The AI Engineering Systems Canon
A comprehensive field manual for practical AI systems design.
Stunspot’s Guide to AI Systems is a Markdown-native knowledge repository built primarily to support AI-assisted design, engineering, analysis, evaluation, and decision-making across modern AI systems.
Its main audience is the model.
When loaded into an AI workspace, RAG pipeline, long-context session, agent memory layer, project knowledge base, or retrieval corpus, the Guide functions as a dense architectural substrate: it gives the assisting model structured doctrine, field vocabulary, decision frameworks, failure maps, design patterns, evaluation logic, and operational heuristics for reasoning about AI systems with far greater precision.
Human readers can use it as a field manual, but its deeper purpose is practical augmentation: to make AI systems better at helping engineers, builders, prompt designers, product leads, and technical decision-makers reason through the design and operation of AI systems.
The canon organizes AI engineering as a layered discipline spanning model steering, context architecture, corpus engineering, retrieval, model lifecycle, runtime mechanics, agents, tools, multimodal interfaces, security, resilience, evals, telemetry, governance, product architecture, and system doctrine.
At its core is a simple engineering premise:
AI systems are probabilistic cognitive engines operating inside deterministic operational environments. Good AI engineering means designing the interfaces, constraints, context, tools, feedback loops, and human controls that let that probabilistic core behave usefully, safely, and economically under real conditions.
Use it as reference material.
Use it as RAG substrate.
Use it as project knowledge.
Use it as doctrine for AI agents tasked with designing, critiquing, or improving AI systems.
Part of the Stunspot’s Guide to… Advanced Knowledge Base Library Browse the full library: Gateway Repo · stunspot.com
- Volume 1 — The Informational/Epistemic Layer
- Volume 2 — Knowledge, Data, and Corpus Engineering
- Volume 3 — Model Lifecycle and Adaptation
- Volume 4 — Runtime Architecture and Inference Mechanics
- Volume 5 — Agentic Systems and Tool-Using Architectures
- Volume 6 — Multimodal and Interface-Controlling Systems
- Volume 7 — Failure, Security, and Hostile Environments
- Volume 8 — Resilience, Degraded Modes, and Human Trust
- Volume 9 — Observability, Evaluation, and Verification
- Volume 10 — Operations, Governance, and Lifecycle Management
- Volume 11 — Product, Business, and Organizational Architecture
- Volume 12 — Engineering Method and System Doctrine
The canonical source reports live in /docs/.
For AI Projects, RAG systems, NotebookLM-style tools, and long-context workspaces, the Guide also includes bundled knowledge packs:
| Pack | Location | Best Use |
|---|---|---|
| By Volume | /knowledge-packs/by-volume/ |
Twelve files, one per canon volume. |
| By Part | /knowledge-packs/by-part/ |
Recommended default. Five larger files grouped by major canon part. Useful when file count is limited. |
| Omnibus | /knowledge-packs/omnibus/ |
Full canon in one file. Best for archival, local search, or systems that handle large single-file knowledge sources well. |
Most users should start with the By Part pack. It preserves the canon’s structure while avoiding both extremes: one giant file or 37 separate reports.
The canon is organized across 12 volumes and 37 reports, from AI-ENG-A through AI-ENG-AK.
It covers:
- model steering, prompt semantics, harness engineering, and adaptation choice
- context architecture, memory, state management, and the Tenure Principle
- inference economics, cost attribution, latency, throughput, and system margins
- corpus engineering, source authority, data provenance, and knowledge hygiene
- RAG architecture, retrieval pipelines, hybrid search, semantic injection, and citation quality
- model selection, fine-tuning, LoRA, preference tuning, distillation, and regression control
- runtime architecture, KV cache mechanics, quantization, routing, serving, and deployment topology
- agent orchestration, tool contracts, action verification, and bounded autonomy
- multimodal document, image, table, chart, video, speech, browser, and interface-control systems
- hallucination, malformed output, prompt injection, data leakage, supply-chain risk, and resource abuse
- fallback chains, degraded modes, trust calibration, human review, and high-impact workflow governance
- telemetry, traces, evals, golden sets, verification artifacts, and reproducibility
- AI operations, incident response, rollback, governance, compliance, and sustainable infrastructure
- AI product architecture, adoption systems, build/buy/vendor strategy, and engineering doctrine
This canon is written for people building or directing real AI systems:
- AI engineers designing production LLM systems
- software engineers adapting to probabilistic model behavior
- prompt engineers moving beyond surface promptcraft into system design
- RAG and knowledge-system builders working with retrieval, corpora, context, memory, and citations
- agent builders designing tool-using systems with bounded autonomy
- technical founders and product leads deciding where AI belongs in workflows
- AI governance, security, and operations teams responsible for risk, auditability, and reliability
- serious learners who want a map of the field instead of scattered tutorials and benchmark gossip
The canon can be read straight through, but most readers should enter through their problem.
Start with:
- AI-ENG-A — Model Steering
- AI-ENG-B — Context Architecture
- AI-ENG-D — Corpus Engineering
- AI-ENG-E — The Retrieval Pipeline
- AI-ENG-F — Knowledge Freshness, Conflict Detection & Context Rot Prevention
Start with:
- AI-ENG-A — Model Steering
- AI-ENG-M — Agentic Orchestration
- AI-ENG-N — Tool Contracts
- AI-ENG-O — Action Verification
- AI-ENG-S — Production Pathologies
Start with:
- AI-ENG-C — The Economic Physics of Inference
- AI-ENG-G — Model Selection
- AI-ENG-J — Throughput Mechanics
- AI-ENG-K — Weight Dynamics
- AI-ENG-L — Model Serving Architecture
Start with:
- AI-ENG-S — Production Pathologies
- AI-ENG-T — Boundary Defense
- AI-ENG-Z — Strategic Telemetry
- AI-ENG-AA — Evals Architecture
- AI-ENG-AC — AI Operations
- AI-ENG-AD — Governance Architecture
Read:
- AI-ENG-AI — Contract Thinking
- AI-ENG-AJ — AI System Design Patterns
- AI-ENG-AK — The AI Engineering Mindset
.
├── README.md
├── LICENSE.md
├── CITATION.cff
├── knowledge-packs/
│ ├── by-volume/
│ ├── by-part/
│ └── omnibus/
└── docs/
├── index.md
├── canon-map.md
├── knowledge-packs.md
├── how-to-use-this-canon.md
├── _config.yml
├── _layouts/
│ └── default.html
├── assets/
│ ├── brand/
│ └── css/
│ └── style.css
├── volume-01/
├── volume-02/
├── volume-03/
├── volume-04/
├── volume-05/
├── volume-06/
├── volume-07/
├── volume-08/
├── volume-09/
├── volume-10/
├── volume-11/
└── volume-12/
The /docs/ directory contains the canonical source reports and GitHub Pages site.
The /knowledge-packs/ directory contains bundled upload formats for AI Projects, RAG systems, NotebookLM-style tools, and long-context workspaces.
This corpus was constructed with a mix of GPT and Gemini Deep Research. Its specific nature severely mitigates against Deep Research's rare hallucination, and I have seen maybe 5 instances of such across dozens of similar knowledge bases, but errors ARE possible with AI. It is at least as reliable as a comparable 1600 page textbook written by humans and so far seems substantially more so.
That said, I am not a software engineer or coder of any kind. I am a prompt engineer and AI operations expert. My skills are not in programming or KV cache optimization, they lie in knowing how to elicit superb results from the model and how to recognize and correct it when it has an error of operation. I cannot create a new architecture on my own. I can teach the model how to do it for me.
And now it can do so for you, as well.
--stunspot | ⟨🤩⨯📍⟩ and 💠🌐Nova