hybrid-loopslisted

# Hybrid Loops ## TL;DR (one screen) > A *cycle* of alternating LLM-and-code layers that *mutually generate each other's working surface* — not just constraining each other, but producing the very inputs the other half operates over. *LLMs bring fluency. Substrates bring discrimination. Code brings restraint.* The LLM writes typed records (often the schema or notation itself); the deterministic layer aggregates and shapes those records into the input the next LLM call sees. They don't just gate each other — they manufacture each other. > > The point isn't LLM-as-pipeline-stage. It's *LLM-as-half-of-a-loop* — and at scale, *layered loops that wrap around each other.* Runtime: one cycle resolving one judgment. Development-time: a critique-patch loop wraps around the runtime, with an LLM-panel reading transcripts of runtime behavior and patching the deterministic layer (or the lens prompts, or the substrate schema) below. The system grows by stacking such loops. > > 5-phase diagnostic: > 1. Find candidate **surfaces** in the project (places where fuzzy judgment is happening or should be) > 2. **Scope** each: A (just call an LLM), B (don't use an LLM), or C (hybrid loop) > 3. Choose **shape**: substrate-as-record (analytical) or substrate-as-vocabulary (interventional) > 4. **Quick design** in 3 questions (input, schema, action) with sane defaults for the rest > 5. **Scaffold** to the surface, not the project. Always include a calibration log and an ablation test. > > Five role