The lights go off. Everything the night shift knew has to be written down first

A dark factory is not a robotics problem — most plants already have the machines. It is a context problem. The night shift carries an enormous amount of undocumented knowledge about which machine drifts, which material behaves oddly and when to stop. Unattended operation means that knowledge has to exist somewhere a system can read.

  • A machine-readable model of the operation. Equipment, materials, routes, constraints and tolerances in a form a system can reason over rather than in a manual.
  • Exception handling defined in advance. What happens when something is out of tolerance, at three in the morning, with nobody there — decided while people are available to decide it.
  • Provenance for every decision. What was observed, what was concluded and what was done, so the morning shift can reconstruct the night.
  • An honest map of the boundary. What can genuinely run unattended and what cannot, which is usually a narrower set than the ambition.
Talk to us about this

What is delivered

  • Context model

    Equipment, capabilities, materials, routings, tolerances and dependencies. The thing an unattended system reasons over, and the artefact this whole service exists to produce.

  • Tacit knowledge capture

    Structured sessions with the shift teams to write down what they know: which machine drifts after a changeover, which supplier lot behaves differently, what a bad sound means. The hardest and most valuable part.

  • Exception taxonomy

    Every deviation the operation actually sees, with a defined response: continue, hold, stop, escalate. Anything not on the list stops and calls somebody, by design.

  • Decision record

    A permanent log of what happened overnight and why, readable by a person in the morning. Without it, unattended operation cannot be audited and therefore cannot be trusted.

  • Read-only integration

    Signals from equipment, MES and quality systems. The context layer reads; it does not write to machines, and anything that actuates stays inside the safety case where it belongs.

  • Unattended-readiness assessment

    Which parts of the operation are genuinely ready, which need instrumentation, and which should not run unattended at all. Frequently the most useful document.

How it runs

  1. 01

    Watch a night shift

    Literally. What people actually do between midnight and six is the specification, and it is almost never what the process documentation says.

  2. 02

    Capture and model

    The tacit knowledge into the context model, with the shift teams reviewing it. If they do not recognise it, it is wrong.

  3. 03

    Run supervised

    The system decides, a person confirms, for as long as it takes to compare its calls with theirs. The disagreements are the remaining work.

  4. 04

    Narrow the supervision

    Reduce human confirmation where the record supports it, one exception class at a time, with a defined route back.

A good fit when

  • You already run automated production and want to extend unattended hours.
  • Knowledge sits with a few experienced people and their retirement is a risk on somebody's register.
  • Unattended runs have been tried and stopped too often to be economic.
  • You need to demonstrate to an auditor or a customer how overnight decisions were made.

Not the right service when

  • The plant is not instrumented. Sensing and data engineering come first — context over missing data is fiction.
  • You want the system to control machines. We do not build inside the control path, and anyone who does should be able to show you the safety case.
  • The real constraint is manual material handling. Automate that first; context will not move a pallet.

Frequently asked questions

Is this robotics or automation engineering?
Neither. It is the layer above them: the model and the rules that let a system decide what to do next. The machines and their controllers stay with the suppliers who assessed them.
Does the system control the line?
No. It reads, decides and records, and where an action is needed it goes through the existing control systems under their own safety case. Anything that can actuate machinery carries obligations a context layer should not claim.
How do you get knowledge out of experienced operators?
Structured sessions built around real cases rather than interviews, and reviewing draft rules with them until they recognise their own judgement. People cannot list what they know, but they can correct a wrong version of it — which is the technique.
What if the exception taxonomy is incomplete?
It will be, and the design assumes it. Anything unrecognised stops and calls somebody. An unattended system that guesses at an unknown exception is worse than one that halts, and the taxonomy grows from each of those calls.
How long does this take?
The assessment is typically three to four weeks. A context model for one line, with tacit knowledge captured and an exception taxonomy, is usually two to four months — most of it spent with the people who work nights.

Keep reading

Other services