TAO Workflow Launch Gate

The skill explicitly gates platform, credentials, image, datasets, compute, dependencies, access checks, and a final launch review with user confirmation. It limits credential handling to presence checks and forbids printing secrets, while declaring Read/Bash access. Deducted 4 points because Bash and remote platforms still enable broad external effects, no general rollback or job-cancellation procedure is defined, and data-flow and sensitive-data handling remain incomplete.

The documentation gives coherent preflight, retry-after-remediation, dependency, SLURM, Docker, remote-storage, and terminal-monitoring rules. Deducted 2 points because helper scripts, model skills, and platform implementations are outside the selected skill directory, so static evidence cannot establish availability, interface consistency, or runnable key paths; there is no skill-specific test suite.

Triggers cover TAO training, evaluation, inference, export, TensorRT, AutoML, and DEFT, with explicit platform selection, root-mode/direct-spec dataset inputs, and monitoring choices. Deducted 3 points because non-fit boundaries are not stated, Chinese-language support is unspecified, mainland-China reachability is unverified, and core platform/object-storage access may depend on environment-specific services.

The skill has a name, description, Apache-2.0 license, author, version 0.1.0, tags, compatibility note, and reasonably layered documentation; repository material also indicates synchronization and update context. Deducted 6 points for missing recommended Instructions and Examples sections, no FAQ or changelog, insufficiently explicit maintenance ownership/update path, and inconsistency between the skill-card author format and top-level license metadata.

The documentation describes an end-to-end launch flow from platform selection and credential collection through image confirmation, dataset validation, resource preflight, user confirmation, and monitoring; the benchmark report records one passing external evaluation task. Deducted 1 point because there is only one evaluation task and no static verification of real launch results, output artifacts, or cross-platform coverage; post-launch result formats and acceptance criteria are underspecified.

Committed evals.json, BENCHMARK.md, version metadata, benchmark metrics, and static-validation findings provide limited auditable evidence. Deducted 1 point because evaluation covered one task with one attempt per agent, the skill has no committed tests covering key platform paths, real behavior cannot be independently reproduced from static files, and benchmark conclusions are primarily report evidence.