This study introduces a verification-centric pedagogy for introductory object-oriented programming in which ChatGPT is positioned as a controlled scaffold rather than an answer source. The design mandates an explain-before-generate routine, a UML-tests-code workflow, prompt-literacy checkpoints, and auditable verification artefacts (unit and mutation tests, AI-usage logs, and oral code walks). Implemented over one academic term across several course sections, the approach surpassed business-as-usual instruction. After covariate adjustment, students achieved meaningful gains on the OOP concept inventory, with the largest improvements in cohesion and explicit invariant specification, as assessed using a design-quality rubric. Verification also strengthened: test coverage and mutation scores increased while defect density decreased. Process telemetry indicated sustained growth in test-first behaviour and in goal- and constraint-oriented prompting. Integrity safeguards were effective, evidenced by low rates of unreproducible segments and frequent interception of AI-introduced faults, without additional time costs. Because the method hinges on tool-agnostic practices and transparent evidence trails, it is readily transferable to other languages and platforms. Limitations include the single-term setting and possible model/version drift. Planned work includes replication, component ablations, analysis of cost and equity, and longitudinal tracking of transfer to advanced coursework and internships.