At some point, another search adds less information than a conversation.
There is a computer on the seat in front of me containing a detailed version of a place that has never agreed to any of its assumptions.
For weeks, the project has existed as information: maps, articles, photographs, traffic estimates, cost assumptions, product ideas and operating scenarios. A physical place has gradually been translated into objects, numbers and relationships inside a model. That is useful, but it is also dangerous, because the more detailed a model becomes, the easier it is to forget what it actually is.
An early spreadsheet looks provisional. A mature model has interfaces, dependencies, scenarios, thresholds and calculated outputs. It begins to look less like a hypothesis and more like a description of reality.
But reality has not signed it.
One of the simplest rules in the model is therefore also one of the most important:
Unknown is not zero.
An unknown rent is not free rent. An unknown staffing cost is not free labour. An unknown repair bill is not a repaired building, and an unknown commercial structure is not permission to insert whichever structure makes the numbers work. Sometimes leaving something unresolved is more accurate than filling the cell.
That sounds obvious until software enters the process. Software prefers completion. Dashboards want totals, forecasts want outputs, and models reward us every time uncertainty becomes a number. AI makes that temptation even cheaper: another scenario can be generated in seconds, another comparison can be found, another argument written, another plausible explanation added until the internal world becomes slightly more coherent.
The problem is that a model can continue improving long after the most valuable next piece of information has stopped being available online.
That is where I am now. There are questions that another hour of searching will not answer. Someone knows why certain decisions were made. Someone knows which plans were serious and which were merely possibilities. Someone knows what the model currently represents as an empty field. And someone may know that an assumption I have spent days refining is simply wrong.
That would be useful.
A five-minute answer should be allowed to invalidate five days of modelling.
If it cannot, the model has stopped being a research tool and started becoming something that needs to be defended. The work was not wasted if it identified the question capable of destroying it.
There is a natural escalation in research. Start with the cheap evidence: read the documents, search the archives, look at the maps, compare similar cases, calculate, model, and make the assumptions explicit. But eventually the marginal value of another search collapses. The next research instrument may be a measurement, a contract, a site visit -- or simply a conversation.
Today, the computer is coming with me for that reason. Not because it contains the answer, but because it contains a structured record of what I do not know.
If the next conversation is useful, some empty fields will gain evidence, some assumptions will disappear, and new unknowns will probably appear. The model may even become less complete in places.
That can be progress.
The spreadsheet has done enough talking. Now it gets to listen.
