Practice Calibration

Use this after attempting independent practice. These are calibration notes for the new cases, not answers to every original open-ended project. The module assessment retains its own scoring and evidence requirements.

Prerequisite calibration

2 MiB/s; changing machine may change hardware, toolchain and workload conditions. Repair units and experiment controls before optimization.

A ready answer includes the example or trace requested in the entry check. A memorized term is not sufficient. Repair the named earlier skill, then change the example and explain it without this page.

Case A

Keep dimensions/data/result and compiler flags fixed; compare checksums, multiple repetitions and access order. Record warmup, distribution and generated-code assumptions. One result does not establish universal cache behavior.

Full practice credit requires the result, assumptions and an explanation of the tempting error. A correctly guessed conclusion without those elements is only partial evidence.

Case B

Stride and working-set interaction can change even at equal element count. State the hypothesis first; measurement decides it for this host.

The key judgment is which guarantee survives the changed condition. Do not transfer a result from the first case without checking that condition.

Experiment calibration

Run the existing locality kit, retain compiler/version/flags, checksums and repeated observations, and inspect generated code. A missed optimization or no speedup is a legitimate result; do not manufacture a performance improvement.

A usable record links the hypothesis to the actual input, observation and interpretation. The negative control must fail for the intended reason, and the repaired setup must pass the relevant check again. Report unexpected outcomes rather than altering the hypothesis afterward. Explicitly state the model, environment and untested boundary; passing a small fixture never proves every deployment behavior.

Recheck

Explain a fresh variant without the key. Retain the original error, repair, help received and delayed result in the existing evidence packet. Self-calibration is not an independent reviewer sign-off or a learner-pilot result. For retrieval-question explanations, use quiz calibration only after attempting the quiz.