A measurement the shop floor can explain is a measurement the shop floor uses. That is what training builds, and it is what remains once we no longer come back.
Three questions come up before an inspection rather than during it: what was this model built on, how do we know it is still valid, and who will be able to say so once the person who built it has left.
A project has succeeded when one simple sign shows: the team no longer needs us.
Training to use and training to understand are two different trades
Training to use
Starting an acquisition, presenting the material correctly, reading a screen, responding to an alarm, completing a record. It is what a system needs to run day to day.
This training answers “how do we do it”. It prepares for no unusual situation.
Training to understand
Knowing why the displayed value is what it is. What moves it, what distorts it without moving it, and what the sensor actually looks at.
It is the only one that lets someone diagnose a drift, decide whether to recalibrate, and defend a method in front of an auditor.
The difference never shows in routine. It shows on the day of a deviation: an atypical batch, a raw material from a new site, an inspector’s question. “Black box” is not an acceptable description of a model in a regulated environment. What you cannot explain, you cannot defend.
What ICH Q9(R1) says about it
ICH Q9(R1) says it for quality risk management, in section 6: “Training of both industry and regulatory personnel in quality risk management processes provides for greater understanding of decision-making processes and builds confidence in quality risk management outcomes.” Training personnel gives a better understanding of decision-making processes and builds confidence in their outcomes. That is exactly what “training to understand” produces.
Five audiences, and each has its own questions
| Audience | What they must be able to do alone | Frequent mistake |
|---|---|---|
| Line operators | Tell a valid acquisition from a doubtful one, and present the material the same way from one shift to the next | Teaching them chemometrics, instead of the gesture and the right to challenge a value |
| Laboratory technicians | Build a usable set of reference samples, and see that the uncertainty of their method bounds that of the model | Treating them as suppliers of values, without discussing the specificity of the reference method |
| Process engineers | Link a signal to a control parameter, read the diagnostic statistics, prepare a recalibration | Handing them the model with no access to the pre-processing, the calibration set or the residuals |
| Quality assurance and validation | Know what a declaration of compliance covers, what ICH Q2(R2) requires, and what belongs in the control strategy | Discovering the subject when the file is being written, two years after the instrument was chosen |
| Management | Know what the system monitors, what it does not, and what happens when a key person leaves. ICH Q10 places human resources and training under the responsibility of management | Funding the instrument and not its upkeep |
The topics we teach
- The principles of non-destructive measurement. A spectroscopic measurement does not analyse a batch: it analyses a few milligrams of material. Repeating the measurement at the same spot only reduces the noise of the instrument.
- Sampling and representativeness. The most profitable module, and the least requested. The literature of the theory of sampling — the work of Pierre Gy, spread by Kim Esbensen — commonly reports a ratio of 10 to 50 between sampling error and analytical error on heterogeneous materials. It is an order of magnitude, not a constant to apply as such to a given process, and the module teaches precisely how to establish it on yours. Biased measurement positions against compliant ones. The sampling pillar.
- Applied chemometrics and its two calibration routes. The pure-component model decomposes the spectrum on pure components: it requires a known, declared composition, and it discovers nothing. The model calibrated on samples regresses on reference values: it requires a representative population, and it ages. Both are chemometrics: the choice between them is a process decision.
- Reading the diagnostic statistics. Hotelling’s T², residuals, detection of out-of-range samples, the gap between calibration error and validation error. A model whose performance collapses in external validation has learnt noise. Parsimony is what passes audits.
- Running a recalibration. When to trigger it, on which samples, what gets documented, under which change control. The rule fits in one sentence: recalibration is the last resort, never the reflex. Keeping a model alive.
- Critical reading of a supplier offer. Covered below: it is the module that most changes a manufacturer’s relationship with its market.
Reading a supplier file is a skill in itself
A manufacturer that can read a technical offer negotiates differently and chooses better. We teach this module even when it works against a sale.
- On a declaration of compliance. Which version, which scope. Is the electronic signature specified in the functional requirements, or only claimed. Is there a method lock. Above all: “21 CFR Part 11 compliant” is not “validated system”: manufacturers themselves state that overall compliance depends on the operator’s procedural controls. What a declaration covers.
- On a spectral library. How many entries, and above all which families. What happens when a material is not in it: rejection, non-identification, or identification by default. Who has the right to extend it.
- On a rejection threshold. How was it set, on how many samples, with what accepted rate of false negatives. A threshold inherited from a demonstration set does not protect a process.
- On a qualification protocol. Does it cover the instrument at rest or the measurement in flow. The first describes the device, the second the product in motion: the two complement each other.
One last reflex to install: two instruments with the same generic name are not interchangeable. A spectral range that stops before the bands of the minor constituent you want to quantify will give a mediocre model, whatever the other specifications. Two instruments, two results.
Who owns the model when the project ends
- Who owns the model. A model is a calibration set, a chain of pre-processing steps, coefficients, and the record of the choices made. Delivered without its data, it is usable, not transferable.
- Who has the right to modify it. Adding samples, widening a range of validity, changing a pre-processing step: under which change control, with which performance re-verification. A team without that right is not autonomous, however competent.
- What happens when the project lead leaves. The critical knowledge lies in the reasons behind the decisions: why this pre-processing, why this constituent was set aside. We document the reasons, and train at least two people per critical audience. Two people who know, and the knowledge stays in the team when one of them is away. That is knowledge management in the sense of ICH Q10 (§ 1.6.1), linked to change management by ICH Q12. The text describes “a systematic approach to acquiring, analysing, storing and disseminating information” and imposes no particular tool.
These questions are prepared during deployment, not at the end.
We train so as to be no longer needed. An installation that only holds because we come back to restart it is a dependency, not an asset. An autonomous team keeps it alive and extends it.
We would rather come back to extend the approach to a new process, at the request of a team that knows what it wants. Integration delivers a system. Training delivers the ability to keep it alive.
Support that sells availability, not hours
The principle fits in one sentence: the grid is strictly additive. Each level takes over the previous one in full and adds one line. No à la carte options, no bank of hours to use up.
| What is added to the previous level | What it changes |
|---|---|
| Remote support | Someone who knows your installation, reachable on the day the value becomes doubtful |
| Training | Bringing teams up to date, onboarding newcomers, recovering after a reorganisation |
| Software updates | Version traceability, and a check that the update has not changed the result returned |
| Periodic review with performance validation | Review of the residuals, comparison with the reference method, a documented decision: keep, watch, recalibrate |
| On-site intervention | A visit when remote diagnosis is no longer enough: fouling, mounting, sample presentation |
| Instrument warranty | A hardware failure stops being a hazard arbitrated in a hurry |
| Replacement instrument during repair | The measurement does not stop. The only line that protects a process driven by the measurement itself |
The higher levels do not sell working time: they sell availability. A volume of hours gets used up and leaves a gap at the end of the period. Availability cover renews without interruption.
What a support contract leaves with you
It covers the instrument, the software and assistance. It does not decide for you that a model has drifted. Reading the control charts and triggering a recalibration remain acts of your team. That is why they are taught.
The awareness workshop, often the real way in
Most projects do not start from a precise request, but from an intuition: “we should be able to measure this better than we do”.
The format is half a day on site, with the people who know the processes, production, laboratory, quality, in the same room. We map the steps, and note for each what is measured and when the result arrives. We then pick out the measurement questions that are genuinely open, and sort them: measurable in routine, worth watching, or to be set aside.
The output is a short, ranked list, with at least one subject explicitly dropped, often the most useful line of the day. The natural next step is a scoping phase on one or two subjects.
Frequently asked questions
Doesn’t training your clients to understand everything cost you work?
On paper, yes. In practice, an autonomous team extends the approach to other processes and buys better-defined projects. That is what brings projects back, and it is in our interest as much as yours.
Will our operators accept it?
The objections are always the same: “the machine will decide for me”, “we will be replaced by sensors”, “what if the instrument is wrong”, “it is one more task”. What works fits in four gestures: show the measurement running on their own product, involve the operators in choosing the probe position, show that the model can flag what it does not know, and count at scoping the samples the in-line measurement replaces. The detail, sentence by sentence.
Is training enough to defend the method in front of an auditor?
No. A competent team with no documentation does no better than complete documentation with no one to explain it. It takes both.
After training — keeping a model alive over time
Tell us who will have to keep the measurement alive after us. We will build the training around those people.
Forty-five minutes are enough to identify the audiences to train on your side, what each must be able to do alone, and the level of support that matches your criticality.