Every demo shows the happy path. What decides whether a system is safe to rely on is the other path, and specifically, what it says when something has gone wrong.
We have collected a few of these the hard way. Each one is a message that was technically accurate and led its reader to the wrong conclusion.
“SAP refused that read”
We pointed the platform at an SAP host that was routable and dead, and watched what it told the agent reasoning above it:
SAP refused that read: TOOL_RAISED, ConnectTimeout
SAP had refused nothing. It was never reached. But “refused” is a statement about the request, so the reasonable interpretation is the request was wrong, and the agent did the reasonable thing. It rephrased. Then it rephrased differently. Nine attempts, a different operator, a different period, and no answer at all, because no rephrasing makes an unreachable host reachable.
The lesson generalises past this system: a failure message that invites a retry, when retrying cannot possibly work, is worse than saying nothing. It names the one action its reader cannot take. An outage and a rejection need different words, because they warrant opposite responses.
“You got fewer rows than exist”
A server is allowed to return fewer rows than you asked for and hand back a pointer to the rest. If you ignore that pointer, you get a short page and no indication anywhere that it was short.
The subtle part is what happens next. The layer that reports truncation compares what it fetched against what it asked for, but it was handed the already short page, so it compares that page against itself, finds nothing missing, and reports a clean, complete read.
Both ends of the system now agree the figure is whole. It is a lower bound wearing a total’s clothes, and no component in the chain is in a position to notice.
“0”
A read that legitimately matches nothing and a read that failed to find the right place both produce the same number, and one of them is an answer while the other is a bug.
Reporting zero for the second kind is the most confident wrong answer available, because zero is a perfectly good finding: “we had no returns last month” is useful information. The platform distinguishes them: an empty result says the entity and window it covered, and where nothing could be read at all, it says that instead.
“Applied your reporting standard”
The most dangerous failure in this category is a wrong adjustment that arrives with a citation attached. A reader who would question a bare number accepts one that names the standard it was computed under, because the citation reads as evidence the figure was checked.
Which is why a rule about order counts in this platform is structurally prevented from adjusting a currency figure. Not instructed to be careful: prevented, because the match requires the kind of measure to agree as well as the subject.
The general shape
Every one of these is the same defect: a message that is locally true and globally misleading. The read did raise an exception. The page was returned in full. The count was zero. The rule was applied.
The test we use now is not “is this message accurate” but “what will a reader do next, and is that the right thing to do?” Those are different questions, and only the second one matters.