Architecture / Clarity Engine
The intent layer before generation and agent routing.
Clarity Engine intercepts ambiguous work before a model or agent acts. It detects missing variables, asks the smallest useful clarification, compresses resolved intent, and routes the work through the right downstream path.
Core layers
How Clarity Engine is structured
01
Intent capture
Classifies the request into intent patterns and identifies subject, scope, timeframe, domain, evidence need, and output shape.
02
Ambiguity resolution
Maps ambiguity to a neutral clarifying question, then records the resolved intent for the session when appropriate.
03
Compressed handoff
Forwards a cleaner prompt, agent route, context request, or human approval path after the user intent is pinned.
Operating flow
The request path through the product
Prompt
Raw request enters the clarity gate
The engine evaluates whether the instruction is clear enough to execute safely and usefully.
Clarify
One question beats a long wrong answer
If the request is underspecified, Clarity Engine asks the highest-value clarification before generation.
Route
Resolved intent chooses the path
The clean intent can then be sent to UCL, Importance Engine, an approved model endpoint, a tool, or a human review route.
Governance
Boundaries the architecture must respect
Naming
Public product, private internals
Clarity Engine is the public name. ICA remains internal implementation shorthand and should not expose proprietary mechanics.
Determinism
Pattern-first boundary
The public architecture is rule and pattern based before generative output; the site should not imply an opaque ML-only clarity layer.
Trace
Clarification cycle visibility
Analytics can report clarification cycles and cache behaviour without publishing raw prompts as product claims.
Integration points
Where it connects to the wider stack
Proxy
OpenAI/Anthropic-compatible gateway
The proxy can intercept ambiguous requests and pass through resolved calls to approved upstreams.
SDKs
Developer integration
TypeScript, Python, CLI, and MCP-facing surfaces support controlled product integration.
KynticAI stack
Context and importance handoff
Resolved intent can request governed context from UCL and route weighted decisions through Importance Engine.
Evidence boundary
What this page proves, and what it does not claim yet
Built
Local product surfaces
The repo contains SDK, CLI, MCP, proxy, Rust API, taxonomy, ambiguity, and clarification test coverage.
Environment
Live deployment proof
Live Redis and upstream-provider proofs are environment work and are not presented as a public production claim here.