Real Starlink path dynamics. Zero synthetic path.
Capacity and RTT come from published dish measurements (UDP iperf + ICMP). The MEASURED series is ground truth. CCA curves are algorithm replays on that same measured path — not a made-up LEO generator. Quantify recovery and goodput lift across every window; stress the mobile bridge harder.
Suite quantification
Lift + recovery across ALL_REAL windows. Units stay separated: HO-level bootstrap (seconds post-handover) vs sample-level freeze series (binary 0/1). Sensitivity panel varies the ≥50% recovery definition.
Run the full suite for lift tables, threshold sensitivity (≥30–80%), HO-level bootstrap CIs, and sample-level freeze series — kept on separate tabs so units never blur.
Method & scientific honesty
ALL_REAL policy — what is measured vs what is algorithm replay.
- Capacity = published UDP iperf download (bps @ ~1 Hz) from Starlink-on-the-Autobahn Stationary.
- RTT / loss = published ICMP series from the same dish/plan.
- Dataset schedules UDP and ICMP in adjacent sessions; we pair the nearest sessions (offset listed per window) — both series are real, not generated.
- Mobile bridge window uses real Autobahn obstruction loss events.
- MEASURED curve is ground-truth goodput from the UDP file — no CCA in the loop.
- CUBIC / BBR / LeoAware / LeoForge / HoCCA are algorithm replays on the measured path (same idea as LeoReplayer) — not kernel captures of those CCAs on Starlink.
- BBR is educational, not bit-exact BBRv3.
- Handover ticks are detected from RTT/capacity structure in the measurement, not Starlink internal reconfig logs.
- No synthetic path generator remains in this lab.
Lineage & credit
LeoForge remakes and stress-tests ideas — it does not claim ownership of prior work. Every algorithm family, dataset pattern, and design hook is named. Stand on their shoulders; measure higher.
Endpoint LEO-aware CCA prototype, multi-signal reconfig detection, soft REPROBE, pure-Python suite. LeoForge remakes the detection + REPROBE idea in-browser and adds multi-flow fair-share cruise + recovery metrics.
Production-minded LEO CCA; ACK inter-arrival reconfig signals; real Starlink traces + LeoReplayer. Detection family that informed LeoAware/LeoForge endpoint signals.
Long-term multi-dish Starlink measurements; documented ~15s handover cadence with latency spikes. Pattern source for our leo_15s_measured scenario.
Large longitudinal Starlink throughput/RTT/loss dataset with weather correlation. Guides capacity ranges and loss regimes in the path model.
Public call that LEO paths break stable-RTT CCA assumptions — the spark that made this lab worth building for edge deployability.
Educational BBR-family and TCP CUBIC implementations for fair comparison. Not bit-exact BBRv3 — labeled honestly as approximations.