zero-framework-cognition

name	zero-framework-cognition
description	Use when designing or reviewing model-owned contracts, output schemas, backend guards, routing, semantic parsing, intent detection, or data fields where application code might duplicate judgment or ask a model for redundant facts.

Zero-Framework Cognition

Core Principle

Models decide; server executes.

Application code must not replace model judgment with keyword routing, regex intent detection, heuristic classification, duplicate semantic checks, or redundant field computations. The backend should execute tools, persist state, and normalize deterministic output shape. Backend correction is a last resort: first prove the prompt/schema/agent contract cannot prevent the bad payload on the real path, then add only the narrowest correction-only invariant.

Reference: Zero Framework Cognition: A Way to Build Resilient AI Applications.

Model Decision Engines

In this repository, "model" means:

• The LLMs used by the application for reasoning, narration, schema output, routing decisions that require context, and semantic judgment.
• The FastEmbed semantic intent classifier where the repository already uses it as a model-backed local classifier.

Do not replace either with new application-code heuristics. New regex, substring checks, weighted keyword lists, or phrase maps for semantic judgment are ZFC violations.

Minimal Semantic Fields

Ask the model for the smallest set of fields that expresses the semantic decision. Do not ask the model to emit multiple equivalent fields that can drift apart.

Prefer:

{
  "current_level": 4,
  "target_level"