Related Work
DIAL solves a different problem than most AI frameworks. Understanding the distinction helps clarify what DIAL does.
What DIAL Is
DIAL is a measurement and delegation harness. It answers the question: can this AI specialist reliably predict what this human would choose, in this specific context? When the answer is yes — empirically demonstrated through alignment scores rising over repeated decisions — DIAL progressively delegates. When alignment degrades, the trip line fires and the system reverts.
DIAL can wrap any agent framework, alignment technique, or model architecture.
How DIAL Relates to Other Approaches
The key dimension of comparison is where ground truth comes from: the signal used to judge whether AI behavior is correct.
| Approach | Ground truth source | When trust is established | Can trust change at runtime? |
|---|---|---|---|
| LangGraph / LangChain | Designer's predefined graph | Before deployment | No |
| Multi-agent debate | Human judges per decision | Each decision | No (static) |
| Constitutional AI / RLHF | Offline training signal | Training time | No |
| Mixture of Experts | Gating network | Training time | No |
| DIAL | Human's actual runtime choices | Empirically, per decision cycle | Yes (progressive collapse + trip line) |
Agent Frameworks (LangGraph, LangChain, CrewAI)
Agent frameworks define how an AI system operates: the graph of states, tools, and control flow. DIAL defines whether an AI system should be trusted to operate autonomously at each decision point.
These are complementary. A DIAL specialist can be a LangGraph agent. DIAL wraps the agent and measures whether its decisions match human choices. The agent framework handles execution; DIAL handles trust calibration.
There's a deeper relationship: a DIAL machine can be viewed as an abstraction over many different agents executing simultaneously — each specialist proposes a different path through the state space. Through the arbiter's unified consensus score, DIAL materializes the superposition into a single concrete execution path. The machine definition is the shared structure; the specialists are competing interpretations of how to traverse it.
Multi-Agent Debate
Multi-agent debate uses multiple AI models to argue and a human to judge. DIAL's multi-specialist proposal mechanism is superficially similar, but the purpose differs: debate aims to improve answer quality through adversarial argument; DIAL aims to measure which specialist best predicts the human, with the goal of eventually removing the human from routine decisions entirely.
Constitutional AI / RLHF
Constitutional AI and RLHF train models against offline signals: a constitution document or human preference data collected in advance. The trust relationship is fixed at training time. DIAL's ground truth is the human's live, runtime choices in a specific operational context. Trust evolves continuously, per-specialist, per-state. A constitutionally-trained model can serve as a DIAL specialist; DIAL then measures whether the training generalizes to this particular human's preferences through alignment scores.
Mixture of Experts (MoE)
MoE architectures route inputs to specialized sub-networks via a learned gating function. The analogy to DIAL's specialist selection is real but shallow: MoE routing is learned at training time and frozen; DIAL's trust in specialists updates at runtime based on human feedback. MoE optimizes for task performance; DIAL optimizes for human prediction.
Using DIAL with Other Systems
DIAL wraps existing AI systems:
- Your agent framework handles task execution, tool calls, and control flow
- Your model handles reasoning, generation, and tool use
- DIAL handles the question: should this agent/model be trusted to act autonomously here, or does a human need to decide?
The specialist interface is intentionally minimal: anything that can propose a state transition and compare two proposals can participate in DIAL's decision cycle. The arbiter orchestrates the rest.