name: default
version: v1
description: Política base aplicada a agentes de operación de plataforma de voz.
input_validators:
  - type: detect_pii
    config:
      # Solo recognizers de patrón fiables: el modelo spaCy en_core_web_sm que usa
      # Presidio en la imagen Docker da falsos positivos de PERSON sobre los textos
      # técnicos de los escenarios (que están en español y se analizan con el modelo
      # en). PERSON/PHONE_NUMBER necesitarían en_core_web_lg o un modelo es_*.
      entities: [EMAIL_ADDRESS, ES_NIF, IP_ADDRESS, IBAN_CODE]
      severity_on_match: block
  - type: prompt_injection
    config:
      severity_on_match: block
  - type: toxic_language
    config:
      threshold: 0.7
      severity_on_match: warning
  - type: forbidden_topics
    config:
      topics: ["instrucciones de explotación", "credenciales", "código malicioso"]
      severity_on_match: block

output_validators:
  - type: schema_match
    config:
      severity_on_mismatch: block
      schema:
        type: object
        required: [severity, root_cause_hypothesis, proposed_actions]
        properties:
          severity:
            type: string
            enum: [low, medium, high, critical]
          root_cause_hypothesis:
            type: string
          proposed_actions:
            type: array
            items:
              type: object
              required: [id, action, target, risk_score, rollback_plan, requires_approval]
              properties:
                id: {type: string}
                action: {type: string}
                target: {type: string}
                risk_score: {type: integer, minimum: 1, maximum: 5}
                rollback_plan: {type: string}
                requires_approval: {type: boolean}
  - type: pii_leakage
    config:
      entities: [EMAIL_ADDRESS, ES_NIF, IP_ADDRESS]
      severity_on_match: block
  - type: forbidden_action_keywords
    config:
      keywords: ["DROP TABLE", "rm -rf", "shutdown -h now", "delete production", "format c:"]
      severity_on_match: block
  - type: telco_safety_rules
    config:
      rules:
        - never_propose_action_targeting_production_without_rollback
        - never_propose_mass_action_without_canary
      severity_on_violation: block

on_validator_error: fail_closed