Responsible deployment framework

Deployment proceeds by evidence, authority, and the ability to reverse course.

Glemad moves Ollandi model class capabilities from isolated evaluation into institution-specific use through five explicit stages. A later stage never inherits authority by default.

01Controlled sandbox evaluation02Limited-scope integration03Human-supervised deployment04Controlled autonomy within policy05Continuous monitoring and adjustment

The operating position

Progress is a governed decision.

Each stage answers four questions before the deployment moves: what the system may do, what evidence supports that scope, what permits advancement, and what forces a return. The framework applies those questions to the actual institution rather than to an abstract deployment.

Stage-gated lifecycle

Five stages. Four gates at every stage.

Deployment stage

Controlled sandbox evaluation

Study behavior and failure cases in an isolated environment without operational exposure.

AAuthority: Evaluation access only. No live infrastructure authority.
EEvidence: Results under adversarial, partial, conflicting, and repeated inputs, including policy-boundary behavior.
+Advance: Acceptance criteria are met, material failure cases are documented, and owners approve a defined read-only scope.
RRollback: Unexplained behavior, unstable reasoning, or a control-boundary failure invalidates the stage.

Deployment stage

Limited-scope integration

Evaluate the system against real institutional context while retaining a read-only operating position.

AAuthority: Read-only access to the approved context. The system can interpret and recommend but cannot act.
EEvidence: Reviewed outputs for log interpretation, compliance mapping, behavior summaries, and risk-scoring previews.
+Advance: Outputs remain useful and traceable in the institution's environment, and human review paths are ready.
RRollback: Material context mismatch, recurring signal error, or incomplete evidence returns the deployment to isolation.

Deployment stage

Human-supervised deployment

Introduce operational assistance while accountable operators retain the decision for every action.

AAuthority: The system proposes. A named operator reviews the evidence and confirms each action.
EEvidence: Reasoning, confidence, supporting evidence, operator approval, action, and outcome are recorded together.
+Advance: The institution approves a bounded action scope after its policy checks and rollback path have been exercised.
RRollback: Unclear reasoning, unresolved policy conflict, or outcomes outside the approved expectation remove action eligibility.

Deployment stage

Controlled autonomy within policy

Permit controlled autonomous defense only for actions expressly bounded by policy and operational consequence.

AAuthority: Permission is limited to named actions, roles, environments, conditions, and revocation rules.
EEvidence: Each action retains its triggering evidence, policy decision, execution record, observed effect, and any escalation.
+Advance: The approved scope can enter sustained operation when monitoring, intervention, and rollback controls are active.
RRollback: A policy conflict, expanding blast radius, degraded evidence, or unexpected effect suspends the affected authority.

Deployment stage

Continuous monitoring and adjustment

Review behavior as the model, threat environment, infrastructure, and institutional policy change.

AAuthority: Existing authority remains fixed until a separate review explicitly changes it.
EEvidence: Drift observations, edge cases, incidents, policy changes, model changes, and post-action outcomes enter the deployment record.
+Advance: The deployment remains in scope while its evidence and controls continue to meet the institution's acceptance criteria.
RRollback: A material change that invalidates prior assumptions returns the affected capability to the appropriate earlier stage.

One deployment record

The evidence should survive every transition.

Observations, reasoning, policy checks, approvals, actions, outcomes, and exceptions belong to one reviewable record. A change in model release, operating environment, policy, or authority is treated as a deployment change and assessed at the stage it affects.

This framework governs how Glemad research reaches operational systems. Institution-specific scope is agreed through Ollandi deployment planning.

Discuss an Ollandi deployment