The previous three left photographs and memories. This one is being documented well enough that its decisions can survive the next teardown.
This aquarium has been torn down and replanted several times. Each version taught me something — which plants sulked where, which equipment was worth the money, what a maintenance routine actually needs to look like — and almost none of it survived in a form I could use later.
What carried forward were impressions. I remembered that something had worked, rarely the conditions it worked under. When the tank was rebuilt, those impressions were the only inheritance, and impressions are exactly the sort of evidence that quietly rearranges itself to agree with whatever you already believe.
Pearl started as a record, not a controller. The first thing it did was read the probes and write down what it saw. Nothing was automated for a good while after that, and the order was deliberate.
Automating a system you do not yet understand does not remove mistakes. It makes them consistent, and it makes them invisible, because a schedule that runs quietly attracts far less attention than a person forgetting to do something.
Automation applied to an unexamined routine just repeats the mistake more reliably.
Parts of this tank are not automated, and parts of the record are incomplete. The first PAR survey does not carry per-position depth. A mounting-height comparison was run and its readings were never preserved as a standalone dataset.
Both are published as they stand. A record that only contains the parts that went well is a brochure, and it would be a worse guide to what building this actually involves.