System One decision cascade

The idea

Bounded choices (route, next action, failure class, done, risk, intent) are typed questions. A rule answers first, then a typed arm, then a small LLM, each with a confidence.

The problem
Asking the generative model every small decision costs a full turn each time.
Why it matters
Small decisions are fast and cheap, and each answer is only a signal that code enforces.
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

  • Shadow first, promote on measured outcomes

    A new decision point starts in shadow. It is scored offline on labelled cases and live on outcomes, then promoted or demoted on that evidence. It acts only on the answer keys measured safe.

  • Fast path and step-boundary fan-out

    At each step boundary, one call asks every relevant purpose. Low-risk reversible next steps can run without a model turn, and read-only steps run while the model thinks.

  • Tier × effort routing from runtime signals

    Each call's model tier and effort are chosen separately. The inputs are the stage, the risk of the pending action, the change size, open findings, failures, repository size and the budget left. Floors tune from verified outcomes.

  • Flow signal triage

    Ideas, requests and problems arrive as signals. Rules decide each signal at once, using the triage employee's accepted lessons, while System One is measured next to them. Each decision becomes held work or started work.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.