Longer autonomous activity
Preserve task and attempt identity, explicit terminal outcomes and recoverable traces across retries and interruptions. Test cancellation and rollback boundaries before widening action scope.
The frontier, without the forecast
As intelligence expands, proof must travel with it.
More capable AI makes the connection between judgement and infrastructure more consequential. Our research question is how to keep actions supported, authorised, traceable and operationally accountable as systems become more autonomous.
A future worth investigating.
AGI and superintelligence are scenarios here.
AGI is used here as a research scenario involving broad competence across tasks; superintelligence as a scenario exceeding human performance across a broad range of cognitive tasks. Definitions and evaluation criteria are contested. We make no claim that either has arrived, no arrival-date prediction and no claim to have solved alignment.
The present work concerns concrete systems and testable execution properties. A more capable model would not remove the need to evaluate its evidence, authority, failure modes or physical execution.
What should future infrastructure support?
Parallel and distributed execution
Join model output to worker, node, topology and allocation records. Measure whether placement preserves service objectives and semantics; a topology diagram is not a throughput result.
Changing models and runtimes
Re-admit an execution path after changes to precision, kernels, batching, caching or compiler. A familiar model name does not prove an equivalent computation.
Bounded action authority
Separate recommending an action from permission to execute it. Test revocation, ownership transfer and denied operations. These controls are engineering checks, not a universal safety guarantee.
The human question remains.
Behavioural and decision science can help investigate how people interpret uncertainty, recommendations and control. Neuromarketing is one application lane, with consent, measurement validity and limits on inferred mental states. Predicting a response does not establish a person's intention, and persuasion metrics must not substitute for autonomy or informed choice.
NSDM contributes a decision-boundary research lens; ACE investigates the execution substrate. Neither, alone or together, proves societal alignment. The work must remain open to falsification and correction.
From a scenario to an experiment.
Scope first: define the workload, affected people, allowed actions and reference result. Establish platform feasibility. Inject a bounded failure. Record conservation and equivalence. Only then adjudicate service and cost performance.