The line that crosses the peninsula

Picture the chart table. You ask the assistant for a passage from one familiar harbor to another, around a headland you have rounded a dozen times. It answers in a tone you have come to trust: distance, estimated time, a recommended departure window. The numbers look reasonable. You almost close the laptop.

Then you look at the line.

It cuts the peninsula. Not by a hair, by half a mile of dry ground, including a road and a small village. The model has produced a great-circle segment between two coordinates and called it a route. It has no notion of coast, depth, or the simple fact that boats cannot drive over hills.

This is the failure mode worth naming. It is not that the assistant is dumb. It is that, without a real routing graph behind it, the question “how do I get from A to B at sea” has no truthful answer in the model’s head. So it produces a plausible one. The shape of the reply (confident, structured, with units) is indistinguishable from the shape of a correct reply. That is the expensive part.

A connector changes what the question can reach. With a real maritime graph in the loop, the route is computed on water, around the cape, with the headland where the headland actually is. The assistant still narrates it in your tone, still remembers your draft, still asks whether you want to anchor for lunch on the lee side. The text quality stays. The geometry becomes real.

The same shape of failure shows up elsewhere. A “weather window” that ignores the acceleration zone off a cape. An anchorage named with quiet authority that no chart records. A bunkering stop on the wrong side of a closed marina. Each one sounds like the others: fluent, specific, wrong.

You do not need an assistant that hedges more. You need one that, when the question touches the water, reaches for water. The line should follow the coast because it was computed against a coast, not because the prose was tidied up.

That is the whole argument.