← Catalogue

Conversational systems04

isanna turnflow

In build

A deterministic kernel for multi-turn conversation, with the model kept outside the execution path.

Turnflow runs a conversation through a seven-stage pipeline: classify, load session, plan, agent loop, inject results, compose response, persist session. The model proposes a plan; a deterministic validation boundary decides what actually runs. That separation is the whole engine — everything else follows from refusing to let generated text reach an executor directly.

C1
The plan is the single execution authority.
C2
No execution from raw user input.
C3
A deterministic validation boundary sits between plan and execution.
C4
The language model is never in the execution path.
C5
The core remains domain-agnostic.
C6
Grounded generation resolves to grounded-with-citations, or refusal.
Two execution classes
Structured mutation and grounded generation are both first-class, and both governed by the same turn invariant rather than by two parallel code paths.
Domain-free core
Domain logic registers through a pipeline configuration. The core carries no knowledge of any product built on it.
Transport adapters
Telegram, WhatsApp and web chat sit behind one adapter interface, so a channel is a plug rather than a fork.
Provider-agnostic gateway
Anthropic, OpenAI or a local Ollama model behind one interface — the provider is a configuration choice, not an architectural commitment.

The engine and its pipeline exist in TypeScript on Deno, with the transport adapters and the model gateway in place. The interfaces are still moving and Turnflow is not yet published.