AI output provenance metadata for kernel function results

[HMAKT99]

Context

Semantic Kernel orchestrates AI capabilities through plugins and functions. When a kernel function generates output — whether from an LLM, a connector, or a chain of function calls — the result carries no provenance metadata:

• Which model generated it?
• What was the confidence level?
• What plugins/connectors were involved?
• Was the output verified or raw inference?

Why this matters

As Semantic Kernel is adopted in enterprise applications, output provenance becomes critical for:

1. Audit trails — regulated industries need to know how AI outputs were produced
2. Trust differentiation — an output sourced from a verified database connector is more trustworthy than one from unverified LLM inference
3. EU AI Act compliance (Article 50, August 2, 2026) — AI-generated content needs transparency metadata

Possible approach

FunctionResult or ChatMessageContent could carry optional provenance:

var result = await kernel.InvokeAsync("plugin", "function");

// Provenance on the result:
result.Metadata["ai_generated"] = true;
result.Metadata["model"] = "gpt-4";
result.Metadata["trust_score"] = 0.85;
result.Metadata["source_tier"] = "verified_connector";

This is already partially supported through Metadata dictionaries — the question is whether there should be a standardized schema for AI provenance fields.

Reference

• AKF defines a compact provenance schema for AI outputs
• The Metadata dictionary on SK results already provides the extension point
• This aligns with Microsoft's Responsible AI principles around transparency

Has the SK team considered standardizing provenance fields on function results?