Skip to main content

Inputs & Outputs

Every AI Product must declare its inputs and outputs.
Inputs define what the product consumes.
Outputs define what the product produces.
Together, they establish the data contract of the AI Product.

Why Inputs & Outputs Matter​

  • Transparency → Consumers know what to provide and what to expect.
  • Discoverability → Products can be cataloged by input/output types.
  • Composability → Enables chaining of products into workflows.
  • Governance → Inputs and outputs may be subject to different compliance rules.
  • Reproducibility → Clear data contracts ensure consistent behavior across deployments.

Input Declaration​

An AI Product must declare:

  • Input Types → Structured (JSON, CSV), unstructured (text, image, video), multimodal.
  • Input Schemas → Machine-readable specifications (JSON Schema, Avro, Protobuf).
  • Preprocessing Requirements → Normalization, tokenization, feature engineering.
  • Constraints → Size limits, language support, format restrictions.
  • Sensitivity → Whether inputs contain personally identifiable or regulated data.

Output Declaration​

An AI Product must declare:

  • Output Types → Classification label, probability score, generated text, embeddings, multimodal outputs.
  • Output Schemas → Machine-readable specifications.
  • Confidence Metrics → Probabilities, error margins, or uncertainty estimates.
  • Postprocessing → Transformations applied before output delivery.
  • Explainability Attachments → Optional explanatory signals (feature importance, heatmaps, rationales).

Contract Characteristics​

  • Deterministic vs Probabilistic → Declare whether outputs are reproducible given the same inputs.
  • Stability Across Versions → Interfaces must remain stable unless declared in versioning.
  • Error Handling → Define error outputs and fallback modes.
  • Governance Markers → Tag inputs/outputs with regulatory relevance (e.g., HIPAA, GDPR).

Example​

AI Product: Sentiment Analyzer

  • Inputs:
    • Type: Text (UTF-8 encoded).
    • Schema: JSON with field {"text": "string"}.
    • Constraints: ≤ 500 tokens.
  • Outputs:
    • Type: JSON object.
    • Schema: {"label": "positive|neutral|negative", "confidence": float}.
    • Confidence Metric: Softmax probability score.
  • Explainability: Feature importance (keywords influencing sentiment).

Summary​

  • Inputs and outputs define the data contract of an AI Product.
  • Must declare types, schemas, constraints, and governance markers.
  • Outputs must include confidence and explainability signals where applicable.

Principle: An AI Product without clear input-output contracts is an unreliable asset — not a trustworthy product.