Glemad institutional charter
Charter 01 / 10 principles
These principles govern how Glemad designs, evaluates, and deploys intelligence for security-critical environments.
Model design · Evaluation · Governance · Product decisionsEvidence, uncertainty, authority, and consequence should remain visible throughout a decision.
The charter
The Ollandi model class is developed for environments where incomplete evidence, adversarial pressure, and operational consequence can converge. Performance alone does not settle whether a system should be trusted. Its reasoning must be inspectable, its authority explicit, and its effects reviewable.
The ten principles
Each principle defines a condition to preserve across research, evaluation, governance, and deployment.
A conclusion must remain attached to its evidence and understandable to the people responsible for the decision.
Separate observation from inference, expose uncertainty, reject unsupported claims, and preserve consistency across extended reasoning.
Security intelligence must retain predictable boundaries when conditions become noisy, incomplete, adversarial, or degraded.
Evaluate repeatability, conflicting inputs, high-load conditions, manipulated context, and the point at which the system must escalate.
Capability does not create authority. Every action must remain inside explicit institutional permission.
Define role-aware constraints, escalation paths, policy hierarchy, revocation conditions, and reversible action wherever consequence permits.
A security conclusion must leave a record that can be examined independently of the system that produced it.
Preserve the decision path, supporting evidence, policy checks, and a clear boundary between source data and model interpretation.
Intelligence must reflect the operating environment, responsibilities, risk thresholds, and obligations of the institution it serves.
Account for access roles, internal policy, regulation, infrastructure dependencies, operational consequence, and local definitions of acceptable risk.
A defensive system must be examined for the new failure paths its own reasoning or action could create.
Test incorrect escalation, signal misinterpretation, speculative conclusions, overconfidence under uncertainty, and propagation of flawed assumptions.
Infrastructure events must be understood across the dependencies through which impact can travel.
Reason across domains, upstream services, downstream consequence, coordinated behavior, and patterns that indicate a cascading incident.
The system should strengthen accountable human judgment and make intervention possible before consequence grows.
Present clear recommendations, confidence, supporting evidence, unresolved conflict, and the precise conditions that require human review.
Trust must be renewed as models, environments, and adversarial methods change.
Use threat-informed stress tests, environment-specific evaluation, adversarial challenge, behavior review, and findings that alter design or deployment boundaries.
Responsibility continues after release because security-critical systems operate through changing conditions and accumulating consequence.
Maintain governance paths, safe defaults, operational boundaries, post-deployment review, and stewardship focused on reducing systemic harm.
Application
A finding should narrow a boundary, revise an evaluation, change a release decision, or require new evidence. Glemad applies this charter to the Ollandi model class and to the products through which that research reaches operators.
Examine our model evaluation work