official/fixed_juice_ralph

ralph_loop with a governor on it: a fresh session every round, and between the rounds the effort moved a rung to hold the agent to juice output tokens per turn of the model.

hmz exec -f official/fixed_juice_ralph -a claude/claude-opus-5:high "$(cat TASK.md)"

official/fixed_juice_ralph

A governor rather than a brake: what it holds steady is the size of an answer. Under the target the agent is asked to think harder, over it to think less — one rung of its own model’s ladder per round.

A governor is not a brake

What it holds steady is the size of an answer. A loop held to 2000 output tokens a turn is a loop that goes on producing 2000 output tokens a turn for as long as anybody leaves it running; what ends it is still the budget.

The two are one quantity read at two scales: juice is what a turn is worth, and budget is what the loop is.

Per turn of the model — one request and the answer to it — rather than per turn of the flow, which is many of those and as many again of whatever the tools took. That average is what an effort actually moves: a model asked to think harder writes more in each answer and takes longer over it. So an agent under the target is asked to think harder and one over it to think less, one rung of its own model's ladder per round, so that the loop settles rather than swings.

Nothing here is a clock. How long a round takes and what it costs an hour are what the model and the work make of it.

What it takes

juice: 2000     # output tokens an average turn of the model is to come out with
over: 300       # how far back that average is taken, in seconds
slack: 0.15     # how far off it may be before the effort moves
rest: 5         # seconds between rounds
budget: 10      # millions of output tokens the whole loop may come to

juice	What a turn of the model is to come out with, in output tokens. 2000.
over	How far back the average is taken, in seconds. 300.
slack	How far off the average may be before the effort moves, as a fraction of the target — 0.15 leaves it alone between 85% and 115%.
rest	Seconds between rounds. 5.
budget	Millions of output tokens the whole loop may come to. 10, 0 for no limit.

What it keeps

rounds, output, and effort — the rung the governor settled at. A loop started again at the top of the ladder walks back down to that rung one paid turn at a time, which is the thing worth keeping.

The rung is kept as the effort's own word rather than as a place on the ladder, because the ladder is whatever the account says its CLI runs today: a model retired and an effort added both move the places. A word that is no longer on the ladder is read as no rung, and leaves the agent where a first run would start it.

The answer size is not kept. It is an average over the last few minutes of turns, and a run starting today has none of yesterday's minutes to average.

See also

• Efforts — what a rung is, and which ladder each model has
• Cost and rate — reading what a run is spending while it runs
• ralph_loop — the same loop, ungoverned