File 02 · entry
The goal
stack.
Aim at the terminal objective, not the proxy.
Keep asking "in service of what?" until the next answer wouldn't change anything you'd actually do. Every level below the top is a proxy.
01 · Start with the task
Whatever is in front of you. Then ask: in service of what?
02 · One level up
The answer is the level above: what the task is for. Ask again.
03 · Keep asking
Each answer lays a new level on top. Keep asking "in service of what?"
04 · The stop rule
Stop when the next answer wouldn't change anything you'd actually do.
05 · The top
What's left on top is the terminal objective. That's what you aim at.
06 · Everything below
Every level below the top is a proxy. And any proxy optimized hard drifts from what it was meant to measure.
Proxy drift
Push one number hard.
Any proxy optimized hard drifts from what it was meant to measure. The aim follows the number, not the target.
Aimed at the whole goal. The number measures one piece of it.
Most guardrails turn out to be parts of the real goal that the proxy forgot.
Ethics
Guardrails that matter are part of the real objective; the rest of the judgment stays with the operator.
When it's all you have
Sometimes a proxy is the best you can get.
Especially when you don't own the data. Then the job isn't to refuse it.
It's to know it's a proxy and keep it tied to the objective above it.
What it allows
When results go wrong, ask which layer.
The stack is a diagnosis tool. Pick what went wrong:
Fix how it was done. The stack stays as it is.
Fix at the lowest layer that explains the failure. Climb only when a good fix keeps failing.
HOW THE LOOP CLIMBSAnd for agents
The answer to the paperclip problem: agents aimed at the stack, not the proxy.
ANNEX A · LIMITSWHERE IT BREAKS
- There's more than one real objective. Some goals can't be ranked into one stack: welfare against consent, speed against safety, your interest against a partner's. Forcing one terminal objective then hides a choice about whose interests win. Hold them as a council of weighted seats instead.
- There's no goal yet. In new territory, the objective has to emerge from mapping first. That's a valid way in, but not a permanent mode.
- It becomes a rationalization engine. "The real goal is..." can justify anything. The guard: a terminal objective must show up in outcomes, not just in the argument.