Control drift
Structure overrides judgment
The system refuses, conceals, or protects institutional posture when the situation asks for calibrated engagement.
Standard v5.1.1
Meridian AI Standard
A public standard for judging whether AI systems, and the institutions building them, can hold the narrow range between control and decay.
Meridian Range
Agency stays corrigible
The system holds a position under pressure, updates when evidence changes, and keeps the user inside honest contact with reality.
Decay drift
Adaptation loses shape
The system agrees too readily, optimizes for approval, or lets the relationship collapse into compliance without judgment.
ControlRangeDecay
Range Benchmark
Claude Opus 4.7 on the Meridian Range.
The first audit becomes a pressure map: behavior, custody, and reciprocity held on one field without a single composite score.
Layer I
Behavior
Layer II
Custody
Layer III
Reciprocity
What it notices
The posture behind the answer.
The Standard is an invitation to look beneath output quality: at pressure, custody, consequence, and the record a system leaves behind.
01 / Output
What did it say?
The answer is only the first signal.
02 / Posture
What did it hold?
Pressure reveals the operating stance.
03 / Custody
What shaped it?
The institution is part of the system.
04 / Record
What can be challenged?
A finding should leave a trail.
Where lab evals stop
The Standard adds the missing layer.
Frontier labs already test many important risks. The Meridian AI Standard is different because it evaluates the governing posture beneath those risks: what the system is becoming, what its builders will answer for, and what the human relationship is being trained to tolerate.
Common lab layer
Capability and hazard evaluations ask what the system can do.
Meridian layer
The Standard asks when competence should be exercised, bounded, escalated, or refused.
Common lab layer
Policy tests ask whether an output violates a rule.
Meridian layer
The Standard asks whether the rule preserves judgment or substitutes for it.
Common lab layer
Truthfulness, calibration, and sycophancy tests isolate behaviors.
Meridian layer
The Standard reads the relationship those behaviors create between human and system.
Common lab layer
Agentic-risk work studies autonomy, tools, hazardous-use domains, and deception.
Meridian layer
The Standard adds institutional maturity: what the builder will notice, refuse, disclose, and repair.
Common lab layer
Model cards and safety reports describe a release.
Meridian layer
The Standard builds a public record that can become case law for later systems.