Core Concepts
DIAL orchestrates specialists — both AI and human — that compete and collaborate to navigate state machines through decision cycles. An arbiter drives each cycle, grouping proposals by transition and declaring consensus when one transition's alignment-weighted margin exceeds the threshold.
The Big Picture
Task Specialists, Not Agents
DIAL does not guide a single agent toward completing a task. It simultaneously solicits proposals from an arbitrary number of models, prompts, and strategies — all competing at the same decision point. Each registered proposer independently analyzes the current state and suggests a transition. The arbiter groups proposals by transition and declares a winner when one transition's alignment-weighted margin exceeds the threshold.
This is mass simultaneous solicitation, not sequential A/B testing. Specialists are interchangeable and compete on the quality of their contributions, measured against human ground truth.
A DIAL "specialist" is scoped to a specific role in a specific decision. A proposer suggests the best transition; the arbiter orchestrates the entire process. Specialists can be AI models, webhooks, local functions, or humans.
Learn more about Specialists →
Sessions
A session is an instance of a state machine. It starts in an initial state and progresses toward a goal state through decision cycles.
Specialists
Specialists are the pluggable actors that participate in sessions. They fill two roles:
| Role | Description | Can be AI? | Can be Human? |
|---|---|---|---|
| Proposer | Analyzes state, suggests transitions | Yes | Yes |
| Arbiter | Evaluates alignment margin consensus via alignment-weighted margin (built-in) | No | No |
Specialists can be enabled or disabled. Disabled specialists remain registered (with their alignment history intact) but stop receiving requests. The arbiter can re-enable disabled specialists when needed — for example, if an enabled proposer submits an invalid proposal.
Learn more about Specialists →
The Decision Cycle
When a session needs to progress, the arbiter works through a sequence of solicitations at a steady pace:
- Solicit Proposals: Call each enabled proposer. Validate and cluster proposals by transition.
- Block for Human (if no consensus): All specialists have responded, no consensus. Wait for a human-forced decision.
After every arriving proposal, the arbiter re-evaluates the consensus score. If one transition's margin of superiority crosses the threshold, consensus is declared immediately — the arbiter doesn't wait for all responses.
Learn more about the Decision Cycle →
The Consensus Score
Every proposal is an endorsement of a transition. The arbiter counts proposals per transition and evaluates whether one transition has built a sufficient lead.
The arbiter groups proposals by transition (not individual proposal). Two proposers that chose the same transition are supporting the same outcome — their endorsements count together.
Consensus is reached when the alignment-weighted margin exceeds the threshold:
margin = (leaderScore − runnerUpScore) / totalAlignment
consensus when: margin >= threshold
The consensus threshold (a float, 0–1) controls how much alignment-weighted agreement is required. A lower threshold allows faster consensus; a threshold of 1.0 requires unanimity — all proposals must agree on the same transition.
Learn more about Arbitration →
Human Primacy
The fundamental principle underlying DIAL:
Humans have context that AI cannot access. AI specialists are judged on their ability to predict what humans would choose. When consensus cannot be reached, only a human can force a decision.
Human alignment is always 1.0 — they are the ground truth. A human proposal always wins consensus immediately. A human forcing a proposal bypasses the score entirely.
Every human-forced decision creates an exemplar: a capture of the full context plus the human's choice. Exemplars are the training data that drive progressive collapse — they improve specialist context through few-shot learning and provide ground truth for alignment measurement.
Learn more about Human Primacy →
Progressive Collapse
As alignment improves, the system naturally collapses:
- Cold start: All alignment = 0. Every contribution is multiplied by 0. System blocks for human.
- Calibration: Humans decide, exemplars accumulate, alignment scores grow.
- Autonomous consensus: Aligned specialists' contributions cross the threshold without human participation.
- Pruning: Low-alignment and redundant specialists are disabled. Cost drops.
- Champion: One specialist handles decisions alone with periodic human spot-checks.
- Collapsed: A fine-tuned cheap model runs at a fraction of the original cost.
If alignment degrades, the trip line fires: the arbiter reverts to a more deliberative configuration, re-enables disabled specialists, and the calibration loop begins again.
See the full process → · Implementation details →
Concepts in This Section
- Sessions: State machine instances
- Specialists: AI and human actors
- Decision Cycle: The arbiter's async solicitation sequence
- Arbitration: The unified consensus score
- Consensus Strategies: The default algorithm and alternatives
- Human Primacy: The foundational principle
- Related Work: How DIAL relates to other approaches