COMPOUND PULSEStart qualification
LUMA · QUALIFICATION SPRINT

Turn an AI infrastructure decision into a qualified resource boundary.

Freeze the accepted-service contract, then let PEX find the minimum PASS, reproduce the adjacent FAIL, identify the binding resource and test the highest-value software/runtime or hardware intervention. The Sprint returns a causal baseline-versus-intervention decision—not just a benchmark score.

WHAT THE SPRINT PROVES

Exactly what is necessary, what fails first and what actually moves the boundary.

Machine-for-service and service-for-machine are evaluated together because a software change can alter the required machine and a hardware change can alter the best software configuration.

01

Accepted service

Quality, context, concurrency, TTFT/throughput, reliability and economics are frozen before the intervention is judged.

02

PEX boundary

Minimum PASS → adjacent FAIL → binding resource. The actual limiter is named rather than inferred from generic utilization.

03

Candidate intervention

Test software/runtime choices such as quantization, KV policy, batching, placement, scheduling or compression, or a bounded hardware change when evidence points deeper.

04

Qualified decision

Re-test the same accepted service and return the change in physical requirements, accepted service, energy/cost and the next newly binding resource.

First intake is non-confidential. Keep checkpoints, raw production traces, credentials and customer-confidential data out of this form. If needed, a controlled technical channel is opened after scope/confidentiality alignment. Need the methodology first? Read the LUMA / PEX FAQ →

SPRINT OUTPUT

A reproducible deployment decision.

The output is a qualified resource boundary tied to the service actually required.

A

Service contract

The exact accepted outcome and acceptance metrics held fixed across baseline and intervention states.

B

Resource Boundary Report

Minimum PASS, adjacent FAIL, binding resource, relevant telemetry and the complete machine state around the boundary.

C

Intervention evidence

Measured or otherwise explicitly qualified software/hardware intervention effect with overheads separated from projection or third-party claims.

D

Next decision

Deploy, retain current state, expand testing, open Economic Control, or move to a deeper design-in only when the evidence justifies it.