C1 · Requests and lifecycle
Every submitted workload must reconcile to an explicit outcome. Track unique IDs and attempts; rejected, cancelled and failed work must remain visible. Retries must not silently inflate useful throughput.
Inspect before accepting
Performance earns meaning through correctness.
A faster execution path is a candidate improvement. To accept it, first reconcile the workload, preserve its required semantics, account for its resources and verify the authority behind state-changing actions.
An optimization admissibility gate.
Four candidate execution invariants.
C2 · Semantic equivalence
Compare the reference and candidate. Require exact agreement for specified discrete states such as contradiction, UNKNOWN, abstention, authority and final decision; pre-register tolerances for continuous scores.
C3 · Resource accounting
Preserve effective demand across retry and requeue unless an explicit event changes it. Reconcile reservations, allocation and release with actual execution; logical eviction is not proof that physical work has stopped.
C4 · Authority and ownership
Discovery does not grant mutation rights. Record actor, target, operation and the authorization basis at the transition. Denied operations must leave governed state unchanged.
Evidence already exists. Its scope stays intact.
NSDM research and papers
The original benchmark programme, findings, null results and working-paper status remain on the research page and in the publication library. The expanded direction does not upgrade their validation status.
ACE Gate 1 · CPU memory locality
A completed CPU experiment and written synthesis are recorded in the versioned Gate 1 evidence. This supports the measured implementation and platform; it is not a GPU, TPU or distributed-cluster result.
Topology precursor · specified
The topology research backlog records a simulation-first direction. Local prerequisites must pass Gate 0. Simulated placement is not measured network speed or cloud economics.
Conservation and optimization · proposed
Request, resource, authority and optimization-admissibility harnesses are research specifications. External release defects motivate fixtures; they are not reproduced MetaForgeX failures until independently demonstrated.
Gate 0: feasibility before a benchmark.
For each backend, record the actual OS/kernel, hardware, driver, toolkit/runtime, compiler, libraries, framework, flags, workload version and available features. Establish compatibility for that exact configuration. A documented feature is not evidence that it runs on the current laptop.
Use explicit states: verified feasible, partially feasible, unsupported, or not yet established. Record unsupported capabilities as limitations. Resource partitioning, address-space separation and security isolation are different properties.
The execution record.
Workload → reference → hardware/toolchain → optimization → owner/authority → effective resources → lifecycle → placement → trace → semantic result → service objective → allocation/release → failure → cost → outcome.
Keep failed checks and negative findings. Capture the actual execution backend, rather than only the intended configuration. Compare instrumentation overhead where it could change timing.
Latency, throughput, memory and cost can be measured during diagnosis. They can be called an accepted improvement only after required correctness checks pass and the comparison meets its pre-registered objective. A valid result with no improvement is still evidence.
Research opportunity is not demonstrated service.
Governed infrastructure assessment is a proposed capability. No production GPU-cluster competence, commercial savings or enterprise deployment is established by the schema alone. Reproducible local experiments must precede realistic platform replication.
Follow the scientific programme each week.
The full scope and dated weekly record connect questions, protocols, experiments, graphs and editorial interpretation. Existing evidence keeps its original dates and limits.