Lifecycle Taxonomy: Structures & Applications

• Lifecycle Taxonomy is a formal, hierarchical structuring of distinct phases that defines a system’s evolution from inception to decommissioning.
• It facilitates method reuse, cross-phase optimization, and clear stakeholder mapping through standardized interfaces and transformation rules.
• Its applications across AI, quantum software, privacy engineering, and more enable systematic performance, security, and risk analysis.

A lifecycle taxonomy is a formal, hierarchical structuring of the distinct stages, activities, or transformations through which a system, artifact, or resource passes from initial conception to eventual decommissioning or evolution. Across domains such as machine learning, quantum software, AI infrastructure, agent skill management, privacy engineering, and software delivery, a lifecycle taxonomy provides a scaffold for analysis, method composition, stakeholder alignment, and cross-phase optimization. Recent works synthesize and formalize comprehensive lifecycle taxonomies to address fragmentation, aid reproducibility, and enable system-level reasoning about performance, security, privacy, and governance.

1. Lifecycle Taxonomy: Conceptual Foundations and Formalism

A lifecycle taxonomy consists of an ordered set (often hierarchical) of phases, stages, or components, each with formally defined responsibilities and interfaces. Formally, one typical representation is as an ordered tuple or set $\mathcal{L} = \langle \ell_1,\dots,\ell_n \rangle$, where each $\ell_i$ is a lifecycle phase (e.g., Initialization, Training, Deployment) (Miraghaei et al., 9 Jun 2025, He, 26 Nov 2025, Xia et al., 2024, Weder et al., 2021). Some frameworks generalize this by allowing for main, optional, and cross-cutting components, as in $L = (M, O, X)$ with Main $M$, Optional $O$, and Cross-cutting $X$ components (e.g., the SLM lifecycle) (Miraghaei et al., 9 Jun 2025). Each stage can be realized as a transformation $T_i : \Theta \to \Theta$ (parameter manifold), and cross-cutting components may inject regularizers, constraints, or auxiliary transformations (e.g., $R_j: \Theta \to \mathbb{R}$ or $U_j: \Theta \to \Theta$).

Lifecycle taxonomies provide a basis for several core functions:

• Compositionality: Method reuse and coordinated adaptation across pipeline stages.
• Interface specification: Standardizing input-output contracts between stages.
• Metric propagation: Formal modeling of interdependencies among metrics and outcomes across the stack.
• Stakeholder mapping: Assigning responsibility and accountability at each phase for complex supply chains (Xia et al., 2024).
• Risk and compliance tracking: Systematic placement of controls and audits.

2. Representative Lifecycle Taxonomy Structures in Technical Domains

A. Small LLMs (SLMs)

The SLM lifecycle is modularized as $L = (M, O, X)$:

• Main Components ($\ell_i$0–$\ell_i$1): Initialization, Distillation, General Lifecycle (includes SFT, RLHF, inference-time transforms), Quantization, Deployment.
• Optional Components ($\ell_i$2, $\ell_i$3): On-Device Learning, Federated Learning.
• Cross-cutting ($\ell_i$4–$\ell_i$5): Data Selection, Evaluation, Efficient Fine-Tuning, Inference Optimization.

Main stages are realized as sequential transformations $\ell_i$6, while cross-cutting and optional components may operate as auxiliary transforms or constraints at one or more main stages (Miraghaei et al., 9 Jun 2025).

B. AI Infrastructure

A six-layer $\ell_i$7 three-domain taxonomy (18 cells) spans the entire AI infrastructure stack (He, 26 Nov 2025):

Layer (i)	Physical (j=1)	Compute (j=2)	Economic (j=3)
1. Grid & Sustainability	PUE, MEF, WUE, CUE	Shiftable Workload Fraction	LMP, Grid Cost
2. Facility	Cooling Power Ratio	Throttling Limits	Cooling OpEx
3. Compute Hardware	TDP, FLOPs/W	Accelerator Utilization	Depreciation Curve, Cost/hr
4. Networking	Network Power	AllReduce Latency	Network CapEx
5. ML Runtime	Idle Energy	MLPerf Throughput	Cost per Training Step
6. Service & Operations	Redundancy Energy	SLI/SLO, MTBF, MTTR	TCO, Lifecycle ROI

Each metric may propagate (formally as a graph $\ell_i$8 of nodes $\ell_i$9 and propagation functions $L = (M, O, X)$0), enabling cost, carbon, and efficacy to be jointly optimized across full asset lifecycles (He, 26 Nov 2025).

C. Hybrid Quantum Applications

Here, the enclosing lifecycle comprises eight phases: Requirements, Quantum-Classical Splitting, Architecture & Design, Implementation (forks into Quantum, Classical, Workflow sub-lifecycles), Testing, Deployment, Observability, Analysis (Weder et al., 2021). Each sub-lifecycle (Quantum Circuit, Classical Software, Workflow, Operations) has its own artifacts and processes, joined at formally defined interfaces and points of integration.

D. Agent Skill Management

The lifecycle of a reusable procedural "agent skill" is structured as:

1. Representation: $L = (M, O, X)$1, with main doc, resources, applicability condition.
2. Acquisition: Human-derived, experience-driven, task, or corpus-based.
3. Retrieval & Selection: Embedding/sparse/hierarchical retrieval; policy-based selection.
4. Evolution: Revision, validation, policy coupling, repository sync, runtime governance.

Transitions between stages are governed by formal eligibility and revision criteria (e.g., success/failure triggers, validation suites) (Zhou et al., 8 May 2026).

E. Privacy Lifecycle in AI and Health Data

AI privacy threat modeling (PriMod4AI) taxonomy: Six sequential phases—Data Collection, Processing, Model Building/Training, Deployment, Inference, Continuous Monitoring—anchored to threat classes from both classical (e.g., LINDDUN) and model-centric (e.g., membership inference, gradient leakage) sources; mapping threat categories onto stages enables full coverage (Savaliya et al., 4 Feb 2026).

Personal health data lifecycles distinguish: Creation, Storage, Access, Sharing, Linking, Learning, Destruction—with taxonomy matrices mapping privacy threats (e.g., impersonation, eavesdropping, manipulation) and applicable technical controls to stages (Bose et al., 2023).

F. Software Delivery and Operations (Everything-as-Code)

The Everything-as-Code (EaC) taxonomy consists of 25 practices grouped into six lifecycle-aligned layers: Infrastructure Provisioning, Platform/Orchestration, Application Design, Data/Database, Security/Compliance, Observability/Analysis. This structure binds practices to SDLC and DevOps phases, supporting controlled code-based management throughout operational lifecycles (Wei et al., 7 Jul 2025).

3. Lifecycle Taxonomy as a Basis for Method Reuse, Co-Adaptation, and Cross-Stage Optimization

Lifecycle taxonomies serve as canonical maps for method co-adaptation and cross-stage optimization:

• Explicit dependency matrices (e.g., $L = (M, O, X)$2 if $L = (M, O, X)$3 influences $L = (M, O, X)$4) guide safe method reuse (e.g., LoRA factorization spanning SLM initialization, PEFT, and inference-time KV cache compression) (Miraghaei et al., 9 Jun 2025).
• Multi-objective optimization is facilitated by propagation graphs (MPG) that formalize how changes in one metric (e.g., grid carbon intensity) cascade through infrastructure or model lifecycles to final cost or risk (He, 26 Nov 2025).
• Seamless method chaining: E.g., quantization-aware fine-tuning must constrain both the training and quantization stages; federated distillation reuses the classical loss at the edge (Miraghaei et al., 9 Jun 2025).

4. Formal Taxonomy Notation and Metric Integration

Lifecycle taxonomies admit formal hierarchical, set-theoretic, and functional definitions:

• SLM lifecycle: $L = (M, O, X)$5, $L = (M, O, X)$6, $L = (M, O, X)$7 (Miraghaei et al., 9 Jun 2025).
• Metric matrix for AI infrastructure: $L = (M, O, X)$8, $L = (M, O, X)$9, with $M$0 as the propagation matrix (He, 26 Nov 2025).
• Agent skill selection: $M$1, with eligibility driven by applicability condition $M$2 (Zhou et al., 8 May 2026).
• Privacy classification: $M$3, a ternary relation mapping each lifecycle stage $M$4, threat $M$5, and technique $M$6 (Bose et al., 2023).

These formalizations are essential for toolchain automation, provenance capture, and lifecycle-aware hyperparameter search.

5. Practical Implications and Theory-to-Practice Impact

Lifecycle taxonomies enable:

• Tooling and automation: Lifecycle-aware orchestration engines, SLMOps systems, and code-as-pipeline infrastructures (Miraghaei et al., 9 Jun 2025, Wei et al., 7 Jul 2025).
• Unified benchmarking: Composite metrics (e.g., PUE-adjusted cost per training step, carbon-normalized throughput) with provable interpretability (He, 26 Nov 2025).
• Adaptive risk management and governance: Placement of secure data factories, supply chain controls, runtime governance, and audit logging within a single reference architecture (Rashid et al., 26 Feb 2026, Xia et al., 2024).
• Systematic identification of research gaps: E.g., automotive data taxonomy reveals under-representation of requirements engineering data (Hohl et al., 1 Oct 2025).

Further, lifecycle taxonomies foster end-to-end traceability, reproducibility (via provenance-capture schemas, e.g., W3C PROV), and closure of DevOps or model management loops.

6. Limitations and Open Research Challenges

Current open problems—surfaced by domain-specific taxonomies—include:

• Cross-stage abstraction leakage and method drift (protecting skill acquisition mechanisms from resource drift, ensuring co-evolution does not silently break governance or safety) (Zhou et al., 8 May 2026).
• Automated, lifecycle-spanning optimization and benchmarking: Jointly tuning parameters (e.g., quantization bit-width ↔ adapter rank) over the full pipeline (Miraghaei et al., 9 Jun 2025).
• Standardization and interoperability of taxonomies: Harmonizing definitions across domains for tooling, compliance, and knowledge sharing (Xia et al., 2024, Rashid et al., 26 Feb 2026).
• Accurate mapping of emerging threat classes (e.g., AI-specific privacy attacks not captured in classical frameworks) onto existing or extended lifecycle taxonomies (Savaliya et al., 4 Feb 2026, Kawamoto et al., 2023).
• Lifecycle-aware system scaling: Ensuring that methods designed for one stage remain robust under scale, failure, or adversarial load across all dependent stages (Souza et al., 2020, He, 26 Nov 2025).

7. Summary Table: Lifecycle Taxonomy Illustrative Structures

Domain / Paper	Taxonomy Structure	Formalization
SLM Lifecycle (Miraghaei et al., 9 Jun 2025)	Main/Optional/Cross-cutting Stages	$M$7
AI Infra. (He, 26 Nov 2025)	6 Layers $M$8 3 Domains (18 cells)	$M$9, MPG Graph
Quantum SW (Weder et al., 2021)	8 enclosing phases + 3–4 sub-lifecycles	Fork/join structure, artifact binding
Agent Skills (Zhou et al., 8 May 2026)	Representation, Acquisition, Retrieval, Evolution	$O$0, skill retrieval, selection
Privacy (Health/AI) (Bose et al., 2023, Savaliya et al., 4 Feb 2026)	6–7 lifecycle stages; matrix threat/technique mapping	$O$1
Software Delivery (EaC) (Wei et al., 7 Jul 2025)	6 layers, 25 code practices	Adjacency matrices, SDLC alignment

Lifecycle taxonomies constitute a fundamental organizing principle for modern computational, AI, and cyber-physical systems, enabling rigorous definition, systematic process improvement, and composable, auditable, and efficiently orchestrated engineering (Miraghaei et al., 9 Jun 2025, He, 26 Nov 2025, Weder et al., 2021, Zhou et al., 8 May 2026, Bose et al., 2023, Wei et al., 7 Jul 2025, Hohl et al., 1 Oct 2025, Xia et al., 2024, Savaliya et al., 4 Feb 2026, Rashid et al., 26 Feb 2026, Souza et al., 2020, Kawamoto et al., 2023, Qian et al., 2019).