Skip to main content
Jira progress: loading…

NLPI

Interpretation Engines

1. Overview

NLP Interpretation Engines (NLPI) are assistive calculation engines within the ZAYAZ Computation Hub that interpret textual or unstructured inputs using natural language processing and large language models (LLMs).

NLPI engines transform human-readable content into structured interpretations such as classifications, scores, labels, or extracted semantic fields.

NLPI is an engine category. Concrete implementations are registered as micro-engine instances (MEIDs) that implement the NLPI type.

2. Design Principles

  1. Probabilistic by Nature
    NLPI outputs are inherently probabilistic and must include confidence indicators.

  2. Non-Authoritative by Default
    NLPI results are advisory unless explicitly approved and governed.

  3. Explainable Interpretation
    Interpretations must be traceable to prompts, models, and inputs.

3. Scope of Responsibility

3.1. What NLPI Engines Do

  • Interpret narrative disclosures and documents
  • Classify text against predefined categories or frameworks
  • Extract structured fields from free text
  • Assign scores or labels based on semantic meaning

Typical use cases:

  • Classifying documents by ESRS topic
  • Interpreting supplier disclosures
  • Mapping narrative statements to structured ESG inputs

4. What NLPI Engines Do Not Do

  • ❌ Perform deterministic computation (CALC)
  • ❌ Validate data correctness (VALI)
  • ❌ Emit canonical USO signals by default
  • ❌ Enforce compliance or policy decisions

NLPI engines interpret meaning, not truth.

5. Inputs

NLPI engines consume:

  • Text documents or free-form user input
  • Interpretation prompts or schemas
  • Contextual parameters (framework, language, scope)

6. Outputs

NLPI engines produce:

  • Structured interpretation payloads
  • Classification labels or extracted fields
  • Confidence scores
  • Provenance metadata (model, prompt, version)

Outputs are typically routed through:

  • CFIL (confidence filtering)
  • ZARA (presentation, governance, approval)

7. Governance & Trust

All NLPI outputs must:

  • be clearly labeled as assistive,
  • carry confidence and provenance metadata,
  • be subject to governance before downstream use.

Human-in-the-loop review is strongly recommended for material decisions.

8. Canonical Identification

  • Engine Type: NLPI
  • USO Code: NLPI
  • Category: Assistive Calculation Engine
  • Layer: Computation Hub
  • EXTR — Extraction Engines
  • CFIL — Confidence Filter Engines
  • ZARA — Orchestration & Explainability

Status: Stable
Owner: Intelligence Governance / ZARA



GitHub RepoRequest for Change (RFC)