Autonomy with boundaries. Systems that preserve their invariants.

Make the promise explicit before action.

Intent governance, actor identity and authority, obligation tracking, evidence sufficiency, conformance and drift.

• Intent authoring, approval and version chain with a distinct approving principal
• Declared obligations, scope and expiry attached to a published intent

Keep systems within their safety boundaries, even when conditions fail.

Formal contract and invariant-verification foundation: verify admissible conditions, constrain state transitions and model safe degradation when operating assumptions break.

• Formalization of machine-enforceable contracts from declared intent
• Deterministic compilation of supported a deliberately bounded class of hard invariants and invariant checks

Translate intent into enforceable requirements.

Translation of declared intent and obligations into machine-enforceable contract fragments and compiler inputs.

• Intent resolution against authority and compiler inputs
• Translation package generation for a published intent version

Admission is a decision, not a side effect.

Independent admission and authorisation, gate evaluation and signed permit issuance bound to the action.

• 13-gate admission evaluation with explicit refusal reasons
• Signed permit issuance with scope, expiry, nonce and action-digest binding

Coordinate the decision, not only the call.

Policy evaluation, decision coordination, execution state and optimisation across the governance lifecycle.

• Policy evaluation and decision coordination
• Dispatch only against a verified permit

Feasible decisions inside hard constraints.

Invariant-Dynamic Machine decision model and its supported solver capabilities for constrained decisions.

• α-fair single-resource allocation within declared bounds
• Opt-in multi-resource projection (prototype)

Uncalibrated confidence, stated as uncalibrated.

In-process evidence-sufficiency scoring and an explicitly uncalibrated confidence object attached to decision metadata.

• Evidence-sufficiency scoring for a decision context (in-process)
• Explicit uncalibrated confidence metadata (calibrated:false, probability 0, method none)

Pressure and quarantine state at the admission boundary.

Pressure/quarantine state and an admission veto within defined policies, plus a containment request shape. Not an autonomous detector.

• Pressure and quarantine state used by admission
• Admission veto when immune policy state requires it

Propose is not dispatch.

Runtime action mediation for registered agent and tool actions, separating a proposal from a dispatched effect.

• Registered action mediation with proposal and dispatch separated
• Permit verification bound to action and arguments

Execute only under a verified permit.

Controlled execution and resource adapters that apply an action and support independent read-back.

• Permit verification before an effect is applied
• Controlled execution through a reference adapter

One governed external boundary, built carefully.

A permit-gated Kubernetes Deployment scale boundary with receipt, plus connector and adapter libraries. Never live-cluster validated in the reviewed environment.

• Permit-gated Kubernetes Deployment scale operation with receipt
• Connector and adapter libraries for external targets

A successful response is not proof of outcome.

Independent observation, expected-versus-observed reconciliation and linked evidence across the audit ledger.

• Observation of external state after an action
• Expected-versus-observed comparison with explicit divergence

A response is not proof of what happened.

Post-action observation of external state feeding reconciliation — with the current independence boundary stated explicitly.

• Post-action state capture and read-back
• Observed state supplied to reconciliation

Linked records, not a certification badge.

Per-component hash-chained decision, permit, execution and outcome records with lifecycle reconstruction and explicit evidence boundaries.

• Per-component hash-chained audit records
• Chain verification and ordered lifecycle reconstruction

Change only within declared invariants.

Control Plane change assurance: policy gates, change lifecycle, commit invariant and enforcement planning for system change.

• Change lifecycle: declare → attest → admit → candidate → commit → enforce → observe → reconcile → audit
• Policy gates over a proposed change

Network change assurance inside the Control Plane.

In-process Control Plane network change and commit assurance: topology, network gates, potential, routing and containment plan surfaces. Not packet-level enforcement.

• Network admission and permit handling
• Certificate and trust material surfaces

A temporal scheduler inside the control plane.

NiZiNG Temporal Controller — an in-process temporal scheduler and cadence controller. Not a causal validation engine.

• In-process temporal scheduling and cadence control
• Temporal status and ingress surfaces

Read-only findings over recorded evidence.

Read-only drift, incident, state and summary analysis over audit chains, using deterministic rules. No forecasting, tuning or decision authority.

• Read-only drift detection over audit chains
• Incident, state and summary findings

Change assurance in a standalone lab.

Separate change-assurance product: no-regression gate, signed evidence, nftables laboratory and PostgreSQL failover testing.

• No-regression gate for a proposed change
• Signed evidence for change-assurance decisions

A development operator surface.

Same-origin backend-for-frontend operator read surfaces with permission checks and no token leakage. Development/demo surface, not a production operator plane.

• Same-origin backend-for-frontend with no open proxy
• Permission-checked read routes over platform services