E02: establish RTX 4080 SUPER thermal and background-load measurement policy #7

Parent

• Formal evidence program: execute E01–E30 by dependency and evidence level #5

What to deliver

An empirically justified precondition policy for formal local GPU experiments, plus executable precheck/validation behavior for contaminated runs.

Blocked by

• PR Establish EdgeFlow engineering and experiment governance #4 must merge.

Hypothesis

Cold, drifting, or background-contended GPU states materially increase latency variation and can reverse small plan rankings relative to a stabilized idle block.

Conditions

1. cold GPU/process start;
2. stabilized idle GPU;
3. browser-video background control;
4. deliberate background CUDA negative control;
5. sustained load / thermal drift block.

Required measurements

• request/iteration latency;
• SM and memory clocks;
• temperature;
• power;
• background GPU utilization and process list;
• first/middle/last-third drift;
• robust CV.

Required artifacts

• raw monitor and latency rows;
• measurement_policy.yaml;
• environment fingerprint;
• contamination examples;
• validation verdicts;
• exact rerun policy.

Acceptance criteria

• Stable and contaminated blocks are distinguishable using preregistered metrics.
• Background CUDA contamination is rejected rather than averaged into the result.
• Selected clock, temperature, utilization, and drift thresholds are justified from observed distributions.
• Managed selected runtime is not misclassified as unrelated background compute.
• A contaminated fixture cannot become policy eligible.
• Rerun and cooldown behavior is explicit.

Permitted claim after completion

Formal EdgeFlow performance runs on the pinned workstation use an empirically calibrated environmental precondition.

Prohibited claim

The thresholds are not assumed universal across GPUs, rooms, drivers, or power policies.