Evaluated skills
Each report is separate; the evidence level is labelled per card, and scores from different suites are never combined.
DSH sandbox attribution
Does a skill make the agent name the sandbox layer that hid a GPU or /dev object, and cite verifiable evidence?
Art creation
Artifact quality of art-creation skill arms, with previous/candidate comparisons and original case evidence.
DSH observability
Exact percentile/tie wording in the skill plus event-log analysis. Strong on large generated traces, but one task family.
DSH plugin validation
Large plugin audit cases are useful controls; no stable correctness separation from the without arm.
DSH runtime guide
Sandbox-awareness and true-resume are controls; runtime-upgrade behavior is candidate-unstable, not a guarantee.
All reports
| Suite | Report | Records / trials | Primary result | Link |
|---|---|---|---|---|
| dsh-skills | Sandbox attribution | 11 sealed records | B evidence 6/6 (100%); A0 0/5 (0%) | dsh-sandbox-attribution/ |
| art-creation | Art creation | 38 legacy trials / 3 reports | candidate 80.5% vs previous 92.7% quality | art-creation/ |
| dsh-skills | DSH observability | legacy documented | case-specific signal | dsh-observability/ |
| dsh-skills | DSH plugin validation | legacy documented | disabled | dsh-plugin-validation/ |
| dsh-skills | DSH runtime guide | legacy documented | disabled | dsh-runtime-guide/ |
Generated from sealed records and public legacy bundles; no model calls. Source: benchmarks/evaluation/.