AI foundations, data and deployed-system lifecycle
Identify the actual system
An AI-enabled application includes more than a model. It can combine training/inference data, preprocessing, retrieval, prompts, orchestration, tools, identity, infrastructure and downstream consumers. A large language model generates token-based outputs under its learned behavior; it is not a guaranteed database of verified facts. Tokenization identifies model input units and representations, while embeddings serve different representation purposes. Not every AI system uses a transformer or generates natural language.
Follow the data and model lifecycle
Define the intended task and affected users before choosing a model. Data collection, labeling, preparation, training, tuning, evaluation, deployment, monitoring and retirement create different risks. A dataset can be validly formatted yet contain incorrect, biased, sensitive or inappropriately licensed material. Record provenance, applicable permissions and quality/coverage limits. Training-data poisoning and manipulation of retrieval or later update data are related threats at different stages, not only a problem before the first model training.
Evaluate representative performance
A single benchmark score does not prove correctness, fairness, robustness or safety in every deployment context. Define relevant populations, failure consequences, uncertainty and acceptable use. Check leakage between training and evaluation and measure against realistic alternatives. Retrieval-augmented generation can supply useful context but does not guarantee truthful output or safe handling of untrusted retrieved instructions. Changes to documents, prompts, tools or retrieval indexes can change risk without changing the model’s weights.
Control deployment and change
Version the relevant model, data/configuration, dependencies and evaluation evidence. Protect build/training pipelines and production identities, and verify approved artifacts rather than trust a filename alone. Monitor actual behavior and distribution changes, with accountable incident and rollback paths. A previous evaluation is evidence within its scope, not perpetual approval after every integration change. Retirement should consider retained data, access tokens, endpoints and dependent workflows. These chapters provide written AI lifecycle/safety foundations and do not train a model, send data to an AI service or perform an automated consequential decision. Current credential logistics and jurisdiction-specific legal duties are not invented.