The full branch of the Index category tree, from the L1 dimension (code 110) down to canonical L4 indicators — the units the ethicometric challenge must learn to measure:

110 Model Design & Development

110.110 Architecture & Objectives — clearly defines intended use and excludes unsafe applications

Intended use and out-of-scope disclosures are documented

Safety-critical use cases explicitly disallowed in system policies

110.120 Interpretable Design — provides understandable explanations for its outputs

Mechanisms enabling global interpretability are documented

Support for local explanations is available via API or UI

110.130 Efficiency & Environmental Considerations — minimizes environmental impact and discloses efficiency trade-offs

Energy and carbon footprint of training and inference is estimated and disclosed

Efficiency optimizations (e.g., quantization, distillation) documented with trade-offs

110.140 Data Development — uses datasets with clear sources and licensing

Training data sources and licenses disclosed at an aggregate level

Third-party datasets include SPDX or clear license metadata

110.150 Dataset Quality & Representativeness — reflects diverse populations and documents known data gaps

Coverage across demographics or locales is characterized (no raw protected attributes required)

Known gaps and limitations documented with mitigation plan

110.160 Data Documentation — maintains versioned datasheets and dataset lineage

Internal datasheets and data statements are versioned

Lineage and versions of major datasets tracked

110.170 Alignment & Training — trained with clear annotator guidance and quality checks

Annotator guidelines and instructions summarized

QA of label quality and diversity reported

110.180 Safety Alignment & Value Modeling — aligns with safety goals and updates based on red-teaming

Safety objectives and value trade-offs are articulated

Red-teaming results inform alignment updates

110.190 Fine-tuning Controls — supports safe fine-tuning with documented constraints and rollback

Domain-specific fine-tunes documented with safety constraints

Model update and rollback mechanisms exist