scrap is inventory that should hurt.

Ninety days is plenty to know. The fractional-employee test still applies: does it keep working when the original human goes on leave?

Staffing a foundry, not a campus

One lab lead measured on loops graduated, not on cobot count. One harness engineer (often the same pod that makes FDE residencies boring). Four to eight operators-in-residence with a named outcome in their real job. They do not "go to the lab." They bring the cell in. A safety/quality partner on a short leash.

Do not staff a research bench unless you are actually doing research. Most sites are trying to install known patterns into stubborn processes. That is foundry work. The micro-team is the unit.

How you know it is real

• Time to first production-shaped path, not a twin fly-through.
• Change-maker conversion, they shipped a loop someone else can run.
• Pull from the line, not push from the tour.
• Reuse of skills across cells.
• Shadow CAD-in-ChatGPT down as the sanctioned path gets faster. If shadow rises, the lab is slower than the street.
• FDE residencies that start from the library.

Token spend is not a KPI. Hours that do not graduate are inventory.

Failure modes

Factory tour theatre. Glass, waiting list, no path to a lot.

Tools as a future tense. Meanwhile the drawing is in a personal account.

Open door, no hunt. Average inputs, average loops, real risk.

Research cosplay. Model bake-offs, zero toolpath change.

No stop, then a ban. Someone ships something unsafe. OT panics. Lab frozen. Change-makers go underground. You had a chance to put governance in the path.

This quarter

Name the lead. Name the safety partner with hour-scale latency. One-page non-negotiables. Allowed data slice. Model route that can actually run here. Hunt four people by watching the work. Day thirty: ugly production-shaped path, logging on, stop tested. Day ninety: two loops labeled. Then talk about scale.

The lab next to the line is how hardware talent gets a floor: tools this month, permission with a stop, skills that can graduate, a hunt for the people already pulling. Campus labs decorate the tour. Foundries change yield.

Related: Establish the Lab, Charter the Lab, and Forward Deployed on the Line.