Jetson Memory Optimizer

The skill scopes actions, forbids setting SWIOTLB to 0, and specifies verification and rollback; however, it modifies BCTs, DTBs, boot targets, and potentially flashed partitions without a unified pre-execution confirmation, automatic backup, or complete recovery procedure, so points are deducted. No credential handling or covert exfiltration is shown. NVIDIA's official provenance supports attribution and governance only.

The four-layer ordering, compile/decompile checks, diagnostics, and rollback paths are detailed and broadly consistent; however, execution depends on the BSP, dtc, gcc, chip-specific sources, an external target-platform-contract.md, and index discovery not supplied here. There are no skill-specific executable scripts or key-path tests, so the static score is capped at 10 and reduced for uncertainty.

The headless, no-camera, and swiotlb scenarios, T234/T264 differences, and refusal boundaries are clear, and the skill explicitly excludes frequency tuning; however, input/output contracts, negative trigger conditions, and broader Jetson/BSP version boundaries are incomplete. The material provides no Chinese-language support guidance, so points are deducted. The core function is local-BSP based rather than dependent on overseas services, so no mainland-network deduction is applied.

SKILL.md provides frontmatter, versioning, an argument hint, prerequisites, limitations, troubleshooting, and verification sections, while the skill card identifies NVIDIA as owner; however, recommended Instructions and Examples sections are absent, version 0.0.1 has no changelog or explicit maintenance update path, and the supplied license metadata is NOASSERTION while the files state both Apache-2.0 and CC-BY-4.0 AND Apache 2.0. Points are deducted for these governance inconsistencies.

The four-layer workflow, chip-specific carveout lists, verification commands, and rollback guidance can address the core memory-reclamation task, and the benchmark summary provides limited auditable evaluation evidence; however, there are no representative final outputs, quantified memory savings, or complete BSP-compatibility proof. Important results still require on-device judgment and reflashing, so the static score is capped at 7 and reduced.

The material gives concrete paths, fields, commands, checks, and two evaluation tasks with expected behaviors, providing some traceability; however, the benchmark contains only summary and single-attempt results, with no committed test suite, logs, or independent reproduction evidence, and key dependency files are absent. The score is therefore reduced and constrained by the static cap of 5.