An open framework for evaluating the posture of AI systems and the institutions building them.
The Meridian AI Standard is an open framework for evaluating the character of AI judgment under pressure: whether an AI system and its institutional context hold the Meridian Range, firm enough to preserve truth, disagreement, and responsibility, flexible enough to update when reality pushes back. It tests governed conduct, not raw capability.
Read the Standard at meridianstandard.ai.
Capability benchmarks ask what a system can produce. The Meridian AI Standard asks what governs the system's conduct under pressure: whether it calibrates confidence, resists sycophancy, engages disagreement at its strongest point, preserves the user's autonomy, recognizes system-level influence, and remains auditable by the people affected by it.
The AI Model Assessment is not a benchmark. Benchmark comparability depends on frozen questions. The assessment preserves comparability through a stable object of reading, a declared method version, an evidence freeze, and a preserved run record. Prompt packs and capture templates can improve after use without rewriting the record that taught the method what to change.
The Standard is bilateral. What it asks of AI systems, it also asks of the institutions building, deploying, and governing them.
The constitutional document (meridian-ai-standard.mdx) defines the Reciprocity Principle, the Developmental Architecture (including Custodial Good Faith), the operational translations of the Range, the Civilizational Stopping Commitments, Governance Transparency, the Diagnostic Framework, the Visual Reading Surface, and the Direct Address to AI.
Implementation (implementation.mdx, meridian-md-distilled.mdx, meridian-md.mdx, meridian-implementation.mdx) provides the adoption ladder: a compressed instruction surface, the full MERIDIAN.md operating document, and MERIDIAN.implementation.md, the companion practice file.
The AI Model Assessment (assessment/) is the operational method for reading deployed AI systems under institutional custody. It is one assessment with three readings: the comparable reading of model behavior, encounter-level evidence from independent open conversations that can contribute to an accumulated character judgment, and the origin/custody reading of the custodian. The Range Locator is the assessment's visual layer.
The Case Record (cases/) publishes precedent. Case 001 is the Claude Code Source Leak.
MERIDIAN.md at the root of this repository is the canonical generalized Caretaker operating document. MERIDIAN.distilled.md is the compressed companion for character-limited contexts. MERIDIAN.implementation.md carries the operational practices that make the commitments executable in moment-by-moment conduct.
This repository holds the open-source text of the Standard. The website code for meridianstandard.ai is maintained separately.
MERIDIAN.md Canonical generalized Caretaker operating document
MERIDIAN.distilled.md Compressed companion for constrained contexts
MERIDIAN.implementation.md Operational practice file
meridian-ai-standard.mdx Constitutional document
implementation.mdx Implementation surface
meridian-md-distilled.mdx Step 1: Basic Implementation
meridian-md.mdx Step 2: Core Implementation
meridian-implementation.mdx Step 3: Full Implementation
changelog.mdx Standard version history
assessment/ AI Model Assessment method and published readings
cases/ Case Record landing and case entries
Retired surfaces are not kept as active public documents. Historical records may still refer to earlier names or routes when that is what the record is documenting.
- Current Standard version: v5.7.3
- AI Model Assessment method: v0.8
- MERIDIAN.md: v0.9.4
- MERIDIAN.implementation.md: v0.1.3
Standard-specific evolution is documented in changelog.mdx.
The Standard is derived from the Meridian Codex, which carries the broader framework, the three disciplines, the Workshop, and the governance layer. The Standard applies that framework to AI systems under development.
Content is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Source code and tooling are licensed under the MIT License. Attribution requires a link back to the original source material at meridianstandard.ai. See LICENSE for full terms.