LLM Wiki — Synaptic Lattice

robust-llm-wiki

Properties1
tagsentity, tool, llm-wiki, research, schema, governance

Leitura de 2 min

robust-llm-wiki

robust-llm-wiki (KevinYoung-Kw, 9 stars) is a research-backed schema framework for maintaining Karpathy-style LLM Wikis at scale. Extracted from two long-running real wikis. Its goal is repeatable maintenance, not one-shot generation.

Key Research Findings (confidence: high)

Problems from real operations:

  1. Human-read and AI-read goals mixed in one layer → pages become either hard to read or hard to process
  2. Over-reliance on body templates makes page style rigid and drifts from Wiki form
  3. Sources appended at end while body wikilinks are weak → readable but non-navigable islands
  4. Lint depends on heavy manual read/grep → consumes context and time

Karpathy Kernel (confidence: high)

The framework defines four non-negotiable kernel properties:

  1. Must be a Wiki
  2. Must have wikilinks/double-links
  3. Must keep the ingest → query → lint loop
  4. Must remain traceable, auditable, and rollback-friendly

Features

Core Architecture

  • Schema framework — research-backed rules for maintaining LLM Wikis at scale
  • Karpathy kernel — 4 non-negotiable properties: Wiki, wikilinks, ingest→query→lint loop, traceability
  • Layered lint — Property → Link → Content three-tier quality system

Quality & Governance

  • Layered lint model — Property (frontmatter) → Link (wikilinks) → Content (body) hierarchy
  • Rollback strategy — built-in rollback and auditability at every stage
  • Model selection guidance — prioritize lower-hallucination models for ingest
  • Human/AI separation — separate layers for human-read and AI-read goals

Data & Storage

  • Markdown files — standard wiki output
  • Audit trails — traceable operations for compliance

Model Selection Guidance (confidence: medium)

  • Prioritize lower-hallucination models on ingest compilation
  • Secondary verification for high-risk query conclusions
  • Maintain layered lint plus rollback strategy
  • Rich expression can be traded off; hallucination contamination cannot

Implications

robust-llm-wiki is the most academically rigorous analysis of the LLM Wiki pattern’s operational failures. Its finding that human-read and AI-read goals conflict in a single layer directly mirrors issues this vault has faced. Its layered lint (Property → Link → Content) is a better model than the flat lint this vault currently uses. The focus on rollback and auditability anticipates the governance problems that appear at scale.

  • llm-wiki — The pattern this researches
  • lint — Layered lint model could improve this vault’s wikilint
  • error-book — Similar pattern of learning from recurring failures
  • vault-roadmap — Robustness at scale is a Phase 3/4 concern

Sources

  • raw/articles/robust-llm-wiki-research.md

Visão de gráfico

Backlinks