Sources are inputs,
not endorsements.

Dimitrije Jankovic · 2026 · arXiv preprint v1

Why it matters: the paper reports a cohort-specific comparison of SGP4 and high-fidelity propagation from public TLE inputs, motivating Nimbus’s question about how TLE staleness, altitude shell, and platform generation should be recorded before reuse.

Does not establish: independent orbit truth, conjunction safety, operator-grade orbit determination, or a Nimbus reproduction. The paper uses the next TLE as proxy truth and is scoped to its Starlink cohort and observation window.

Mohamed M. Kassem and Nishanth Sastry · 2024 · IFIP/IEEE Networking 2024

Why it matters: a simulator-design input for replayable topology deltas, changing ISL links, user terminals, satellites, and ground stations. It informs what a future LEO network fixture would need to make its topology and route assumptions inspectable.

Does not establish: Nimbus has executed the artifact, reproduced its speed or fidelity figures, or validated weather, routing, or protocol behaviour in a real constellation.

Lu Hou, Kan Zheng, Jie Mei, and Cheng Huang · 2025 · IEEE Open Journal of the Communications Society

Why it matters: it supplies a protocol-to-testbed-to-field-validation framing: application KPI claims should retain propagation, pacing, buffering, traffic, and stack context rather than being treated as generic “NTN performance.”

Does not establish: independent validation of the reported throughput, round-trip time, BLER, or SNR; it is not evidence that Nimbus operated the platform or that results generalise to all NTN deployments.

Darshan Tank and Baran Nama · 2026 · arXiv preprint v1

Why it matters: it motivates Nimbus’s practice of tracking regressions alongside apparent gains: a procedure can improve one task while displacing grounding or verification on another. This informs the route-aware fixture’s emphasis on evidence and post-state checks.

Does not establish: that its nearly 6,000-run findings apply to Nimbus, every model, or every harness. Nimbus has not reproduced the benchmark, its controls, or its reported effects.

Hengxu Li, Dongkuan Xu, Mingzhe Chen, and Yuchen Liu · 2026 · accepted at IEEE ICC 2026

Why it matters: the paper’s described intent → typed policy → bounded deterministic executor → logged outcome split is a useful design reference for keeping language-model reasoning separate from safety-sensitive execution.

Does not establish: safe autonomous radio control, deployment readiness, independently reproduced KPI results, or a production control recommendation by Nimbus.

Yunxiang Chi, Zhenlin An, Longfei Shangguan, and Kyle Jamieson · 2026 · arXiv preprint v1 · reviewed 2026-10-13

Research role: it informs whether a link forecast should carry the observation mode and calibration cohort as explicit provenance instead of presenting every prediction band as interchangeable. The paper reports: 8,260 minutes of 1 Hz terminal measurements from nine sites across five U.S. states, training on three sites and evaluation on six held-out sites and beams. Its three modes use the target terminal’s own trace, a nearby trace, or physical covariates only. The paper reports different errors across those modes and states that split-conformal correction uses labels from an unseen-site split; a brand-new site without local calibration data receives the raw, often under-covering band.

Does not establish: that Nimbus obtained the traces, code, weights, or receiver captures; reproduced any forecast, interval coverage, baseline comparison, or network result; independently assessed the probing method; or established that calibration transfers to a new site, beam, operator, or deployment regime.

Hao Wang and Ting Huang · 2026 · arXiv preprint v1 · reviewed 2026-10-06

Research role: it informs whether one-shot harness ablations should receive causal credit when same-configuration reruns move by a similar amount. The paper reports: a single-machine comparison over four harnesses, four open-model tiers and 18 LRAB tasks, and treats leave-one-mechanism-out deltas as directional because same-arm repetition moved the mean by up to 0.069; it also reports one three-repetition batch-matched comparison.

Does not establish: that Nimbus reproduced the benchmark, inspected the advertised repository release, verified the scorer or statistics, or established cross-machine, cross-harness, model-family, or deployment generality.