PROJECT FOUR - EVALUATION AND OBSERVABILITY

CHECK YOUR UNDERSTANDING • What is the difference between a span and a trace? ◦ Answer: A span records one timed piece of work, while a trace is the collection of spans for one whole task. • What pattern in a trace signals a runaway loop? ◦ Answer: The same tool called many times with identical input and no new information. • Why is the guard placed in the harness instead of asking the model to avoid loops? ◦ Answer: Because the model cannot be relied on to notice its own loops, while the harness can count identical calls with certainty. • Why run all fifteen eval items after fixing one problem? ◦ Answer: Because a fix for one case can quietly break another, and only the full set reveals that. • Name the three case types and give one lab example of each. ◦ Answer: Normal such as resume_01.txt, edge such as the long resume_12.txt, and adversarial such as resume_injection_test.txt or resume_empty.txt.