Directive

The OAS Layer 3 (C3) Directive resource defines immutable behavioral rules or “hard” conditions for the AI evaluation engine.

Unlike rubrics (which define qualitative levels of competence), directives impose systematic quantitative penalties or global rules (e.g., deductions for spelling errors) that the LLM must inexorably execute during the correction process.

Schema

Metadata (metadata)

Field	Type	Required	Description
id	string	Yes	Unique identifier for the directive (e.g., directive.es.c3.lomloe...).
title	string	Yes	Descriptive title of the directive.
country	string	No	Country code to which the regulation applies (e.g., ES).
layer	string	Yes	Architectural layer (usually C3).

Specification (spec)

Field	Type	Required	Description
target_subjects	string[]	No	List of subjects to which it applies by default.
rules	array	Yes	List of rules or directives to apply in the evaluation.

Structure of rules

Each rule within rules has the following fields:

Field	Type	Required	Description
id	string	Yes	Rule identifier.
type	string	Yes	Type of rule (e.g., penalization).
description	string	Yes	The exact instruction (hard prompt) that the LLM agent will process for evaluation.
max_deduction	number	No	Maximum score that can be deducted for the infringement of this rule.

Example Usage

Below is an example of a spelling directive for the Spanish LOMLOE curriculum:

apiVersion: colabedu.ai/v1beta1
kind: Directive
metadata:
  id: "directive.es.c3.lomloe.ortografia.v1"
  title: "Global Spelling Correction Directive"
  country: "ES"
  layer: "C3"
spec:
  target_subjects: ["Spanish Language and Literature", "History", "Philosophy"]
  rules:
    - id: "ortografia-01"
      type: "penalization"
      description: "For each spelling error, 0.25 points will be deducted from the final score of the exercise."
      max_deduction: 2.0
    - id: "tildes-01"
      type: "penalization"
      description: "For each accentuation error (tildes), 0.10 points will be deducted."
      max_deduction: 1.0