AI Engineering Rubric

Dimension	Repeat	Pass	Strong
Baseline and evaluation	Demo examples only	Frozen set and meaningful baseline	Segmented evaluation with uncertainty and error analysis
Reproducibility	Undocumented manual run	Pinned code, data, model, and configuration	Clean-room reproduction succeeds
Safety and security	Generic risk prose	Concrete abuse tests and mitigations	Residual risk is measured, monitored, and rehearsed
Operations	No production evidence	Latency, cost, telemetry, fallback, and runbook	SLOs and failure drills drive revisions
Defense	Cannot explain model limitations	Defends value, tradeoffs, and failures	Clearly rejects inappropriate AI use cases

Passing requires at least Pass in every row. A strong average cannot compensate for missing safety, evidence, or fallback behavior.