Resubmit NVIDIA NIM on K8s for CNCF AI Conformance under updated layered-product guidance

[yuanchen8911]

Summary

Resubmit the NVIDIA NIM on Kubernetes CNCF AI Conformance PR, aligned to the updated layered-product guidance merged in cncf/k8s-ai-conformance#95.

Context

• Prior PR: cncf/k8s-ai-conformance#79 — "NVIDIA NIM on EKS AI Conformance for v1.35" (merged 2026-04-21). The PR was reworked from an AICR submission to a NIM-on-EKS submission after WG feedback (Rita Zhang, Jack Francis) that conformance certifies a product on a Kubernetes platform, not standalone tooling.
• New guidance: cncf/k8s-ai-conformance#95 — "Update guidance for layered product seeking conformance" (merged 2026-04-23, two days after Harden workflows for OpenSSF scorecard #79). It formalizes how a product layered on an existing Kubernetes Conformant distribution may pursue AI Conformance, and adds explicit per-requirement attribution rules.

The new guidance landed after #79 was merged, so the existing submission predates the formal rules and should be resubmitted (or extended) to satisfy them.

What the new guidance requires

From the updated faq.md:

1. Per-requirement layer ownership. For each MUST/SHOULD feature, the submission must identify which layer owns it (base K8s/infrastructure vs. layered runtime/add-ons) and supply a reference, evidence link, or "doesn't apply".
2. Reference the base distribution's Kubernetes Conformance entry (i.e., the existing K8s Conformance certification of the underlying distribution) when applicable.
3. Re-test for full AI Conformance. Layered components can change AI-conformant behaviors not exercised by the base platform's standard K8s Conformance tests.
4. Unique contribution. The layered product must contribute at least one of: a MUST replacement, a meaningfully different variant of an existing MUST/SHOULD, a SHOULD addition, or a roadmap feature.
5. Distinguishing language must describe what the layered product contributes, not what the base lacks.

Proposed scope of work

• Re-frame the submission per the new layered-product taxonomy: per-requirement table identifying owner layer (EKS base vs. NIM-on-EKS layer) and contribution kind (replacement / variant / SHOULD addition / roadmap).
• Audit the 9 MUST evidence items from Harden workflows for OpenSSF scorecard #79 against the new per-layer attribution rules; update evidence captions/descriptions to make ownership explicit.
• Add the unique-contribution rationale (KAI Scheduler vs. Volcano, NIM Operator as a robust controller variant, etc.) using contribution-focused wording.
• Re-run evidence collection on a current EKS cluster to satisfy the "re-test for full AI Conformance" requirement.
• Open the follow-up submissions previously called out in Harden workflows for OpenSSF scorecard #79's "Next steps": NIM on GKE, AKS, OKE — each under the same layered-product framing.

References

• New guidance: Update guidance for layered product seeking conformance cncf/k8s-ai-conformance#95
• Prior submission: NVIDIA NIM on EKS AI Conformance for v1.35 cncf/k8s-ai-conformance#79
• Related issue: feat: add NIM Operator recipe for CNCF AI Conformance certification #473 (feat: add NIM Operator recipe for CNCF AI Conformance certification)
• Evidence framework: docs/conformance/cncf/

[yuanchen8911]

Submitted upstream: cncf/k8s-ai-conformance#105