Four ways a model is confidently wrong about physical things
For about a year I have been rebuilding an old off roader, researching every part the way I research everything else: written down, in documents, with revisions, reasoning kept next to conclusion.
It works. It is faster and far more thorough than the forum and phone call method it replaced. It also fails in four specific ways that are much easier to see here than in software, because a wrong answer arrives as a box containing a part that does not fit.
Here is the catalogue, and what catches each one.
One: a number that was an estimate becomes a number that is a fact
A winch sat in my notes at four thousand seven hundred lei. That figure came from a note about a different model. It was marked as an estimate, and then it was summed into a budget total as though it were a quote. The real price was seven thousand.
The catch: estimates and confirmed prices must not live in the same column. Addition destroys the distinction instantly and silently. I keep two columns now, and the total only ever adds one of them.
Two: a specification deduced from a general rule
The engine oil specification was declared wrong by reasoning from a principle: this car has a particulate filter, therefore it needs the oil that particulate filters need. Clean reasoning, confidently delivered, and wrong for this engine.
The catch: a general rule applied to a specific machine is a hypothesis, and it reads exactly like a fact. So the question I now ask every time is "is that from the manual for this engine, or is that a category rule". The phrasing of the answer tells you immediately.
Three: a causal chain stated backwards, with a part attached
One document had soot deposits causing a flap problem. The next reversed it. On the strength of the first version, a part costing nearly two thousand lei had been proposed as the first thing to buy.
The catch: before buying anything that follows from a diagnosis, ask what measurement would distinguish the two directions. In this case the readings had been taken with the engine off, which means nothing. The same car read warm and moving reversed the diagnosis within a day.
Four: an answer that does not exist, because only half the market was searched
Twice I was told a particular steel wheel in that bolt pattern does not exist. It does. It is an ordinary factory spare for about a hundred and eighty dollars. The search had been through aftermarket off road catalogues and never once through the manufacturer's own parts list.
The catch: before accepting that something does not exist, ask which sources were searched. The aftermarket is loud and well indexed. Factory parts catalogues are quiet and barely indexed, and a model trained on the open web has read far more of the first than the second.
The whole wheel saga ended with a used set of original wheels, which solved from the start everything that had held us up for days, because the bore is correct by definition, the load rating is the one the car was designed around, and clearance over the brakes is not a question anybody needs to ask.
The two rules underneath all four
A forum heuristic is not a specification. Get the measurement first. The brake question flipped twice in one evening between two confident internet heuristics and was settled permanently by a tape measure in four minutes. Three hundred and fifty millimetres at the front, three hundred and thirty at the back, and the question was closed forever.
A written conclusion is only reopened by new evidence, not by new pressure. I lost two days re-litigating a decision my own research file had settled six weeks earlier, with the reasoning attached. The file was right. One ambiguous forum comment reopened a question that six consistent sources had closed, because doubt expressed confidently feels like information.
That second one is the expensive one, and it is not a tooling problem. It is me. The documents are the defence, and they only work if I trust them more than I trust the last thing I read.