AI Authoritative Source Ledger

Key jargon

Term	Plain-language meaning
Primary source	The original paper, specification, standard, dataset, or official product documentation.
Authoritative source	A recognized standards, government, or security body providing normative or curated guidance.
Source claim	A statement attributed to what a source's authors report, without treating it as universal proof.
Retrieval date	The date a changing source was consulted.

Key concepts

Use the strongest source type available for the kind of claim being made.
A source can be authentic yet stale, narrow, vendor-specific, or unsupported beyond its own experiment.

Concept map

flowchart LR
    A["Identify claim type"] --> B["Select primary or authoritative evidence"]
    B --> C["Record scope and date"]
    C --> D["Recheck when behavior may change"]

Area	Source	Supports
Transformer	Attention Is All You Need	Transformer and attention architecture origin
Retrieval	Retrieval-Augmented Generation	Parametric plus retrieved non-parametric knowledge
Tool reasoning	ReAct	Interleaving reasoning traces and environment actions in evaluated tasks
Tool learning	Toolformer	Self-supervised learning to invoke external tools
Adaptation	LoRA	Low-rank parameter-efficient adaptation
Preference optimization	Direct Preference Optimization	Direct optimization from preference data without an explicit reward model
Instruction following	Training language models to follow instructions with human feedback	Supervised instruction tuning plus human-feedback pipeline and reported evaluations
Risk management	NIST AI RMF	Govern, Map, Measure, Manage lifecycle
Generative AI risk	NIST AI 600-1	Cross-sector generative-AI risk profile
Secure development	NIST SP 800-218A	Generative-AI and dual-use foundation-model practices extending the SSDF
Adversarial ML	NIST AI 100-2 E2025	Terminology and taxonomy for adversarial machine-learning attacks and mitigations
AI security	NIST Cybersecurity, Privacy, and AI	AI-specific security/privacy programs and publications
AI management system	ISO/IEC 42001:2023	Requirements for establishing and improving an AI management system
Application security	OWASP GenAI Security Project	LLM and agentic application risk guidance
Adversary knowledge	MITRE ATLAS	AI-system adversarial tactics and techniques
Tool protocol	Model Context Protocol specification	MCP roles, lifecycle, transports, and capabilities
Agent design	Anthropic — Building Effective AI Agents	Workflow versus agent distinction and common orchestration patterns
Framework	LangChain documentation	Current framework capabilities and APIs
Stateful agents	LangGraph documentation	Graph/state/checkpoint orchestration concepts
Open models	Hugging Face Transformers documentation	Model, tokenizer, training, and inference interfaces
Tokenization	Hugging Face Tokenizers documentation	Current BPE, WordPiece, Unigram, normalization, and pre-tokenization interfaces
Decoding	Hugging Face generation strategies	Greedy, sampling, beam-search, and decoding controls
Neural networks	PyTorch tutorials	Official practical tensor, autograd, model, and training material
Local inference	Ollama documentation	Current Ollama runtime and API behavior
Automation	n8n documentation	Workflow, credentials, execution, and deployment behavior
Evaluation	OpenAI evaluation best practices	Task-specific evals, datasets, graders, and continuous evaluation
Judge evaluation	Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena	Author-reported study of model judges and documented limitations
Graph retrieval	From Local to Global: A Graph RAG Approach	Author-proposed graph-based approach for corpus-level, query-focused summarization
Telemetry	OpenTelemetry generative-AI semantic conventions	Evolving trace and metric attribute conventions for generative-AI systems

Use rule

Original papers support what their authors proposed and reported on their experimental setup; they do not prove universal production behavior. Official product documentation is authoritative for a dated product interface, not an independent quality comparison.

Staleness

Check model names, prices, API schemas, context limits, framework syntax, benchmark leaders, and security lists at time of use. Record retrieval date in any operational decision.