Proposal: Portable Evidence Artifact Layer (.epi) for Guardrail Decisions

[mohdibrahimaiml]

Problem

NeMo Guardrails provides strong runtime control over LLM behavior (blocking, redirecting, enforcing policies).

However, the resulting decisions are only observable via logs or internal state.

There is no portable, self-contained artifact that captures:

• what input was received
• what guardrails triggered
• what was blocked/allowed
• why the decision was made

This makes external auditing, compliance validation, and offline review difficult.

---

Proposal

Introduce a portable artifact layer (e.g. .epi) that records guardrail decisions as a structured, replayable file.

Each .epi file would contain:

• input prompt
• guardrail policies evaluated
• decision trace (allowed / blocked / modified)
• final output
• metadata (timestamps, config, versions)

---

Why this matters

• Enables auditability (compliance, governance)
• Allows offline verification of guardrail behavior
• Makes decisions portable across systems
• Bridges runtime control → verifiable evidence

---

Integration Approach (Non-invasive)

This can be implemented as:

• optional export layer
• hook into guardrail evaluation pipeline
• no change to core decision logic

Example:

NeMo execution → decision → .epi artifact emitted

---

Prior art / Prototype

I’ve implemented a working prototype of this idea as:

EPI Recorder (GitHub link)

It records LLM interactions into a portable artifact format.

This proposal extends that concept specifically to guardrail decision systems.

---

Open Questions

• Would an artifact export layer align with NeMo’s roadmap?
• Is there an existing effort toward auditability/export that this could integrate with?
• Preferred format: JSON-based vs domain-specific schema?

---

Happy to contribute a PR or prototype integration if this direction makes sense.