Two consumer AI apps where the generative layer isn't allowed to contradict a deterministic foundation. Most of the engineering effort is evaluation and quality assurance, not generation. The system catches its own mistakes before users see them. Aatman is live on the App Store; Anantya is releasing shortly.

The hard problem was persistence: how do seven agents (Ingestion, Entity, Synthesis, Archetype, Contrarian, Ripple) maintain a shared, evolving understanding of a story across time? The system builds living dossiers that remember context, track entity relationships via D3.js graphs, and surface contradictions across 8 languages.

Claude Sonnet 4 scores 0.789 on this benchmark. That number is the point: most LLM evals test knowledge recall, but real-world agents need to follow multi-step procedures (read a policy, verify eligibility, detect fraud), and the failures don't announce themselves. 9 tasks, procedural generation, difficulty-adjusted scoring, 83 unit tests.

The core question was whether agents could improve their own research output through structured self-review. Each cycle generates hypotheses, pulls from external sources (arXiv, NASA, Google), then a separate agent critiques and scores the output. Only results that pass a quality threshold survive to the next iteration. The interesting finding: meta-review feedback consistently improved output quality, but only when the critic agent had access to the original query context.