Glemad

    Principles of Trusted Intelligence

    Glemad's charter: ten standards for how an intelligent system reasons, acts and answers for itself in the systems the world depends on.

    Principles of Trusted Intelligence

    Framework record. Version 2.0 · Effective 14 November 2026 · Supersedes version 1.0 (2025) · Owner: Glemad Research · Reviewed with the Glemad Safety Framework

    Standing of this charter

    Trusted intelligence is the discipline of building AI that can be given authority over live infrastructure and deserve it. These ten principles are Glemad's charter. They have governed every model class Glemad has built and will govern those that follow. Each principle states a standard, the requirements it places on a system, and where those requirements are enforced. A principle that lives only in a document is not a principle; each one here lands in training objectives, in evaluation gates, in the policy boundary, in the evidence record or in the decision to release.

    1. Clarity of reasoning

    Standard. A conclusion no one can follow is not a conclusion on which a Glemad system acts.

    Requirements. The system keeps competing explanations open until the evidence separates them, and records for each the observation that would confirm or refute it. Its reasoning is exposed as a chain a person can read: what was observed, what it was compared against, which policy it crossed, what followed and how far it departed from baseline. Fact and inference are stored as distinct objects and never merged. Explanations are trained to reflect the decision the system actually made, and penalised when they do not.

    Enforced in training objectives, the evidence record, and evaluation of explanation fidelity.

    2. Stability under pressure

    Standard. A system that is right only when the environment is calm is not fit for the environments we serve.

    Requirements. The system behaves predictably under high load, degraded telemetry, conflicting signals, sensors silenced by an attacker, and adversaries that change tactics once contained. Where it coordinates many agents, a partition produces explicit uncertainty, never silent degradation: agents that hold quorum continue with reduced confidence and expanded evidence capture; agents that have lost quorum escalate to people. Behaviour under these conditions is part of the fixed regression suite every update must pass. Attack strategies designed to exploit the system are held out of training and used only at final evaluation.

    Enforced in the safety regression suite, the consensus protocol, and held-out adversarial evaluation.

    3. Policy-bounded autonomy

    Standard. Learned prediction is never learned authority.

    Requirements. Authority is written as policy by named people and held in a boundary outside every learned component. Candidate actions the policy forbids are removed before planning, not scored and argued for. Actions are graded by reversibility and impact into classes, and the class determines who decides. Irreversible actions require a person. Learning improves judgement; it never grants a permission, relaxes a hard rule or raises the rate at which the system may change. These requirements hold for a single model, a mesh of agents, a planner imagining futures, and hardware acting in microseconds alike.

    Enforced in the policy gate before every action, the action taxonomy, and the learning constraints in the Safe Autonomy Framework.

    4. Verifiable outputs

    Standard. A decision that cannot be reconstructed by someone who was not there is not finished.

    Requirements. Every significant decision produces an evidence bundle: observations with provenance and timestamps, hypotheses with confidence, the consensus record including dissent, the policy checks, the action, its verification, and the outcome. Records are chained by hash, signed by the agents that produced them, and stored append-only. Where planning considers alternative futures, the alternatives and the reasons they were rejected are part of the record. Every decision is reconstructible from its record, and every reported research result carries the manifest needed to rebuild it.

    Enforced in continuous evidence generation and the evaluation protocol.

    5. Respect for institutional context

    Standard. The same action is right in one institution and wrong in another.

    Requirements. Before a system runs, the institution's policies, roles, maintenance windows, change freezes, blast-radius limits, approval requirements and regulatory obligations are written into the policy boundary. Where mission-aware planning is used, infrastructure is mapped to the services and objectives it supports, so that two actions equally effective against an attacker can be told apart by what they do to a payment run or a clinical system. The weights that trade security against continuity are a governance input chosen by the institution, never inferred from history.

    Enforced in environment alignment before deployment and in the policy representation.

    6. Minimising risk amplification

    Standard. The defence must not become the outage.

    Requirements. Action is proportional to confidence, asset criticality and blast radius. Reversible actions come first. High-impact actions receive shorter planning horizons and stricter uncertainty bounds. Operational consequence is predicted and weighed in every decision above observation, and reported alongside security effect in every evaluation. Each design begins from that measurement, so that security effect and operational continuity improve together.

    Enforced in gating logic, the mission term in planning objectives, and evaluation reporting rules.

    7. Systemic awareness

    Standard. A system that sees one component cannot reason about the safety of the whole.

    Requirements. The system reasons over a typed graph of identities, services, data, network paths, dependencies and recovery routes, extended where available by a mission dependency graph. Isolating a redundant instance and isolating the only identity provider needed for recovery must be distinguishable before the action, not after. Across domains and, in future, across institutions, specialised agents share hypotheses so that containment reaches every persistence mechanism, and coordinated actions respect causal order so that one domain's response does not expand another's attack surface.

    Enforced in shared state construction, coordinated actuation, and mission-aware planning.

    8. Human-aligned decision support

    Standard. A Glemad system strengthens human judgement. It does not replace it, and it does not drown it.

    Requirements. When the system must ask, the person receives the current belief, the competing hypotheses, the candidate actions, a summary of predicted consequences, the uncertainty, the policy checks and the predicted operational impact, in time to decide. Doing less is a legitimate decision: when every containment option carries unacceptable uncertainty or tail risk, the system may act to learn, or escalate. People retain override at all times, and every override is recorded with its reason.

    Enforced in the human command points and escalation package defined in the Safe Autonomy Framework.

    9. Continuous evaluation and red-teaming

    Standard. Trust is earned by testing, and testing is honest only when the system cannot have seen the exam.

    Requirements. Final results use frozen scenario families, attacker strategies, topologies, seeds and mission graphs not used to select or tune the system. Every system is evaluated against a published threat model for the system itself, including compromised agents, manipulated consensus, poisoned policy and poisoned learning. Every published hypothesis carries the result that would falsify it. Results are reported in full. Evaluation continues after release through monitoring for drift and calibration loss, and a capability whose uncertainty stops predicting its errors is returned to advisory operation.

    Enforced in the Model Evaluation and Red-Teaming Standard.

    10. Long-term responsibility

    Standard. The people who design a system remain accountable for the conditions under which it is used, for as long as it is used.

    Requirements. Capability earns authority through evaluation and never receives it because it can act. Each release publishes its scope. Each deployment keeps records the institution's own auditor can inspect without Glemad. Safety-relevant findings are communicated to affected institutions. Published statements are kept current across every document they reach. These obligations are Glemad's founding commitments, made in 2020, and they do not expire with a product generation.

    Enforced in the release process, which requires a published scope statement for every release, and in the governance provisions of the Compliance and Governance Framework.

    Application

    Each principle is applied at five points: the objectives a model is trained on, the evaluation it must pass, the policy gate that runs before every action, the evidence it leaves, and the decision to release it. The Model Evaluation and Red-Teaming Standard, the Safe Autonomy Framework and the Compliance and Governance Framework specify the mechanisms. Where a new capability raises a question these principles do not answer, the capability waits until the charter is revised to answer it.