メニュー
Grades a skill or verifier directory for required files, valid manifest structure, side-effect declarations, validation gates, paired-verifier resolution, fixtures, and authoring hygiene. Engineering verification only.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Deterministic verifier for nv_segment_ct_finetune evidence packs. Audits the finetuned checkpoint, training trajectory, and dataset audit signals already recorded by the wrapper. Engineering verification only.
Audits nv_segment_ct evidence packs for label-map readability, anatomy plausibility (organ volume bounds, fragmentation, bilateral symmetry, liver-larger-than-spleen), and optional per-class Dice/IoU against a referenced ground-truth label map. Engineering verification only.
Deterministic paired verifier for nv_generate_ct_rflow evidence packs. Audits per-sample image/mask geometry, CT-HU plausibility, mask label-set sanity, declared output-label coverage, and lung-lobe HU plausibility. CPU-only; does not re-run synthesis.
Used to verify dicom-metadata-extract evidence packs for source gate success, header fact completeness, PHI-flag consistency, and scope disclosure. Not for de-identification.
Audits dicom_series_preflight evidence packs for pass/warn/fail preflight verdicts with explicit corruption, orientation, PHI, and consistency findings.
Used to verify dicom-series-to-volume evidence packs by comparing JSON geometry evidence with the emitted NIfTI artifact. Not for clinical validation.
| name | skill-completeness-v1 |
| description | Grades a skill or verifier directory for required files, valid manifest structure, side-effect declarations, validation gates, paired-verifier resolution, fixtures, and authoring hygiene. Engineering verification only. |
| license | Apache-2.0 |
| allowed-tools | ["Bash"] |
target_skill; outputs are completeness_report.scripts/grade.py on a target skills/<name> or verifiers/<name> directory.run_script, use run_script("scripts/grade.py", args=["skills/<name>"]).| Script | Purpose | Arguments |
|---|---|---|
scripts/grade.py | Primary verifier entrypoint declared by skill_manifest.yaml. | TARGET_SKILL_OR_VERIFIER_DIR |
SKILL.md and skill_manifest.yaml to receive a meaningful structural audit.LLM_VERIFIER=1 and NV_INFER_TOKEN; leave it unset for deterministic local checks.https://inference-api.nvidia.com.| Error | Cause | Fix |
|---|---|---|
not a directory | The argument is not a skill or verifier directory. | Pass skills/<name> or verifiers/<name>. |
| Missing manifest/frontmatter | The target is not skill-shaped or has malformed metadata. | Add the required files before rerunning. |
| Advisory findings | The target is runnable but less agent-friendly. | Fix the listed SKILL.md, fixture, or manifest hygiene gap. |
Meta-verifier for spec quality.
python verifiers/skill_completeness_v1/scripts/grade.py skills/dicom-metadata-extract
python eval_engine/run.py verifiers/skill_completeness_v1 \
--fixture skills/dicom-metadata-extract \
--out runs/audit_dicom_metadata
Tier 1 checks required files, frontmatter, manifest fields, entrypoint paths,
and output schemas. Tier 2 checks side effects, gate presence, paired verifier
resolution, fixture presence, import declarations, nontrivial sanity checks,
and SKILL.md hygiene. The report also derives a non-authoritative lifecycle
status (draft, runnable, gated, verified, or published) from those
checks plus paired-verifier, trusted-run, behavior-eval, benchmark-note, and
curated-evidence signals.
Optional advisory LLM review is disabled by default. Set LLM_VERIFIER=1 and
NV_INFER_TOKEN only when explicitly running that experiment.
This verifier is itself audited by the same eval_engine.