# EM-LND-14 — Independent adjudication ## Verdict **AIResearchLandscape — reject as new subject type; accept as a named, tenant-scoped, as-of landscape declaration plus reproducible projection.** It has no identity, state or lifecycle independent of the masters it cites; its content is exhausted by references, pins, scope, completeness and freshness policy. **AIUsageNetwork — reject as new subject type; accept as a second viewpoint (traversal/impact projection) over the same declaration.** WM-XCT-037 already owns typed edges, direction, determination method, evidence, propagation licence and completeness. A second edge plane would fork mastership, exactly as EM-LND-12 rejected a separate Service Dependency Graph. Both are governed declarations and reproducible graph projections, not subject roots. Neither may mint an artifact, run, system or dataset identifier. No identifiers allocated here. ## Evidence Precedent is consistent: EM-LND-12 made Service Landscape a view on an existing graph host and refused a rival dependency graph; EM-LND-13 made Data Landscape a projection and Analytical Lineage View a profile of WM-DAT-006. The AI subject masters are already reserved (WM-AI-001..010, WM-SFT-004, WM-DAT-001, WM-MAT-008), and WM-XCT-037 supplies the edge grammar. Nothing in the contour's three questions requires a new root: all three are traversal, applicability and quantity-kind questions over existing masters. ## Identity/mastership Every fact names one owning master. Landscape view identity is artifact identity (viewpoint, tenant, graph revision, as-of, completeness perimeter, crosswalk edition, policy version) — never business identity. Digest verifies bytes, not governed identity; equal names or aliases never prove equal weights. Titles, endpoints, scores, aliases and dates are correlation values, not keys. Retired identifiers are tombstoned, never recycled. ## Landscape/usage network The declaration binds: scope (which masters, which tenants, which phases), pinned source revisions, as-of instant, completeness perimeter, freshness policy, projection/redaction policy. The usage-network viewpoint answers "which systems, deployments, agents and runs consume which artifacts, configurations, datasets and benchmarks" by traversing WM-XCT-037 edges under their declared propagation licences, bounded by depth, phase, class and tenant. Reaching a node yields an exposure statement, not an impact verdict. Absence of an edge is no assertion unless a covering closed-world completeness declaration exists. Withheld edges are reported as withheld, never as absent. The host should be the graph host already used in EM-LND-12 (WM-XCT-039) only if its boundary generalises beyond managed IT service; that is unverified here, so the host binding is held. ## Research/data/rights Research programme and research question are unallocated; only WM-ACT-036 research study is referenced (PRODUCES WM-DAT-001, candidate). Dataset revision is WM-DAT-001 — dataset, version and distribution distinct. Benchmark version is WM-AI-009, distinct from its dataset snapshot and from the source dataset master. Rights, consent, purpose limits and export terms attach per input and per resulting artifact, and survive every derived edge: training bindings, checkpoints, adapters, quantizations, merges, registry records and endpoints. A shared endpoint merges no lineage and relaxes no restriction. ## Training/artifacts/registry Training run (WM-AI-006) is an execution aggregate: objective, authority, immutable input bindings, topology, attempts, checkpoints, metrics, outcome. Effective configuration is run-contained and pinned as one immutable digest (code revision, hyperparameters, seeds, dependencies, environment). Model artifact (WM-SFT-004) owns payload, serialization, manifest, digest, licence. Registry record (WM-AI-007) owns discoverability, registration and approvals. Checkpoint is run-contained; promotion to artifact is a separate act, and selection, release and deployment are three further separate acts. Model family and architecture are classification, not roots. ## Evaluation/benchmark/safety WM-AI-003 holds one bounded instance: applied plan registered before results, run, scoring, results, uncertainty, findings, triggers. WM-AI-009 holds the benchmark contract and release. WM-AI-008 holds one dated residual-risk judgement for a pinned system, purpose, population and context. Deployment decision is profiled, not new: rationale (WM-KNW-010), authorized act (WM-ACT-024), issued instrument (WM-REC-010). Measurement ≠ assessment ≠ authorization; a benchmark result is not universal deployment permission. ## Systems/deployments/runs AI system (WM-AI-001) is the accountable unit; agent (WM-AI-002) the durable actor; inference run (WM-AI-004) one bounded execution pinning agent and exactly one effective configuration digest. Deployment and endpoint have no frozen model in scope and remain external masters. One endpoint may serve many artifacts and one artifact many endpoints; endpoint change creates no run, no artifact revision and no new authorization. Incident is WM-AI-010 (reporting case) over WM-ACT-019; authority receipt proves neither causality nor remediation. ## Withdrawal/impact Withdraw dataset revision D2. Traverse declared bindings: training runs pinning D2, artifacts produced by those runs, evaluations citing D2 as a split, registry records for those artifacts, systems and deployments binding them. - **Proven** — pinned D2 digest in a run or evaluation binding, declared method, within a complete perimeter. - **Potential** — reachable only by derivation (adapter, quantization, merge, distillation) or by series-head rather than pinned binding, where the licence permits traversal but rights effect is unadjudicated. - **Unknown** — no covering completeness declaration, missing or stale lineage edges, unresolved references. Effect: mark rights withdrawn and applicability invalid; evidence goes stale, never deleted; assertions are superseded additively; tombstones retain identity. Each deployment decision keeps its own authority, scope and validity — withdrawal triggers reassessment, not automatic revocation. "Not affected" is publishable only with a covering completeness declaration. ## Compute/product boundary Requested compute is a run-owned plan; granted compute is an external reservation (EM-WRK-05 pool/reservation) with authority and expiry; observed consumption is WM-FLW-015 with quantity kind, unit, interval, method and uncertainty. None is inferred from another; GPU-hours, elapsed time, energy, emissions and currency are separate quantity kinds and never summed. A grant implies no product, offering, registry entry, deployment or agent identity. Research may consume compute with zero registry entries and zero commercial products; commercial product remains WM-DAT-008 / the EM-PRD-01 Offering candidate. ## Time/provenance/access Effective, observation, ingestion, approval, release and knowledge times stay distinct; views answer as-of a stated instant. Evidence is revision-bound and stales by declared rule when any pinned digest changes. Every edge records kind, nature, phase, method, evidence and confidence; inferred never becomes asserted without a new accountable assertion. Access is deny-by-default and inherits source rights; weights, protected data, checkpoints, hidden tests and confidential dossiers stay out of the projection. ## Governance Named owner for the declaration, distinct from the masters it cites. Unknown mastership blocks mutation. Cross-tenant merge requires identity evidence. Views are immutable once issued and superseded, not edited. Landscape output carries viewpoint, scope, freshness, completeness, exclusions and policy version or it is not citable. ## Acceptance scenario D2 is withdrawn. Runs R1 and R2 pinned D2 and produced artifacts A1 and A2 with distinct digests (proven). A3, quantized from A2, is potential pending rights adjudication. Evaluations E1 (D2 as test split) lose applicability; E2 (unrelated benchmark) does not. Registry records for A1–A3 are flagged, not deleted. Systems S1 and S2 each retain their own authorization instrument with separate validity; S1's expires and returns to unauthorized, S2's remains valid pending reassessment. Part of the tenant has no completeness declaration, so that segment is unknown rather than unaffected. No evidence is destroyed. ## Invariants 1. Run is not artifact; checkpoint is not release. 2. Benchmark result is not deployment permission. 3. Input rights survive every derived edge; endpoints merge nothing. 4. Research checkpoint creates no product, offering or registry entry. 5. Digest verifies bytes, not identity. 6. Landscape and usage-network views mint no subject identity. 7. Silence is no assertion; negatives need covering completeness. 8. Every run pins one immutable effective configuration digest. 9. Requested, granted and observed compute remain distinct. 10. Withdrawal invalidates applicability without deleting evidence. 11. Measurement, assessment and authorization remain separate acts. 12. Corrections are additive; issued records are immutable. ## Minimal model set WM-AI-001, 002, 003, 004, 005, 006, 007, 008, 009, 010; WM-SFT-004; WM-DAT-001; WM-MAT-008; WM-XCT-037; plus referenced externals for research study, decision triad, compute pool/flow, deployment/endpoint and product. Add one landscape-view element on an existing graph host; add no AI-specific graph, dependency or usage root. ## Holds Missing roots: research programme, research question, deployment/endpoint, metric definition, training recipe, reusable evaluation protocol, standing authorization instrument — all unallocated. Contradictions: WM-AI-007 names WM-SFT-004 as parent with no frozen edge; WM-AI-009→WM-DAT-001 unapproved; WM-AI-006 PRODUCES WM-SFT-004 candidate only, so the lineage edge the withdrawal test depends on is not frozen; WM-XCT-037 is registered as mixin, reclassified relationship, and its scope forbids traversal while its own functions traverse — the usage-network viewpoint cannot be relied on until that is resolved; WM-XCT-037 also carries an unresolved strong-versus-weak edge identity conflict and empty composition contracts; WM-AI-003/WM-AI-009 overlap on protocol and metric ownership; registry entry_kind diverges from subject kind for WM-AI-001 and WM-MAT-008. Assurance: WM-AI-006, 007, 009, 010, WM-MAT-008 and WM-XCT-037 are single-provider drafts; all sources unverified live; no fixtures run. Host binding to the EM-LND-12 graph host is unverified. Nothing here claims canonical completeness, installability or publication readiness.