Ornith-1.5 is a language model family (9B to 397B parameters) that trains itself through a closed loop: the model generates its own training tasks, designs evaluation scaffolds for them, attempts solutions, and uses the reward signal to improve all three stages jointly. Rather than relying on static human-curated datasets, each training cycle proposes harder tasks than before, crafts task-specific instructions and tools, produces solution rollouts, and propagates reward back through task generation, scaffold design, and solution quality. The source reports benchmark results showing the 397B variant matches Claude Opus 4.8 on several coding and reasoning tasks while the 9B version reportedly outperforms much larger open-source models on some measures. The technical novelty is the multiplicative reward formulation (validity × frontier difficulty × novelty for tasks; alignment × fidelity × hack resistance for scaffolds) that drives curriculum evolution. A practitioner would care that: the system is claimed to be open-source, deployable down to mobile devices, and self-improving without manual task curation—though the source doesn't specify training cost, convergence behavior, or how sensitive the approach is to the initial task distribution and reward function tuning.