arXiv:2610.41707v1 Announce Type: new Abstract: Load balancing in best-effort datacenter fabrics faces a tradeoff. Order-preserving schemes such as Equal-Cost Multipath (ECMP) retain loss semantics but use path diversity coarsely, while packet spraying balances load well but removes the ordering signals needed for fast, precise loss detection. We present Ordered Chaos, an end-host design that preserves enough structure for reliable loss detection without giving up fine-grained multipath routing. Reactive Subflow Spraying (RSS) divides each flow into ordered, dynamically routed subflows. Precise Fast Loss Detection (PFLD) combines this ordering with probes to distinguish reordering from congestion, corruption, tail, and retransmission losses within one RTT. Across simulations, AI and HPC workloads, and a DPU prototype, RSS+PFLD improves tail flow completion time by up to $3\times$, eliminates up to 99% of timeout-based recoveries, and accelerates collective and training communication by up to 35%.