Saimaa

A person wanted to cross a corner of Lake Saimaa, in Finland, by packraft, and decided to plan it with Claude, an artificial intelligence. For two months Claude read, measured, programmed and wrote; the person asked the questions and made the decisions.

The starting point was a GPX track downloaded from the internet: a kayaker's, from 2018, 104.5 kilometres between the two ends. A track is what somebody felt like paddling one particular day. The first thing was not to believe it.

Looking at an online map, a human question: couldn't you cross this peninsula through the middle? The project had it noted as "4.6 km shorter" and dropped. Claude measured it over the water, with a program rather than by eye: the earlier measurement started two kilometres late. It was 11.84 km shorter, a whole stage.

Here the track went down four drops of water. Claude had left a second outlet of the lake noted as a hypothesis. A human question brought it back, and measured it turned out 5.07 km shorter, with no rapids and a carry of 106 metres from water to water.

The rule that saved the most work: no figure is printed without its method. Claude wrote a program for each calculation — shortcuts over water, fetch, the point of no return, road exits — on open data: some 24,000 official map tiles and 61 described sources. The 104.5 kilometres of someone else's track came down to 80.53.

Every question ended in a report with its sources: thirty-three. From them came the notebook's boxes, the road exits and the stretches inside official fairways. And two more documents: a register of what we did not know, and an agenda written by an agent that was asked not to research but to raise the questions nobody had asked. It raised 84.

To review the notebook, Claude launched four critic agents over three rounds, each with a different lens. They produced 115 findings and a judge accepted 80. Acceptance rose round after round — 63, 76, 74 % — and no rejection was for being false. A human question stopped it: do the critic agents have too much power?

Some findings were true, and serious. One of them: the notebook stated that the route had no rapids, and the route went right here, through them. The AI itself had written that sentence on day one.

Eight refuters were added with the burden of proof reversed: not "is it plausible?" but "can I prove it false?". Of the 80 accepted findings, 35 carried a factual error. In gold, what was confirmed; struck through, what did not hold. Since then, to get onto the page a change has to win, not tie.

Refuters are not infallible either. Here a critic reported a camp "in the water" and a refuter found that both of them used a tool that counted rock as water. But the refuter itself got the wind wrong: measured later on the official water polygon, the critic was right. In the end the measurement rules, not an agent's verdict.

The most dangerous errors give no warning. Twelve trains printed in UTC, three hours off, with the right numbers. A server that answered "empty" with a success code. A deletion of 46 notes that did not change a single page of the notebook. Every time one slipped through, a guard was written so it would not happen again.

The notebook was printed twenty-four times and, from the second on, none overwrites the previous one, so they can be compared. When something was wrong it was fixed in the data it comes from, not on the paper: a false timetable ended up copied in nine places because it was fixed where it showed, not where it was born.

Every session ended by writing down what had been learned. The errors that could happen again in another project became more than thirty named lessons. That is the part of this work that serves any other. The result — the crossing, its seven stages and its notebook — is the case that proves it.