What counts as a
measurement, and
what is a guess.
A number is a measurement only if whoever produced it could see the thing being measured. Our plant diagnostic asks who is answering before it asks anything else, and when a question about the shop floor is put to someone who does not work the shop floor, it offers them a way to decline and then declines to compute a score rather than filling the gap.
The failure this exists to prevent
Ask a finance lead how many levels sit between an operator and the plant head and you will get a number. Ask a shift manager the same question and you will also get a number. They arrive in the same field, of the same type, and every system downstream treats them identically.
That is the whole problem. A guess and an observation are not distinguishable once they are both integers. Whatever the diagnostic writes to a record, whatever feeds a score, whatever gets quoted back in a meeting six weeks later, carries no trace of which one it was. Nothing breaks and nothing warns.
Who is answering is asked first
The role question used to be twelfth. It is now first, and the reason is mechanical rather than conversational: asked last, it arrives after the floor questions have already been answered and scored. Asked first, it decides how those questions are put and how the result is qualified.
Roles that work with the systems in question are treated as firsthand: an owner or managing
director who is on the floor most days, operations and production, IT and systems, quality,
maintenance. An owner or managing director who is mostly away from the floor is treated as
secondhand, and so is anyone in finance or general management. Anyone selecting something else
is unstated.
Four questions are about the floor, and three of them carry most of the score
The index is 100 points across four components. Four of the twelve questions ask what actually happens on the floor rather than what policy says, and this is how the points fall:
| Floor question | Component it scores | Points |
|---|---|---|
| How many machines report their own data without a person entering it? | Machine data entered by hand | 20 |
| Where do machine and shift logs actually live? | Where logs are kept | 20 |
| Hours per week your team spends moving data between systems by hand? | Manual data movement | 25 |
| How many levels sit between an operator and the plant head? | Not scored. Context only. | 0 |
| Of 100 index points, carried by floor questions | 65 | |
The remaining 35 points come from which systems are running and which of them exchange data on their own, which anyone in the business can answer. The hours question additionally produces the only figure in the report expressed in time rather than points.
A respondent who cannot see it can say so
When a floor question is put to a secondhand or unstated respondent, the question says so on its face, and an extra option appears beside the others: I would be guessing. It is not a hidden escape. It is a first-class answer, offered by the tool, and choosing it is treated as more informative than answering.
The option is shown only to people it applies to. Someone who works the floor is not offered a way out of a question they can answer.
What the score does then
- The index is not computed. It is null rather than partial, because a number that omits 25 of its 100 points and does not say so is worse than no number.
- The report says which components are missing, by name, and how much of the 100 points is actually held.
- The result is flagged as not a measurement. A score counts as one only when every component is present and the respondent was firsthand. Every consumer of that record is expected to read the flag before treating the number as a fact.
- The hours figure disappears entirely when the hours question is declined, rather than being estimated on the respondent's behalf.
Three states, and the third one fails closed
There are three answer states rather than two. Firsthand, secondhand, and unstated. The third exists because a respondent we cannot classify is not the same as a respondent we can vouch for, and it is treated exactly as secondhand for every refusal above.
That is a deliberate choice about which way to be wrong. Treating an unknown respondent as reliable produces a confident number nobody can audit. Treating them as unreliable produces a smaller answer that is true.
Why a partial picture beats a complete guess
A partial result can be finished. It names what is missing, so the next step is obvious: ask the person who would know. It can also be acted on immediately, because what it does contain was observed.
A complete result containing guesses cannot be repaired, because nothing marks which parts need repairing. It will be copied, quoted and compared against later results as though all of it were measured. The error does not stay where it was made.
The weights above are a considered judgement about what makes plant information hard to read across. They are not a calibrated model. Nothing here has been validated against outcomes, and no claim is made that a score of 70 predicts anything in particular.
The index has been computed three times in total, and all three were our own runs while building and checking the tool. No outside respondent has completed it yet.
The answer-basis design described on this page went live on 17 August 2026, which is after all three of those runs. So the decline option has never yet been used by anyone, and no recorded score currently carries a basis at all.
We would rather say that than imply a validation that does not exist. The same standard is the point of the whole page: a stated limitation is what makes the rest of a number worth reading.
Where this came from
It came from finding the defect in our own tool. The role question sat twelfth, four floor questions were being answered by whoever happened to open the page, and the resulting index was written to a record and fed into scoring with nothing attached to say who had produced it. The tool did not fail. It produced a well-formed number of exactly the right type, which is the hardest kind of error to notice from outside.
Run it on your own plant
Twelve questions about how your systems, machines and registers actually hold information. No signup. You get a fragmentation score, an estimate of the hours it costs every week, and the first fixes worth making.
Your answers and your score are recorded when the report renders on your screen, which happens before any email is asked for. The page says so before the first question.
Take the diagnostic