Back to the observability problem
Observation
Warrant: Foundation
When reasoning costs nothing, waiting for reality is the expense.
A shutdown happens once. The opportunity to observe it exists once. No quantity of later reasoning reconstructs what nobody was there to see.
01 · The class of non-repeatable windows
Some events emit their evidence once, and then stop.
Industrial work is not uniformly informative. Most of it is routine, repeated on a weekly rhythm, performed under conditions already understood. A narrow class of events is different in kind. Each one assembles people who have never worked together, on equipment that is opened rarely, against a critical path that removes every option to rehearse. For a few days or a few weeks such an event emits more about what a workforce can actually do than a year of ordinary operation emits. Then it closes.
The class is small, and every member of it has the same shape: a window whose schedule is set by physical necessity rather than by anyone's convenience, which opens, emits, and does not reopen. The refinery turnaround is the clearest single case, and it is set out at length in the asymmetry. It is not the only one.
| Event | What is observable only then | What is lost afterward |
|---|---|---|
| Nuclear outage | Work held to the tightest acceptance criteria in industry, under continuous surveillance, on a critical path where every hour carries a cost. | Who held the standard while the clock was running, and who was carried by the people around them. |
| Offshore hook-up | Competence under motion, confinement, weather and shift rotation, on connections that will not be opened again. | Any distinction between the crew that performed and the crew that was merely present. |
| Shipyard erection | How separately fabricated work behaves at the moment it is first brought together. | The attribution of a misfit back to the fabrication that produced it. |
| Commissioning window | The first occasion a system is asked to do the thing it was built to do, with the people who built it still on site. | The link between a build decision and the behaviour that decision produced. |
| Pipeline tie-in | Work performed once, in a trench or a bay that is closed immediately afterwards. | Everything except the finished joint and the paperwork that travels with it. |
| Unplanned repair | Judgement under genuine time pressure, on an incomplete brief, with no opportunity to prepare. | The only unrehearsed sample of how a person works once the plan has already failed. |
| First-of-class fabrication | The point at which a written standard stops describing the work and a person's interpretation of it begins. | Which interpretations held, and which were quietly corrected somewhere downstream. |
Presence at an event cannot be bought after the event.
02 · Why nothing computes its way back
Warrant: Derivation
No model is present at an event that has already happened.
Reasoning is reproducible. Run it again, run a better version of it next year, and the same input yields more, faster, at a lower price. That property is what makes reasoning abundant, and abundance is why it stops deciding outcomes. The argument is made in full in reasoning compounds, evidence does not.
The drawing is checked. The weld is the part that was executed, and it can be observed once.
Observation has the opposite property. An observation is a physical event with a location and a time. It is made by an instrument that was there, or it is not made at all. No operation performed on the record afterwards manufactures a reading nobody took.
This is not a weakness in any particular model, and no better model repairs it. A system asked about an event it holds no observations of does not stop and say so. It interpolates from the events it does hold and returns something fluent. The failure is invisible at the point of use and becomes visible only when a deployment goes wrong, which is the most expensive place in the sequence for it to surface.
An organisation can buy more computation on any afternoon it chooses. It cannot buy a place at a window that has already closed. Price does not move this constraint, because the constraint is not scarcity of a good that somebody else holds. It is the direction of time.
Everything else in industrial work can be acquired late. Presence cannot.
Observation
The only surviving trace of a once-only event is the record somebody chose to make while it was open.
Which places the decision that governs what can ever be known about an event before the event, in the hands of whoever decided what was worth measuring.
03 · The reversal of allocation
Argument
Time in contact with reality outranks time computing.
The argument is short enough to set out in five lines, and each line follows from the one above it.
- 01 Suppose recursive reasoning becomes arbitrarily cheap Take the trend to its limit. Nothing in what follows depends on how quickly it arrives there.
- 02 Then waiting for reality becomes the expense When one input to a decision falls toward nothing, the price of the decision is set entirely by the other one.
- 03 The informative events are non-repeatable by construction The window opens, emits, and closes. It cannot be rescheduled to suit a budget cycle, and it cannot be run twice to reduce the noise.
- 04 Therefore an hour of contact outranks an hour of computation An hour spent where evidence is being emitted is worth more than an hour spent working over evidence already held, because the second hour can be repeated at will and the first cannot.
- 05 Today's allocation is the reverse of this Industrial organisations spend heavily on analysis of what they already hold and almost nothing on being present when something new is emitted. That ratio is inherited from a period when reasoning was the scarce half, and it has not been revisited since the scarcity moved.
The argument needs no forecast. It needs only that the cost of reasoning keeps falling.
04 · What cheap sensing changes
Warrant: Conditional
Cheap sensing industrialises the recording. It does not industrialise the inference.
Cameras, wearables and telemetry are getting cheaper on the same curve as everything else digital. Over the next several years they industrialise the acquisition of raw industrial observation in the way general-purpose processors industrialised the execution of code. That is a tailwind, and it is worth stating plainly rather than defending against.
It is a tailwind for one half of the problem. Recording an event and knowing what the event predicts are different operations. The step that carries the value is the causal link between something observed on a specific day and a real outcome arriving later, under conditions nobody had seen when the observation was taken. That link is not assembled retroactively out of a larger archive. It is established by having been present, having written down a claim while the window was open, and having gone back afterwards to check it.
Cheapening the recording half therefore raises the relative value of the calibration half rather than replacing it. As the abundant input falls in price, the scarce one sets the price of the whole. That holds on one condition, and the condition is not automatic.
- Recording
-
The capture of what happened, in a form that can be stored, indexed and queried. Getting cheaper every year and heading toward free.
Exposed. Necessary plumbing, and not defensible on its own.
- Inference
-
The causal step from an observed characteristic to a later, non-repeatable outcome, under conditions that had not occurred when the observation was made.
Protected. It requires having been present, and it cannot be manufactured after the fact at any price.
- The condition
-
Calibration is grown deliberately, in absolute terms, year on year, while the recording half is automated as hard as it will go.
If investment follows the cheap half because the cheap half returns fastest, the protected half erodes quietly and the tailwind reverses. How much of what is called verification anywhere in this industry is in fact low-level recording is an open question, and it is published as one rather than settled by optimism.
05 · What follows
Presence is engineered, not awaited.
If the informative events are non-repeatable, the question of which events to be present at stops being an operational detail and becomes the central allocation decision the institution makes. Work is chosen against information as well as revenue, by a standing rule rather than by whichever engagement arrives first. That rule is set out in how work is chosen.
Presence is only worth having if something checkable is written down while the window is open. Every assessment leaves here carrying a dated, checkable claim about what will happen next, and the claim is revisited. That is a condition of the work, not a feature of it. A window that was observed and never checked against what followed produced a record, not a measurement, and that difference is the whole of the calibration ledger and the loop that closes on it.
It is also why the edges of the record are published. An institution whose argument rests on being present cannot be vague about where it has not been. What is not yet known is set out in stated limits.