# EM-LND-14 provider comparison

Claude and Grok agree that AI Research Landscape is a tenant-scoped declaration plus reproducible projection and AI Usage Network is an impact/traversal viewpoint over the same declaration and WM-XCT-037 edges. Neither is a root or edge master, and no model or runtime identifier is added.

Both preserve separate identity and mastership for research, datasets, revisions, training runs, configurations, artifacts, registry records, evaluations, benchmarks, safety assessments, systems, deployments, inference runs, incidents, compute and products. A projection is not an artifact; a run is not an artifact; benchmark success is not deployment permission; a checkpoint creates no product.

Both require source rights to survive every derived edge and viewpoint access to be no broader than the intersection of source rights. Tenant isolation, pinned as-of inputs and complete edge snapshots are required. Missing edges never prove absence.

Grok sharpened the withdrawal classification. Recorded consume, derive or basis edges establish proven participation, including mixed-input artifacts with sole-cause false. Evaluation and safety records retain identity but have proven applicability impact. Systems and endpoints are potential operational impact until current binding and effect are established. Missing or incomplete edges, unofficial copies and undeclared fine-tunes are unknown. Deployment decisions remain separate and are flagged for review, not automatically reversed.

Missing or unproven dependencies remain outside EM-LND-14: Research Programme/Question, Dataset Revision, Deployment Decision, Deployment/Endpoint, Compute, Commercial Product and a Rights/Licence Grant. The reconciled result remains a held declarative profile pending edge-model correction and canonical dependencies.

## Frozen-audit reconciliation

The single frozen audit accepted the profile with eight limits. The final candidate separates proven participation/applicability from operational harm, keeps impact sets re-derivable, gives scope to one declaration, prevents derived-edge promotion, closes inference-level rights leakage, forbids an unaffected class, uses timed overlays and requires bounded review of potential deployment impact. No second audit was run and no identifier was added.