Hydra - One cortex, many heads local-first AI control plane

Hey Product Hunt 👋 Every AI router I tried optimized for one thing: cost. None of them could answer the question I actually cared about: how confident are we that this answer is correct, and what's the blast radius if it's wrong? So I built Hydra: a local-first Trust Control Plane for AI. Three ideas run the router: Cost/quality Pareto routing: single models are dots; Hydra's routed curve hits equal-or-better quality at lower cost by picking the best head per task. Optimal stopping (SPRT): it polls heads until a sequential probability ratio test crosses your target confidence (e.g. 0.95), then stops early and banks the unspent heads. Percolation-κ blast radius (Molloy–Reed): κ = ⟨k²⟩/⟨k⟩ over the code graph. A hub file with many transitive dependents is cascade-capable, so Hydra raises the confidence bar before touching it. All of it runs on-device with ~0ms routing overhead and zero network calls to decide. Local Qwen is the always-on terminal fallback, so a dispatch never hard-fails on a rate limit. It's MIT-licensed. I'd love to get torn apart on the details: Is SPRT the right stopping rule here, or would you calibrate confidence differently? And is κ a fair proxy for blast radius?