Core Concepts

Not a Chatbot — A Composition Engine

Ctrl AI is not a chatbot wrapper. It's a composition engine that chains expert-verified reasoning units into deterministic workflows. The AI is constrained by verified structure — it cannot hallucinate outside the boundaries set by your domain experts.

Five Unit Types

Every domain's reasoning decomposes into five irreducible types:

Two Brains

• Brain 2 — Deterministic execution engine. A JSON DSL interpreter that takes the same inputs and always produces the same outputs. Zero hallucination. Used for Program units.
• Brain 1 — LLM constrained by verified unit structure. Produces professional prose, but bounded by expert-verified reasoning. Used for Framework, Pattern, Argument, and Protocol units.

Given / When / Then

Every unit is defined in a structured format readable by domain experts (not just engineers):

• GIVEN — Prerequisites and context (what must be true before the unit fires)
• WHEN — Trigger conditions (what input pattern activates this unit)
• THEN — What the unit produces (typed outputs with verification level)

This format is verifiable element-by-element and executable by the composition engine.

Typed Inputs & Outputs

Every unit declares typed inputs and typed outputs. The composition engine chains units by matching output types to input types:

Unit A (Program): income_data → verified_income (number)
Unit B (Framework): verified_income (number), credit_history → risk_factors (object)
Unit C (Protocol): risk_factors (object) → decision (approve | deny | escalate)

This type matching enables automatic composition — the engine builds a DAG (directed acyclic graph) of units that can satisfy a query.

Workflows

A workflow is a named, versioned composition of units forming a complete process.

Example: "Customer Credit Risk Assessment"

1. Income verification (Program) → verified_income
2. Credit history analysis (Framework) → risk_factors
3. Regulatory compliance check (Program) → compliance_status
4. Fraud pattern detection (Pattern) → fraud_flags
5. Decision protocol (Protocol) → approve/deny/escalate
6. Rejection justification (Argument) → regulatory citations

Each unit has typed I/O. The workflow is the auditable artifact — show it to any regulator.

Trust Gradient

Every claim in every response is tagged with a verification level:

The trust gradient gives every stakeholder — from the end user to the auditor — immediate visibility into how much of the AI's response is backed by verified reasoning.

Domains

Domains are organizational tags for knowledge areas (Finance, Legal, HR, Clinical). They are used for:

• Library organization — filter units by department
• Team assignment — who reviews what
• Accountability — who owns this unit
• Relevance ranking — composition prefers same-domain units

Domains are not composition boundaries — the engine searches all units in the org when building a response.

Element Primitives

Units are built from 7 element primitives: