Open Source Framework
DIAL
Dynamic Integration between AI and Labor
Measure the cost of AI. Automate what's proven. DIAL is a coordination framework for AI and human specialists making decisions together within state machines.
The question most AI deployments skip.
Given any task modeled as a state machine, how do you know, in dollars, time, and quality, exactly what it would cost to turn that task over to a minimally competent AI decision-maker?
Most organizations deploy AI by hoping it works. They invest in models, build pipelines, and launch. Months later they discover the AI makes different decisions than their humans would have made, at costs they didn't anticipate.
DIAL takes the opposite approach. AI has no role by default. Tasks are assumed too difficult for artificial intelligence until proven otherwise, one decision at a time. Rather than asking "can AI do this?" in the abstract, DIAL provides empirical, per-decision-point data.
Three Principles
Built on a foundation of earned trust.
Human Primacy
Humans remain authoritative not because they are infallible, but because they possess contextual knowledge inaccessible to AI systems. Lifetime embodied experience, institutional knowledge, and real-time sensory input are things specialists simply cannot access within bounded context windows.
Progressive Collapse
Repeated measurement of how well AI predicts human choices causes the multi-agent deliberation structure to gradually "collapse" into deterministic execution. As AI alignment proves itself, the system progressively delegates to the most trusted specialist, with automatic reversion if alignment degrades.
Empirical Trust
Trust develops through demonstrated alignment between AI recommendations and actual human decisions. Confidence is built through measurable performance, not initial assumptions. Every decision point generates dollar-precise cost data, alignment rates, and latency measurements.
How It Works
Eight steps from human-operated to AI-automated.
- Define states, transitions, and decision prompts as a state machine
- Create interfaces for humans to view context, select transitions, and document reasoning
- Develop LLM prompts that draw on session history and examples
- Recruit and educate human operators on decision criteria
- Register multiple AI specialists using different models and prompt approaches
- Humans lead while AI submits parallel proposals in shadow mode. Humans retain approval; AI learns
- The arbiter tracks alignment and deactivates underperforming specialists
- When confidence reaches thresholds, AI assumes autonomous decision-making with periodic spot-checks
Technical Details
Built for real workflows.
State Machine Architecture
Every workflow is modeled as a state machine with discrete decision points, named transitions, and decision prompts. Define machines in JSON and run them from the CLI, MCP server, or TypeScript library.
Specialist Roles
Proposers suggest state transitions with reasoning. Arbiters evaluate consensus using alignment scores and configurable thresholds. Both can be human or AI, with cost tracking per decision.
Alignment Scoring
Wilson score lower bounds measure specialist alignment, accounting for sample size. Specialists compete at each decision point. Register multiple models and let empirical performance determine the winner.
Three Access Methods
CLI (npx dialai machine.json), MCP Server (npx dialai --mcp with 12 tools), or TypeScript library for direct programmatic integration. PostgreSQL storage available.
Constitutional AI Governance
A formal Constitution ensures specialists behave as participants, not authorities. Hard constraints on fabrication and manipulation enforce a priority hierarchy placing human alignment first.
Model-Agnostic Competition
Register GPT-4o, Claude, Llama, or any other model at the same decision point. DIAL doesn't pick favorites. It measures which model actually aligns with human judgment for each specific task.
Use Cases
Any workflow with discrete decision points.
DIAL works wherever decisions can be decomposed into states and transitions, and where you want empirical evidence before trusting AI.
- Content review and publishing pipelines
- Multi-stage approval chains
- Triage and classification workflows
- Quality assurance inspection processes
- Customer service escalation routing
- Any process where you want to measure whether AI can replace human judgment at specific decision points
Need help implementing DIAL?
DIAL is open source and MIT-licensed. We offer enterprise implementation, custom integration, training, and ongoing support for organizations deploying DIAL in production.