Flagships for Every Device: The Wave C Training Program

One flagship model per THOX device, trained, gated, and published under a single factuality standard. The Wave C program replaced an accumulated model lineup with a defended slate of four base models, trained all four on one 982-row factuality corpus through a single training plane on Hugging Face Jobs, and judged them by one 49-question disclosure-conformance and verified-recall gate (GATE49) under a publish-on-improvement rule. Two flagships end promoted (ThoxAir Wave C.2, 47/49; ThoxEdge Wave C, 46/49); ThoxMini ships as a documented baseline; ThoxClip stands on its prior baseline. A constant decline responder scores 25/49 on the same gate, and every publish bar is read against that floor. Post-publication served-path measurements (section 8.6) show the shipped Ollama tags scoring materially below the merged weights they were built from — an artifact of serving configuration, closed by a configuration fix and re-measured. This edition (v0.5.0) is a Process v1.0 reissue of v0.4.2 (2026-10-30) in the THOX Research Paper Template v1.0, content unchanged.