One exam and its mathematics lesson had four legitimate student counts: 521, 530, 522 and 523. How request identity, applicability filters, seen-blank versus unseen coding and a silent library helper decided who and what entered an IRT calibration.

Three new exams showed far lower maximum IRT discrimination than the archive. Before blaming the questions, I checked the 2PL model, a stored Lambda, a misleading 3pl name and two code paths, and learned to read a coefficient as the output of a pipeline.

Structuring a Minimal API service that stays readable at fifty endpoints: feature folders, route groups, TypedResults, validation, ProblemDetails, output caching, rate limiting, health checks, an upgrade path from .NET 6/8 and integration tests.