Loop & Scoop: Four Loops, One Till

// key_findings

1. 01.

   Loop 1 is the one everyone builds and the one that quietly lies. Automating the four reasons on the laminated card clears 62.70 and leaves 62.60 unexplained, faster than before and no more correct.

2. 02.

   Loop 2 is four lines of Python, not a model. The cone has to add up to the gap and every void has to be labelled 'ask Marco'. A deterministic checker that returns jump_to: model is cheaper and stricter than any LLM judge.

3. 03.

   Loop 3 removes the human who starts the run, not the human who thinks. Thirty unattended nights is the win. Thirty nights of asking the same two questions is the tell that something upstream never learned.

4. 04.

   Loop 4 is the loop nobody ships: the agent proposes a patch to its own memory and prompt, an eval suite scores it against a held-out split, and an acceptance gate rejects anything that is not a strict improvement.

5. 05.

   The acceptance gate matters more than the optimiser. Disable it and the score wanders. Drop the generalise constraint and the training set hits 100% while held-out never moves, which is overfitting made visible rather than assumed.

6. 06.

   Some minutes are waste and some are the job. The over/short habit gets learned once. The void run stays in front of a person forever, because a shop where nobody looks at the till voids is not efficient, it just has a problem it has not found yet.