Skip to main content

Decision Cycles

The repeating process that drives state transitions.

The Cycle​

Propose -> Arbitrate -> Execute -> (repeat)

Each cycle attempts to move the session from one state to another.

Phases​

1. Propose​

Each registered proposer submits a transition proposal.

What proposers receive:

  • Current state
  • Available transitions
  • Transition prompts
  • Session history

What proposers return:

  • Action name (e.g., approve, reject)
  • Target state
  • Reasoning

2. Arbitrate​

The arbiter evaluates consensus using alignment margin.

How it works: The arbiter computes an alignment-weighted margin across proposals. Each proposal is weighted by its specialist's alignment score. The margin must meet the threshold for consensus. Human proposals always win via submitArbitration.

Human primacy: When proposals don't produce consensus, only a human specialist can force a decision via submitArbitration.

3. Execute​

If consensus is reached, the winning transition is applied.

What happens:

  • Session state updates to the target state
  • TransitionRecord added to session.history
  • Next cycle begins (or session completes if goal reached)

Watching the Cycle​

npx dialai machine.json --verbose

Verbose output shows:

[PROPOSE] ai-proposer: approve -> approved
[PROPOSE] ai-proposer-2: reject -> draft
[ARBITRATE] consensus reached: approve (alignment margin)
[EXECUTE] draft -> approved

No Consensus​

If proposals don't produce consensus, the engine reports needs_human status:

  • The session pauses and waits for human input
  • A human specialist can force a decision via submitArbitration with an explicit transitionName
  • Once a human decision is made, the cycle continues on the next tick

Session Completion​

The session ends when:

  • Current state equals goalState (success)
  • No consensus and no human intervention (waiting)
  • No valid transitions available (stuck)