Measurement-driven failover has a floor: probe interval × misses + smoothing. Fast platforms fall through it. The prediction layer feeds what the radios and the geometry already know — signal trends, positions, velocities, terrain — into the same per-link risk, and moves traffic before the failure. Opt-in, and advisory by principle.
At defaults: 250 ms probes × 5 misses ≈ 1.25 s to a hard verdict, with EWMA smoothing damping the early signal. That’s excellent for a substation and marginal for an airframe opening range at 20 m/s — it travels 25 m between “fine” and “gone”.
A fading RSSI trend, a diverging position track, a ridge sliding into the path — each predicts the failure seconds early. Atlas already carries all three signals: adapters feed radio telemetry, LSAs carry position and velocity, and the terrain crate reads the elevation model.
It biases scheduling and routing costs; it never declares a link dead and never falsifies a measurement. Probes remain the sole authority on death — prediction just makes sure there’s nothing left riding the link when they finally say so. Dashboards always show the measured truth.
Adapter-reported RSSI is smoothed and its slope projected 2 s forward with a 6 dB margin. Projected below the floor (−88 dBm default), the scheduler sees inflated loss — capped at 75%, so the link is never fully abandoned by prediction alone — and traffic migrates before probes miss. Rising trends are never penalised; a link recovering is left alone.
prediction_source = "position": nodes share GNSS position and velocity inside the routing flood. Both endpoints project each other along their tracks; links about to stretch out of range get a rising loss penalty locally and an SPF cost penalty (capped ×4) mesh-wide — so both packet scheduling and multi-hop routes bend before the geometry breaks.
One global range model is wrong the moment the bond mixes radios. [[link]] max_range_m (with an optional warn radius, default 60% of max) scores each radio against its own reach: a 300 m Wi-Fi leg and a multi-kilometre telemetry-radio leg to the same airframe get different risk from the same distance — so the scheduler walks traffic onto the longer-legged link as range opens, before the short-range radio drops a single probe.
Range is not reachability. Two ground stations twelve kilometres apart with a massif between them have no path at any transmit power — and the only nodes that can carry the traffic are moving. The ridge is the one thing in the scene that never moves.
The elevation profile between two positions, sampled: does the straight path clear the ground, and by how many metres.
Radio needs more than a line. A “clear” sightline grazing the ridge still starves the link when the first Fresnel zone is cut.
How much signal survives when an edge cuts the path — collapsed, with the other two, into one terrain risk per link, the same shape as range risk.
nexus-atlas-terrain is built and unit-tested, and it drives the Atlas Sim mission visualiser today; wiring its risk output into live scheduling and SPF is the in-development step. Tracked on the roadmap; the simulator is on the demo index.Prediction doesn’t only choose between the links on one hop — it chooses which neighbour to route through. Relay B is the lower-cost hop and is opening range; relay C is marginal and closing. Route metrics are sticky by design, so a reactive mesh holds B until the path actually fails and pays for everything in flight; the projected margin gives the SPF cost a reason to move ten seconds early instead.
It could — and it would flap. Route dampening exists because re-converging on every telemetry wobble costs more than it saves, so a link-state mesh keeps its next hop while that hop is usable and re-runs SPF when something breaks. Prediction resolves the tension without loosening the damping: the projected margin is just another term in the same per-link cost, so the route moves once, deliberately, at a moment of the operator’s choosing rather than at the moment of failure.
Range risk, signal risk and — once wired — terrain risk fold into a single per-link score with a .max(): whichever threat arrives first drives both the per-packet scheduler and the multi-hop route costs.
One mental model, one tuning surface, one number to read in the stats snapshot. And the fold is open by design — the next layer under research puts declared interference zones — EW and EMCON — on the same map as terrain.
Strip away the features and one substitution is left: the scheduler’s inputs stop being measurements of the recent past and start being projections of the near future. Signal trend, closing velocity, range envelope — and, once wired, terrain shadow — collapse into one per-link risk score that runs seconds ahead of the radio, and both the per-packet scheduler and the SPF costs consume it.
That is the reframing: not “find a path that works”, but a control loop holding a model of its own next few seconds. And the discipline that keeps it honest hasn’t moved — prediction biases costs, probes decide truth, and every projection is inspectable in the stats snapshot.
If your platforms move, reactive failover is already too late. Bring a mission profile and we’ll walk the prediction pipeline against it.