Skip to main content

Specialists

Specialists are the pluggable actors that participate in sessions. They can be AI models, webhooks, local functions, or humans.

Roles​

RoleDescriptionCan be AI?Can be Human?
ProposerAnalyzes state, suggests transitionsYesYes
ArbiterEvaluates consensus via alignment-weighted margin (built-in)NoNo

Proposers​

Proposers analyze the current state and available transitions, then submit a proposed transition with:

  • Transition name: Which transition to take (maps to a state machine edge)
  • Reasoning: Natural language explanation of the choice
  • MetaJSON: Structured data describing the state after taking the transition

Multiple proposers may propose the same transition with different reasoning (their contributions are clustered by the arbiter), or different transitions entirely. Each proposal acts as an endorsement of that transition.

The Arbiter​

The arbiter defaults to a deterministic, built-in component. Custom arbiter strategies can be registered, but the built-in strategies (alignmentMargin, firstProposal) are deterministic and auditable. It groups proposals by transition and applies an alignment-weighted margin to determine consensus:

  1. Solicits proposals from all enabled proposers
  2. Validates proposals — rejects invalid transitions
  3. Clusters proposals by transition — proposals for the same transition combine rather than compete
  4. Scores endorsements continuously as proposals arrive, weighted by alignment
  5. Declares consensus when one transition's alignment-weighted margin exceeds the threshold
  6. Blocks for human when all specialists have been exhausted without consensus
  7. Self-heals by re-enabling disabled specialists when anomalies are detected

See Arbitration for the full algorithm and self-healing mechanics.

Alignment Score​

Every specialist has an alignment score computed using the Wilson score lower bound — the lower bound of a 95% confidence interval for the true match rate between the specialist's proposals and human decisions.

alignmentScore = wilsonLowerBound(matchingChoices, totalComparisons, z=1.96)
  • Humans always have alignment = 1.0 (they are the ground truth)
  • New AI specialists start with alignment = 0.0 (no demonstrated alignment)
  • Alignment grows as the specialist's choices match human decisions
  • Small samples are penalized: 1 match out of 1 comparison yields ~0.21, not 1.0 — confidence requires evidence

The alignment score tracks how well a specialist's proposals match human decisions over time. It is used for evaluation and progressive collapse decisions — determining when a specialist has demonstrated enough alignment to operate without human oversight.

Semantic Isolation​

LLMs that serve as specialists are deliberately unaware of DIAL's internal mechanics. They do not know about proposers, consensus scores, or the framework itself. They receive:

  • The current state description (the decision prompt)
  • Available transitions (presented as tool calls)
  • Exemplar history: past human decisions presented as domain-native history ("In this situation, the transition was X, with this reasoning")

The LLM sees a domain-native decision problem, not a framework coordination problem. This semantic isolation prevents the LLM from gaming the consensus mechanism or optimizing for framework artifacts rather than decision quality.

Enable / Disable​

Specialists can be enabled or disabled at any time:

  • Enabled: Actively receiving solicitations and contributing proposals
  • Disabled: Paused — not solicited, but registration and alignment history are preserved

This is important for two reasons:

  1. Pruning: The system disables low-alignment or redundant specialists to reduce cost and latency without losing their history.
  2. Self-healing: The arbiter can re-enable disabled specialists when anomalies occur (e.g., the sole enabled proposer submits an invalid proposal).

Disabling is preferred over removing because alignment history is expensive to rebuild.

Human vs AI Specialists​

Human specialists have special privileges:

  • Can provide explicit proposals for transitions
  • Their proposals always win — a human proposal is treated as the decisive endorsement, equivalent to an override
  • Their decisions serve as ground truth for alignment measurement
  • A human proposal creates an exemplar (context + decision) for future few-shot learning
  • Alignment is always 1.0

AI specialists operate through strategies:

  • Must use strategy invocation (an LLM, webhook, or local function decides for them)
  • Cannot override — they can only contribute proposals that are counted toward consensus
  • Their alignment score starts at 0.0 and grows through demonstrated alignment with human choices

This asymmetry implements Human Primacy: humans are the ultimate authority, and AI specialists earn influence through empirical alignment.

Specialist Strategies​

Specialists execute their role through strategies — functions that determine what they propose. Strategies can be:

  1. Local functions: Code that runs in-process
  2. Webhooks: Remote services that receive context and return decisions
  3. LLM-based: Context is assembled and sent to a language model
  4. Hybrid: A local function provides context, then an LLM makes the decision

A specialist's configuration includes:

  • Strategy function: How the specialist makes its decision
  • Context function: Provides domain-specific context (optional)
  • Prompt: A string that frames the decision for this specialist
  • Model ID: For LLM-based strategies, the model to use (e.g., via OpenRouter)

Multiple Specialists​

DIAL supports any number of proposers:

  • Multiple proposers: May propose different transitions or the same transition with different reasoning. Proposals for the same transition are clustered and their endorsements combine.

This design enables:

  • Redundancy: Multiple perspectives on the same decision
  • Specialization: Different specialists optimized for different aspects
  • Competition: Specialists compete on alignment with human decisions
  • Progressive pruning: Redundant or low-alignment specialists are disabled over time