Capability Driven Taxonomy

• Capability-driven taxonomy is a formal, multi-dimensional framework that defines and organizes systems by measurable capabilities, dispositions, and functions.
• It decomposes system functions into explicit axes, such as ODD, SAE, ARL in automated driving, and multi-dimensional constructs in agentic SDLs and AI safety zoning.
• This framework enables standardized evaluation, gap identification, and ontological integration across heterogeneous domains for improved governance and performance.

A capability-driven taxonomy is a formal multi-dimensional framework for classifying, comparing, and engineering systems according to their measurable abilities, dispositions, and operational parameters. Underlying such taxonomies is the systematic decomposition of system functions into explicit capability axes—whether in autonomous driving, agentic laboratories, AI safety protocols, or capability ontology engineering. These frameworks support rigorous cross-domain analysis, gap identification, standardized evaluation, and ontological integration of heterogeneous capabilities. Key references include Betz et al. (automated driving) (Betz et al., 2024), Chen et al. (SDLs) (Chen et al., 25 Jan 2026), capability ontologies (Beverley et al., 2024, Silva et al., 2024), and zoning-based AI safety (Pistillo et al., 2024).

1. Formal Ontological Foundations

The ontological core of a capability-driven taxonomy is the distinction between dispositions, capabilities, and functions, as formalized in “Capabilities: An Ontology” (Beverley et al., 2024). In Basic Formal Ontology (BFO):

• Dispositions are internally grounded properties of material entities, realized in processes, and tied to physical changes in the bearer.
• Capabilities are dispositions whose realizations are of interest to some organism or group, formally:

$$\text{Capability}(x) \iff \text{Disposition}(x) \land \exists o,p\,(\text{OrganismOrGroup}(o) \land \text{Process}(p) \land \text{Realizes}(p,x) \land \text{hasInterestIn}(o,p))$$

• Functions are a subclass of capability: historically selected or designed dispositions in artifacts or organs.

This structure supports mapping any capability by bearer (individual, aggregate, artifact), grounding (intrinsic, extrinsic, historical), domain (physical, cognitive, social), and acquisition (innate, trained, emergent).

2. Multi-Dimensional Classification Schemes

A capability-driven taxonomy formalizes and organizes capability spaces along explicit axes, tailored to the domain:

Automated Driving Systems (Betz et al., 2024)

• Operational Design Domain (ODD): A tuple $(c | u | r | e | v | a)$ with country, road users, road types, environmental conditions, velocity bands, and additional requirements.
• SAE Automation Level: Discrete levels $0 \ldots 5$ characterizing responsibility from human-driven to fully autonomous.
• Automation Readiness Level (ARL): Nine maturity stages from concept (ARL 1-2) to commercial deployment (ARL 9).
• Taxonomy space: $$\mathrm{Tax} = \mathrm{ODD} \times \mathrm{SAE} \times \mathrm{ARL}$$.

Agentic SDLs (Chen et al., 25 Jan 2026)

• Six Axes: Decision Horizon, Uncertainty Modeling, Action Parameterization, Constraint Handling, Failure Recovery, Human Involvement.
• Systems mapped as points in the six-dimensional capability space.

AI Safety Zoning (Pistillo et al., 2024)

• Dangerous Capability Zones: Four ordinal zones (Green, Yellow, Orange, Red) representing proximity to high-impact capabilities, defined by distance metric $d(c, C^*)$ and thresholds $\tau_1 > \tau_2 > \tau_3$.

Capability Ontologies (Silva et al., 2024)

• Capability hierarchies encoded in OWL DL, with formal relations (hasInput, hasOutput, hasPrecondition) and SHACL-based verification of logical and relational completeness.

3. Mapping, Embedding, and Evaluation Criteria

Capabilities are embedded into taxonomy spaces via explicit classification rules:

• Mapping functions assign system $f$ to $$(\mathrm{ODD}(f), \mathrm{SAE}(f), \mathrm{ARL}(f))$$ in automated driving (Betz et al., 2024).
• In SDLs, each dimension is quantified (e.g., myopic vs. long-horizon decision, homoscedastic vs. heteroscedastic UQ) and the system located accordingly (Chen et al., 25 Jan 2026).
• Zoning taxonomy assigns capability profiles to zones using a configured $d(c, C^*)$ and evaluation suite $\{T_j, \theta_j\}$ (Pistillo et al., 2024).
• Capability ontologies are verified via syntax checks, logic reasoners, and relation completeness (CR metric) (Silva et al., 2024).
• For each candidate capability in BFO, classification protocol: bearer identification, grounding analysis, process mapping, stakeholder linkage, acquisition history (Beverley et al., 2024).

4. Structural Relationships, Hierarchies, and Ontology Integration

Capability-driven taxonomies establish hierarchical and relational structure through partial orders, subclass relations, and ontological constraints:

• ODD attributes in automated driving are partially ordered (e.g., $H \subset H^{+} \subset *$ for road types; $v_0 \subset v_1 \ldots \subset *$ for velocities), yielding nested hierarchies (Betz et al., 2024).
• Capability ontologies construct trees and DAGs of capabilities, allowing inheritance, disjointness, and compositional relations as shown in Turtle/OWL (Silva et al., 2024).
• BFO taxonomy: Disposition $\supset$ Capability $\supset$ Function, with subaxes for bearer, domain, stability, etc. (Beverley et al., 2024).
• SDL capability dimensions are continuous, not strictly hierarchical, enabling multidimensional placement of lab agents (Chen et al., 25 Jan 2026).

This structure supports integration of heterogeneous capability datasets, cross-domain queries, and standardized analytics (e.g., manufacturing “elasticity,” defense “strategic mobility,” clinical “responsiveness”) (Beverley et al., 2024).

5. Worked Examples and Practical Templates

Empirical mapping and templates demonstrate the taxonomy’s function:

Automated Driving:

• Example: Waymo Robotaxi in SF, mapped as

$$f_{taxi} \mapsto ((US,*,\{H^{+},U\},\{N,R\},v_3,\{onlySF\}),4,9)$$

• Example table (excerpted):

System/Function	ODD	SAE	ARL
Highway-pilot truck	(US, *, H+, N∧R, v4, ∅)	4	6
Waymo Robotaxi	(US, *, H+∪U, N∧R, v3, onlySF)	4	9
Mercedes DrivePilot	(DE+US, *, H, L∧D, v3, vehicleAhead,noGlare)	3	9

SDLs:

SDL Example	Decision Horizon	Uncertainty Modeling	Human Involvement
Hydrogel BO system	One-step	Homoscedastic GP	None
Nanoparticle multi-stage platform	Long-horizon	Heteroscedastic/Ensemble	Mixed-initiative

Zoning Taxonomy (AI safety):

• Template:
  1. Define target $C^*$
  2. Decompose into precursory skills $P_1,\ldots,P_3$
  3. Construct $T_j$ and thresholds
  4. Compute $d(c,C^*)$, assign zone $Z(c)$, trigger disclosure if $I_p(c)$ flips for $p \in P_2 \cup P_3$
  5. Share/report as per zone (Pistillo et al., 2024)

Capability Ontology (LLM4Cap):

• Prompt engineering, OWL DL output, SHACL verification, human review (Silva et al., 2024).

6. Capability Comparison, Gap Analysis, and Benchmarking

Capability-driven taxonomies are used for systematic comparison, benchmarking, and identification of gaps (“white fields”):

• In automated driving, unpopulated ODD-SAE-ARL cells (e.g., SAE 4 urban service in Germany at night) highlight research and regulatory opportunities (Betz et al., 2024).
• SDL taxonomy supports comparison by axis—sample efficiency, constraint violation rate, robustness to drift—across benchmarks (e.g., LCST transition, mechanical tradeoffs, actuation with degradation) (Chen et al., 25 Jan 2026).
• Zoning taxonomy mandates tracking progress toward high-impact (“red line”) capabilities via early precursors, with escalating disclosure and security protocols (Pistillo et al., 2024).
• Capability ontologies enable automated, verifiable generation and maintenance of large capability trees from natural language, with downstream usability for planners/execution engines (Silva et al., 2024).
• BFO-centric taxonomies unify siloed capability data, supporting queries and analytics across medicine, manufacturing, defense, and education (Beverley et al., 2024).

7. Governance, Safety, and Future Directions

Capability-driven taxonomies underpin governance models (especially in AI safety) and inform both regulatory processes and technical roadmaps:

• Zoning taxonomy for pre-deployment AI capability sharing: defines routes for early reporting, role of AI Safety Institutes, risk-proportionate information security, and international exchange protocols (Pistillo et al., 2024).
• Regulatory bodies and developers can target gaps by designing experiments or extending the maturity (ARL) of systems (Betz et al., 2024).
• Taxonomies provide infrastructure for reproducibility, provenance, and auditable policy updates (e.g., structured tool calls, logging, mixed-initiative operation) (Chen et al., 25 Jan 2026).
• Automated capability ontology generation reduces error and manual overhead, increasing robustness and interoperability of capability data (Silva et al., 2024).

These frameworks are foundational for systematic assessment, safe deployment, compliance, and cross-domain interoperability of complex systems, from automated vehicles and laboratories to AI governance and federated ontology engineering.