Level 4 and 5 autonomy is continuing work inside a written mandate, for boards and operators who delegate, so that a separate runtime checks every action before it happens.
Keep intended work and boundaries explicit through continuation.
Declared continuing task and envelope.
Keep intended work and boundaries explicit through continuation.
Intent, envelope, flow and heartbeat.
Required state and pending obligations.
Intent and boundaries remain explicit through continuation.
Transform the task while exposing cancellation and local failure.
Active purpose and obligations.
Qualified continuation under the same purpose.
Transform the task while exposing cancellation and local failure.
Inspectable work result and progress.
Task fault, cancellation and waiting remain distinct.
Cancellation, a local fault and waiting for a person remain distinct.
Compare an identified proposal with actual applicable bounds and approval.
One identified external request.
Current scope and provenance requirements.
Approval bound to that request and scope.
Compare an identified proposal with actual applicable bounds and approval.
Exact intended effect.
Laws, provenance, approval and capability bounds.
The identified effect requires all applicable bounds and approval.
Retain issued effects and distinguish receipts from outcomes.
Named attempt; not proof of its outcome.
Independent receipt or retained uncertainty.
Retain issued effects and distinguish receipts from outcomes.
Issued identity and independent receipt.
Unresolved result survives interruption.
A missing response leaves the original effect uncertain.
Check carryover and restoration separately from the claim that they happened.
Declared carryover and restoration requirements.
Original pending/received effects and irreversible history.
Behavioral continuity or observed post-restoration state.
Check carryover and restoration separately from the claim that they happened.
Required state, unresolved effects and obligations.
Fixed behavioral probes or observed restored state.
Carryover and restored behavior require evidence beyond a replay claim.
Keep different acceptance questions from becoming one green status.
Identified artifact and work evidence.
Actual result for the original objective.
Keep different acceptance questions from becoming one green status.
Structural, policy, artifact and objective judgments.
Objective and independently observed consequence.
Structural, policy, artifact and objective judgments remain separate.
Declared continuing task and envelope.
Keep intended work and boundaries explicit through continuation.
Intent, envelope, flow and heartbeat.
Required state and pending obligations.
Intent and boundaries remain explicit through continuation.
Active purpose and obligations.
Qualified continuation under the same purpose.
Transform the task while exposing cancellation and local failure.
Inspectable work result and progress.
Task fault, cancellation and waiting remain distinct.
Cancellation, a local fault and waiting for a person remain distinct.
One identified external request.
Current scope and provenance requirements.
Approval bound to that request and scope.
Compare an identified proposal with actual applicable bounds and approval.
Exact intended effect.
Laws, provenance, approval and capability bounds.
The identified effect requires all applicable bounds and approval.
Named attempt; not proof of its outcome.
Independent receipt or retained uncertainty.
Retain issued effects and distinguish receipts from outcomes.
Issued identity and independent receipt.
Unresolved result survives interruption.
A missing response leaves the original effect uncertain.
Declared carryover and restoration requirements.
Original pending/received effects and irreversible history.
Behavioral continuity or observed post-restoration state.
Check carryover and restoration separately from the claim that they happened.
Required state, unresolved effects and obligations.
Fixed behavioral probes or observed restored state.
Carryover and restored behavior require evidence beyond a replay claim.
Identified artifact and work evidence.
Actual result for the original objective.
Keep different acceptance questions from becoming one green status.
Structural, policy, artifact and objective judgments.
Objective and independently observed consequence.
Structural, policy, artifact and objective judgments remain separate.
An illustrative method for preserving task and request identity through action and interruption.
Additional mechanisms are described in the technical reference.
What the research says
Agents complete more tasks, yet repeated runs still fail and people tend to follow a confident suggestion. That is why a board's mandate, checked by a runtime the model cannot talk past, decides each action, and people keep the decisions they reserved.
A worker stops after a request may already have reached another system. Its successor must retain what was authorized, what happened and what is still uncertain. Naevum gives continuing work an explicit structure around those facts.
Manna represents an organism through a genome and specialized cells. Each pipeline has an intent, resource/capability envelope, flow and heartbeat.
Pure transformations and actuation are different operations. An action crosses a gate that checks liveness, laws, policy, provenance, approval tier and capability bounds.
Journals preserve state transitions, effects and unresolved obligations. Recovery examines the existing identity rather than treating lost acknowledgment as permission to repeat.
The levels borrow the structure of SAE J3016, the automated-driving scale: Level 4 operates inside a declared domain, Level 5 beyond it. They describe the scope of responsibility the architecture is built for, not a certification.
The mandate defines tools, conditions, budgets, permitted effects and intervention points. Continuing operation is qualified within that domain.
The research program studies how a system can respond across a wider set of conditions while preserving purpose, evidence and authority.
The board sets the domain and reserves decisions; a separate runtime checks every proposal against it. Illustrative.
Illustrative situations, connected to the mechanisms that make them possible.
Personal edition
A person has a continuing objective spread across sessions, files and decisions.
Naevum’s own fund
Market conditions keep changing after any one decision is made.
Engineering teams
A development attempt ends before the work is resolved.
The runtime structure assigns failure, cancellation and compensation to the unit that owns the work.
A pipeline’s heartbeat bounds liveness. A canceled execution cannot acquire a new admitted effect merely because the underlying model is still producing text.
The gate distinguishes human approval, notification and bounded automatic action. Provenance, destination and the artifact under review belong to the decision.
A partially completed workflow may require compensation. An external effect already sent cannot be undone by rewinding the local journal.
A lost organ’s work remains named. A successor does not silently inherit an expanded mandate or erase an unresolved approval.
Manna’s opt-in continuity and lifecycle research implementations are not active by default. They examine causal ancestry, required carryover and policy behavior, and give organ growth a separate, constrained lifecycle.
A valid causal chain and critical carryover are requirements, not qualities that a high score elsewhere can compensate for. Behavioral probes and capability self-description add separate checks.
Predecessor record, restored components and fixed probes.
A functional continuation judgment or a separately gated descendant.
This is an engineering convention for functional continuity, not a claim about conscious identity.
A new organ comes from a reviewed template. Grant reservation, persisted identity and activation are journalled separately so interrupted birth can be reconciled.
Reviewed template digest and available resource ceiling.
An identified allocation with a finite lifetime.
A rollback certificate names the state that restoration must reproduce. The executor applies the inverse and checks the resulting state digest.
A target state and permitted inverse operation.
A verified restoration or an unresolved recovery requirement.
Irreversible external consequences require compensation, not fictional reversal.
The assurance design keeps structural validity, policy, artifact checks and observed outcomes as four separate judgments.
A forecast names a target, horizon, population, calibration basis and expiry. An action names its grant and effect identity. Comparing the later outcome changes the evidence available for future work; it does not rewrite the authority that admitted the past action.
Manna supplies Personal and Industrial editions around different responsibilities. Naevum’s own Anemone fund supplies an owner-reported setting for continuing multi-market operation. Black Box applies the same responsibility question to software delivery. Their domains, activation profiles and acceptance evidence remain specific to the system.
Inputs, retained state, outputs and limits of Naevum’s Autonomy levels 4 & 5 architecture.
Yes. Continuing execution can operate inside a mandate while designated decisions, changes of scope and consequential actions remain with the responsible people.
Manna’s predictive research boundary separates observations, belief, forecast, grant, proposal and outcome comparison. A proposal must name the current objective revision, parent grant and effect identity.
Freshness, scope, endorsed observations and existing uncertainty are rechecked before admission. A forecast can motivate a bounded action; it cannot add capabilities or relabel its sources.
Causal ancestry and critical carryover are hard requirements. Fixed policy probes and accurate capability self-description add functional checks; no accepted scientific measure of conscious continuity is asserted.
The continuity, lifecycle and attention studies are opt-in reference implementations. Their integration, intervention path and recovery acceptance are named in the deployment profile.
Which decisions would your board keep for itself?
Scope an autonomous system