Hypothesize

To say in advance what you expect a change to do, so that afterwards you can tell whether it did it.

One sentence, written before the change is installed: if this line changes, that answer will change in this way. Then the test — the same question, in fresh conversations, asked two or three times, because answers vary from one run to the next.

It is load-bearing because without a prediction there is no way to tell a repair from a coincidence. An agent's account of why it acted is a reconstruction, not a readout; it cannot see its own training, and trained habits have no lines to point to. So its diagnosis is a guess worth testing, and the hypothesis is what turns the guess into something that can be wrong. This is the ordinary discipline of learning coupled tightly to action, run at the scale of one line of text.

A Player+ predicts, changes one thing near the end of a session of edits (gather several early, fine-tune one at a time late), tests in fresh conversations, adds a different case to see whether the behavior generalised, and adds a routine task to check the fix has not made the agent timid. Then they keep what helped and undo the rest.

The habit travels beyond agents. Any change to how a Cell works can be stated the same way before it is made, and a season later the Circle knows whether it was the change or the weather.

Also called: hypothesis Stands on: Agent engineering · Tightly Coupled Learning and Action · Set-up Opens onto: Change log · Soak · Drift · Improvement log · Look-back In play: beyond Sources: Player+ Modules, Agent Engineering, Advanced Agent Engineering · The DNA of Heaven, Part IV (Tightly Coupled Learning and Action). Open: none found.