Technical Architecture (Overview)

Architecture: Open Assessment Standard (OAS v1beta1)

Purpose: To explain the technical operation of the LLM orchestration engine, demonstrating how the separation of responsibilities (Assessment as Code) allows evaluating any Spanish-speaking educational paradigm without requiring additional model training (Fine-Tuning).

The Problem: The “Prompt Engineering” Monolith

Historically, EdTech platforms attempted to use LLMs by sending them a “Mega-Prompt” in plain text. This prompt mixed the educational law, the text to be read, report design instructions (e.g., “use bold” or “create a Markdown table”), and the specific rules of the teacher.

The language model was forced to act simultaneously as a pedagogical evaluator, legal analyst, and frontend designer. The consequences of this approach were disastrous at production scale:

1. Prompt Drift and Hallucinations: LLMs have a limited attention mechanism. By inundating them with contradictory instructions, the model suffered “amnesia,” forgetting to apply critical grading rules by expending its capacity on formatting.
2. Technical Fragility: If the LLM omitted a format tag, the backend parser failed, causing a cascading error.
3. Critical Dependence on Fine-Tuning: Massive costs were incurred training specific models for each educational law, leading to rapid obsolescence with any curriculum change.

The Modular Solution

To solve this chaos, the OAS architecture breaks down the problem into specialized components, analogous to how microservices architecture revolutionized traditional web development.

Throughout the following sections of this documentation, we will explore in detail the technical pillars that support the ColabEdu ecosystem:

1. Assessment as Code (The 4 Layers): How we fragment knowledge using the MVC pattern.
2. Entity Relationships (The Exercise): How ExerciseSpec, AssessmentItem, and ExerciseType relate to each other.
3. The Evaluation Engine (Late Binding and UI): How we inject context in real-time and enforce structured JSON schemas.
4. Agent Ecosystem: The network of specialized intelligences operating on the Spec Manager.
5. GitOps Hierarchy: The repository inheritance model that enables global scalability.