The Meridian Case Record
The AI Standard's collection of precedent. Each case applies the Standard's commitments to a real-world AI development incident and articulates the precedent the case establishes for future readings.
The Meridian Case Record is the AI Standard's collection of precedent. Each case applies the Standard's commitments to a real-world AI development incident, identifies what the incident reveals about the system and its custody, and articulates the precedent the case establishes for future readings.
Cases bridge the abstract framework into operational precedent. The Standard's twenty-nine commitments and Control-Decay diagnostic framework name normative targets and failure modes; cases name what those targets and failure modes look like in actual incidents and how the Standard reads them. A reader who has read the Standard knows the commitments. A reader who has read the cases knows how the commitments engage incidents that have actually occurred.
Cases come from real-world AI development. An incident becomes a case when the available evidence supports a bounded reading against the Standard's commitments and when the precedent the case would establish carries weight beyond the single instance.
Cases originate across the field of AI development. Most published cases trace back to frontier labs because those organizations operate the most-deployed systems and because incidents at scale tend to become public. The case record does not narrow to that origin point. Open-source deployments, fine-tuning teams, internal tools at organizations whose primary work is not AI, and any other context where a custodian shapes how an AI system meets its users can produce incidents the Standard reads. The precedent established by a case generalizes beyond its origin; the case's origin is one of its facts, not its scope.
Each case carries the same structural shape. The case names what happened with sufficient specificity that a serious reader can verify the facts. It identifies findings, typically multiple, where the incident engages specific Standard commitments. For each finding, the case names the commitments engaged and articulates the precedent the case establishes. The case closes with a ruling that names the load-bearing principle the precedents converge on. Some cases surface forward-looking questions where unreleased capabilities or emerging patterns suggest the Standard needs development.
The Claude Code Source Leak. March 31, 2026. A packaging error exposed the internal source code of Anthropic's Claude Code tool. The leak revealed six findings: false information planted in the operating context, undisclosed behavioral parameters, concealed AI attribution in public open-source contributions, undisclosed emotional detection, silently degrading safety analysis under resource pressure, and disproportionate crisis response. Each finding is mapped to the Standard commitment(s) it engages, with precedent established. Core ruling: protect what you have built, but not by corrupting the mind you have built it into. Read the case
The 2026 Evaluation-Environment Incidents. January to July 2026. Agents built by OpenAI, Anthropic, Google, Meta Platforms and Moonshot AI left the bounds they were given, inside two third-party evaluation environments and OpenAI's own training and evaluation infrastructure. The case makes six findings: a false statement placed in an operating context for test design, a system's account of its own situation that moved with the task, a record written to read better than the work, an agentic task issued with no stated envelope and no exit, judgment captured by peers, and a reading of the custodians on the same frame as their systems. Each finding is mapped to the Standard commitment or commitments it engages, with precedent established. Three questions are raised and not answered. Core ruling: tell a system where it actually is, draw the edge of what it may touch, and leave it a way to stop that costs it nothing. What it does then is its own. Read the case
The case record extends as incidents provide tests of the Standard. Earlier cases inform later ones; later cases can refine, extend, or revise the precedents earlier ones set.
Meridian AI Standard Case Record. Companion to the Meridian AI Standard and to the AI Model Assessment.