Meridian
A personal AI operating system. Live in production. Built and governed solo across 47 sessions.
Meridian is a personal knowledge system — email corpus, knowledge graph, AI extraction pipeline. The data is private. The governance architecture is public. The patterns that emerged from 47 sessions of building a production AI system from scratch are the artifact.
How to establish a governing operating contract for AI-assisted work that enforces fabrication zero-tolerance, source attribution, and pre-generation verification at the behavioral layer.
Extraction is proposal, not fact. The complete staging pipeline from AI output to production state — with every bypass route explicitly closed at the schema, API, and authentication layers.
One credential per runtime class. Four independent revocation units. Surgical revocation without collateral damage. The pattern that prevents a compromised MCP server from taking down the rest of the system.
Tool descriptions as compiled behavioral artifacts. Five required elements per tool definition that prevent agentic non-determinism at runtime.
How to govern a multi-session AI-assisted project so decisions don't drift, context doesn't evaporate, and a new session picks up where the last one ended in under two minutes.
How to evaluate AI models for enterprise deployment using system cards as behavioral specifications — not benchmark scores. Worked example: the Opus 4.7 to Sonnet 4.6 migration.
Nine principles. Each one applies cross-project, to every engagement, without exception.
- Fabrication is not mitigated by correcting it afterward — it is prevented by architecture.
- Extraction is proposal, not fact — human authorization required before production state.
- Every runtime class holds exactly one credential with minimum required scope.
- Tool descriptions are compiled behavioral artifacts, not documentation.
- Decisions are locked with rationale and rejected alternatives — never quietly reversed.
- Pre-flight before execution — read-only analysis gates every code-changing prompt.
- Governance before code — the legal and architectural questions precede the sprint.
- A system you can trust must be a system where you can surgically revoke access.
- Model selection is a behavioral contract evaluation, not a benchmark comparison.