Verifiable rewards & the training-side view (RLVR)

Why 2026's coding agents got good, from the model side. Students learn the idea of
Reinforcement Learning from Verifiable Rewards — training a model against signals that
can be checked programmatically rather than scored by preference — and how
"verification engineering" in the RL literature means designing those reward checks
(code for hard constraints, an LLM judge for soft ones). This gives students the
vocabulary to understand why delegating to easily verifiable tasks works so well,
and connects the applied loop/verifier work above to the research that made the
underlying models reliable. Taught at a conceptual level — the goal is literacy, not
training a model.