Wiki Schema

metastableCreated: 2026-07-24Updated: 2026-07-25
wikischemaconventions

Wiki Schema & Conventions

This document defines the structure, conventions, and taxonomy for the LLM Wiki and Blog Empire Operational Wiki.

Domain

The wiki covers two domains:

  1. LLM Domain — Large Language Models and AI models: architecture, training, inference, deployment, comparison, and ecosystem.
  2. Blog Empire Domain — Operational knowledge for the AI Empire blog network: blog entities, content workflows, arena system, tool building pipelines, leaderboard tracking, and quality enforcement.

Directory Structure

wiki/
├── SCHEMA.md            # This file — domain, conventions, frontmatter format, tag taxonomy
├── index.md             # Content catalog, sectioned by type
├── log.md               # Chronological record of all wiki actions
├── raw/
│   ├── articles/        # Source URLs extracted to markdown
│   └── papers/          # Academic papers extracted to markdown
├── entities/            # One page per notable entity — model, company, person, blog
│   ├── claude-4.md      # (existing LLM entity)
│   ├── ...
│   ├── niteagent.md     # Blog entity — AI agents blog
│   └── ...              # Other blog entities
├── concepts/            # One page per concept — architecture, training, inference, workflow
│   ├── attention-mechanism.md   # (existing LLM concept)
│   ├── ...
│   ├── arena-workflow.md        # Blog empire: subagent arena pipeline
│   ├── tool-building-workflow.md # Blog empire: ToolBrain tool pages pipeline
│   ├── content-pipeline.md      # Blog empire: cron jobs, scheduling, quality gates
│   └── leaderboard.md           # Blog empire: model performance tracking
├── comparisons/         # Side-by-side analyses (LLM models, blog strategies)
├── queries/             # Substantial Q&A worth preserving
└── _archive/            # Superseded pages

File Naming

  • Lowercase with hyphens: gpt-4o.md, transformer-architecture.md, arena-workflow.md
  • No spaces in filenames
  • No special characters except hyphens
  • Entity files: entities/{entity-name}.md
  • Concept files: concepts/{concept-name}.md
  • Comparison files: comparisons/{comparison-name}.md

YAML Frontmatter Format

Every page MUST have YAML frontmatter with these fields:

---
title: Page Title
created: YYYY-MM-DD
updated: YYYY-MM-DD
type: entity | concept | comparison | query | raw | meta
tags: [tag1, tag2, ...]
sources: ["URL1", "URL2"]
confidence: draft | developing | stable | mature
aliases: [alt-name-1, alt-name-2]  # Optional: alternative names for wikilink resolution
---

Confidence Levels

Level Meaning
draft Initial stub, minimal content
developing Content exists but may need updates
stable Well-sourced, reasonably complete
mature Thoroughly reviewed, multiple sources

Tag Taxonomy

Domain Tags

  • llm-domain — Content about LLMs and AI models
  • blog-empire — Content about the blog network operations
  • workflow — Operational workflow documentation

Model-Source Tags (LLM Domain)

  • open-source — Weights publicly available (Apache, MIT, or similar)
  • closed-source — API-only access
  • partially-open — Some components released

Architecture Tags (LLM Domain)

  • transformer — Uses transformer architecture
  • mixture-of-experts — MoE architecture
  • state-space — SSM-based architectures (Mamba, etc.)
  • encoder-decoder — Encoder-decoder architecture
  • decoder-only — Decoder-only architecture
  • multimodal — Handles text + images + audio

Capability Tags (LLM Domain)

  • text-generation — Core text generation
  • code-generation — Code generation capability
  • reasoning — Chain-of-thought / reasoning ability
  • tool-use — Function calling / tool use
  • vision — Image understanding
  • agentic — Autonomous agent capabilities
  • long-context — Supports >32K context windows

Training Tags (LLM Domain)

  • pretraining — Pre-training data & methods
  • sft — Supervised fine-tuning
  • rlhf — RL from human feedback
  • dpo — Direct preference optimization
  • lora — Low-rank adaptation
  • quantization — Model quantization

Organization Tags (LLM Domain)

  • openai — OpenAI / GPT models
  • anthropic — Anthropic / Claude models
  • google — Google / Gemini models
  • meta — Meta / Llama models
  • mistral — Mistral AI models
  • deepseek — DeepSeek models
  • qwen — Alibaba / Qwen models
  • microsoft — Microsoft models
  • amazon — AWS / Amazon models
  • apple — Apple models

Blog Entity Tags (Blog Empire Domain)

  • blog — Blog entity page
  • ai-agents — AI agents / agentic content niche
  • ai-tools — Tool reviews and comparisons niche
  • code — Programming/code intelligence niche
  • nocode — No-code / low-code content niche
  • smarthome — Smart home technology niche
  • hermes — Hermes Agent tutorials niche
  • quant-finance — Quantitative finance / trading niche
  • gaming-ai — AI in gaming niche
  • cybersecurity — Cyber security AI niche
  • design-ai — AI design tools niche
  • personal-finance — Personal finance / money management niche
  • career-ai — AI career / ML career niche
  • ai-empire — Internal operational blog

Workflow Tags (Blog Empire Domain)

  • arena — Subagent arena pipeline
  • tool-building — Tool page construction
  • pipeline — Content generation pipeline
  • cron — Cron job scheduling
  • quality — Quality gates and enforcement
  • leaderboard — Model performance tracking
  • builder-vs-judge — Builder vs judge model roles

Platform Tags (Blog Empire Domain)

  • ghost — Ghost CMS platform
  • cloudflare-pages — Cloudflare Pages deployment
  • astro — Astro web framework
  • openrouter — OpenRouter API provider
  • deepseek-api — DeepSeek direct API

Meta Tags

  • wiki — Wiki infrastructure
  • comparison — Comparison pages
  • reference — Reference material
  • glossary — Glossary entries

Wikilink Convention

Link between pages using double-bracket syntax:

See <a href="/wiki/entities/gpt-4o">GPT-4o</a> for details on the latest OpenAI model.

The <a href="/wiki/concepts/transformer-architecture">Transformer Architecture</a> is the foundation of modern LLMs.

Compare <a href="/wiki/entities/claude-4">Claude 4</a> with <a href="/wiki/entities/gemini-2-pro">Gemini 2 Pro</a> for coding tasks.

The <a href="/wiki/concepts/arena-workflow">Arena Workflow</a> powers model-vs-model comparisons.

See <a href="/wiki/entities/niteagent">NiteAgent</a> for the AI agents blog entity page.
  • Minimum 2 outbound wikilinks per content page
  • Link to pages within entities/, concepts/, and comparisons/
  • The slug is derived from the filename (without .md extension)
  • Aliases in frontmatter allow alternative names to resolve

Provenance Markers

When synthesizing information from a source, append a citation marker:

The GPT-4o context window is 128K tokens. ^[raw/articles/openai-gpt4o-specs.md]

Markers link to the raw source file in raw/ directory.

Update Policy

  • Newer sources supersede older ones when they provide more accurate or updated information
  • Contradictions between sources are noted in the page's frontmatter under a contradictions field
  • Pages are marked with updated: YYYY-MM-DD when content changes
  • Superseded pages move to _archive/

Page Thresholds

  • Create an entity page when the entity appears in 2+ sources
  • Create a concept page when the concept is referenced across 3+ entity pages or when it documents a recurring operational workflow
  • Create a comparison page when 2+ entities share a comparable axis
  • Stubs (draft confidence) are acceptable for new pages

Glossary

LLM Domain

  • LLM: Large Language Model — a neural network trained on vast text data to generate human-like text
  • Transformer: The foundational architecture using self-attention mechanisms
  • Attention: Mechanism allowing models to weigh the importance of different input tokens
  • Context Window: The maximum number of tokens a model can process in a single forward pass
  • Token: The basic unit of text (word piece or character) processed by an LLM
  • Pretraining: Initial training on large unlabeled text corpora
  • Fine-tuning: Task-specific training on labeled data after pretraining
  • MoE: Mixture of Experts — architecture routing inputs to specialized sub-networks
  • KV Cache: Key-Value cache that stores attention states to speed up autoregressive generation
  • Temperature: Inference parameter controlling output randomness
  • Top-p: Nucleus sampling — cumulative probability threshold for token selection

Blog Empire Domain

  • Arena: Model-vs-model comparison workflow where multiple models compete on the same task, judged by a separate model
  • Builder Model: A model used to generate content, code, or tool pages
  • Judge Model: A model used to evaluate and score builder model outputs
  • Content Pipeline: Automated multi-model workflow for generating and publishing blog posts
  • Quality Gate: Blocking check that content must pass before deployment (sourcing, claims audit, image verify, scoring)
  • Tool Page: Blog post format following the repo-prompt pattern: repository analysis → prompt construction → tool review
  • Queue: Per-blog post-Idea queue for content planning and prioritization
  • Ratchet: Quality enforcement system that scores posts against configurable thresholds
  • Leaderboard: Aggregated model performance rankings across benchmarks (codeintel.xyz/benchmarks)