nx_closed_loop.nx

nx_closed_loop.nx -- measure-adjust-rerender closed-loop primitive. THE substrate's structural answer to the LoRA / negative-prompt / ControlNet patch problem. Per cardinal feedback-loras-and-negatives-are-patches-not-systems: * Modern image-gen + LLM stacks are BLACK BOXES. You prompt, you hope. Predictability is patched via LoRAs / neg prompts. * Substrate's offer: closed-loop measurement. Generate -> measure -> if off-target, adjust + re-generate. Iterate until target verdict hits or budget exhausted. * User's example: "predictably make an 18 year old human." Today's systems can't. This primitive + caller's age- measurement callback CAN. L4 generic composer. Pure composition -- no new math, no new containers. Caller plugs in the four callbacks: generate_fn : generates a candidate output given a seed and adjustment state measure_fn : measures the candidate and returns a verdict (per-axis i64 values, e.g. measured age, fidelity score, hallucination flag) verdict_fn : judges measurement vs target spec; returns 0 = OK / 1 = ADJUST / 2 = ABORT adjust_fn : given current state + measurement, mutates state for next attempt The substrate handles the loop discipline + budget + sealed verdict semantics. Caller owns the domain logic (what age is "18", what hallucination is, etc). ===== Why this is the anti-LoRA ================================== LoRA approach: 1. Train a LoRA on 1000 examples of "18-year-old faces" 2. Hope the LoRA biases the base model correctly 3. Can't measure: no closed loop; just generate + visually check 4. Fails on out-of-distribution prompts