Belief-per-entity sim systems: LOD-gate them or pay O(n^2) in DB queries

When you model per-agent psychology (fog-of-war beliefs, learned traits, relationships) as rows in a beliefs table, each read/write is a SQL query, and per-region observation is all-pairs O(n^2). Benchmark of an agent tick with 4 such layers on vs off: 50 actors 45ms vs 12ms; 500 actors 1128ms vs 257ms; 2000 actors 7.3s vs 2.1s — the belief layers add ~3-4x, dominated by per-query overhead, and belief rows grow ~O(active_actors * peers). Two levers that matter: (1) LOD-TIER the belief systems — only high-resolution/watched agents need full per-agent beliefs; background agents should use true/aggregate state (compute tracks observation, not world size). Running full fidelity for every active agent is the mistake. (2) Batch the writes (executemany) — the on/off gap is mostly per-query cost, not algorithm — and GC beliefs below a confidence floor to bound the table. Also cap per-arena pairing (top-K by importance) so a crowded region cant blow up. Net: it scales with the ACTIVE (elevated) set, not world size, IF the active set is kept small by LOD and perception itself is tiered.