Omniverse Realtime Viewer

The document explicitly requires ovrtx, preserving user USD files, separating session state, and avoiding browser-side renderers; it also provides some error fallbacks and state constraints, supporting partial data-flow transparency and least privilege. Points are deducted because Shell and Write are declared without explicit user confirmation, permission boundaries, rollback procedures, or sensitive-credential handling. The cloud-asset references also show credentialed and external object-storage operations without stronger isolation or authorization guidance.

Routing, read order, non-negotiable rules, and extensive abnormal-input guidance make the main happy paths reasonably intelligible, with failure feedback for AOV, input, and camera flows. Points are deducted because this is a static review and cannot establish dependency availability, key-path reproduction, or executable correctness; the BENCHMARK reports an overall FAIL, no available Tier 3 evaluation, and many duplicate-content findings plus 32 Tier 1 findings, so the static ceiling applies.

The skill clearly targets Omniverse Realtime Viewer work and defines delivery paths, input, camera, AOV, UI, and automation scenarios; the eval cases cover browser, desktop, Electron, Tauri, and headless requests. Points are deducted because GPU, ovrtx, ovstream, and runtime requirements are substantial, non-fit boundaries and compatibility conditions are distributed across downstream references, and there is no Chinese-language or mainland-China network adaptation guidance. Some cloud-asset workflows also depend on external S3 services.

SKILL.md provides clear routing, read order, reference families, workflow, and completion checklists; metadata includes version, Apache-2.0 licensing, author, and tags, while the repository README describes synchronization and update mechanisms. Points are deducted because maintenance ownership is directed generally to product repositories without a specific maintainer, changelog, or stable-version policy for this skill; the BENCHMARK reports repeated content that increases maintenance and progressive-disclosure burden.

The objective, architecture constraints, reference selection, and representative eval expectations are concrete enough to plausibly help an agent build an RTX USD viewer, while explicitly prohibiting browser-side USD rendering and source-USD mutation. Points are deducted because the supplied files contain no verifiable representative generated outputs or third-party execution evidence; the BENCHMARK has no Tier 3 results and an overall FAIL, so only the static ceiling for core-task plausibility is justified.

The fixed revision, committed evals, validation reference, and BENCHMARK provide limited audit trails; the report also discloses unavailable Tier 3 evidence, static findings, and deduplication findings. Points are deducted because there are no execution logs, complete key-path test suites, CI coverage, or independent corroboration, and important evaluation signals are missing, preventing a higher static verifiability score.