1 · Human purposes and use cases
Behavioural science, neuromarketing, finance, healthcare and industrial systems define outcomes, affected people and acceptable risks. A conversion rate alone does not establish human benefit.
The expanded journey
Intelligence needs a path to defensible execution.
NSDM asks whether an action is justified. ACE asks whether its execution preserves the evidence, semantics, resources and authority that made it admissible. Behavioural and decision science help define what a useful outcome means.
Distinct origins. A shared engineering question.
NSDM developed as an independent decision-systems research programme. Neuromarketing is a separate application domain; it is not the origin of NSDM. Their connection is a workload-level question: how can AI support human decisions without confusing prediction, persuasion and justified action?
Six layers, with ACE across the execution path.
2 · Decision and behavioural science
Study uncertainty, incentives, cognitive limits and decision quality. Validate measures for the population and context; psychological explanations are hypotheses to test.
3 · Intelligence and models
Neural, symbolic and hybrid systems propose interpretations or actions. Model capability and confidence do not confer action authority.
4 · NSDM decision boundaries
Investigate support, contradiction, insufficient specification, abstention and authority before an action is admitted. Retain the existing papers and benchmark limitations.
5 · Execution evidence
Connect workload identity, reference output, actual backend, lifecycle, allocation, trace and terminal result. An optimization remains a candidate until required checks pass.
6 · Accelerated infrastructure
CPUs, GPUs, TPUs, memory, networks, cloud and data centres supply execution capacity. Select them by workload evidence, rather than treating more compute as better judgement.
ACE: the framework for testing execution.
Accelerated Compute Platform & AI Infrastructure Efficiency Engineering
ACE is the expanding engineering programme for studying how a justified workload travels through its compiler, runtime, accelerator, memory, scheduler and telemetry. It is a cross-cutting evidence framework, not a completed infrastructure product or a replacement for NSDM.
The proposed measure is useful, correct, authorised, traceable work delivered within an agreed service objective, relative to measured infrastructure cost. Utilisation is a diagnostic variable; it is not sufficient proof of efficiency.
Choose the execution substrate by the question.
CPU and edge
Useful references for control logic, small workloads and local processing. Measure memory locality and deployment constraints before assuming acceleration is warranted.
GPU
A candidate for parallel numerical work and model execution. Test numerical precision, memory capacity, concurrency and host/device transfers against a reference.
TPU and cloud accelerators
Future comparators when access, supported operators and compiler/runtime behaviour are established. No TPU benchmark is claimed here.
Cloud and data centre
Add placement, network topology, queueing, isolation, data residency and cost attribution. Simulated placement can test provenance; only physical experiments can establish fabric performance.